Haven’t you heard? White-collar jobs are going away, decimated by AI. Waves of layoffs in the tech sector (most recently at Coinbase and Meta and Cisco) are said to presage what will soon come for all of us knowledge workers. But before you quit your job as a software developer or financial analyst—or tech journalist—and look to join the plumbers’ union, it’s worth considering today’s economic research on whether artificial intelligence has actually begun to devour white-collar work.

Analysis of the data gathered for the US Bureau of Labor Statistics (BLS) shows that the unemployment rate for the jobs potentially most affected by AI is actually lower than that for occupations less exposed to the technology. And, critically in the mind of economists, there are no signs that large numbers of people are shifting from jobs threatened by AI to supposedly safer ones, such as those involving mostly manual labor.

While the current labor statistics don’t preclude a sudden job upheaval in the coming years, they do throw doubt on the inevitability of the doomsday scenarios and the pace at which they’d unfold. Everyone in the AI community, it seems, is predicting that the technology will soon wipe out jobs, and everyone, it also seems, knows some young wannabe workers who can’t find one. Perhaps we haven’t seen any major disruption in the labor market statistics yet, people often say, but just wait. 

But maybe we should pay attention to what the data is showing us. And right now, the numbers paint a picture of a relatively stable labor market in which AI disruptions remain largely speculative.

“It could be disruptive, but the data is telling us right now that disruption is not yet here, and we have time to plan.”

“All of the available evidence to date suggests that AI’s impact on current labor market conditions is likely small right now,” says Erika McEntarfer, a labor economist who headed the BLS until President Trump fired her last fall after a jobs report that displeased the administration. (Not surprisingly, BLS reports of sluggish job growth have continued since her dismissal.)

McEntarfer, who is now a fellow at the Stanford Institute for Economic Policy Research, says the relatively small impact that AI is having so far on today’s labor market “surprises many people, but it shouldn’t. What we know from history is that it takes time for innovations to work their way through changes in industries and changes in occupations. AI is unlikely to transform labor markets until it first transforms businesses.”

McEntarfer points to US Census data showing that only one in five companies are using AI in any business function. “The data are a great reality check on the fear that AI will be enormously disruptive,” she says. “It could be. It likely will be disruptive, but the data is telling us right now that disruption is not yet here, and that we have time to plan.”

Things ain’t great—but the question is why

The US job market, to be sure, sucks for many, especially younger would-be workers. Unemployment rates for recent college graduates stand at around 5.6%, well above the level for all workers. It’s a rate not seen since the pandemic and the years immediately after the 2008 recession. Even more troubling is that hiring rates have been particularly dismal during the post-covid economy, a trend that hits hard at young people trying to enter the workforce. If you’re a recent college graduate and looking for a tech job, no one, it can seem, is hiring.

There are signs that AI is contributing to the pain for the 22-to-25-year-olds seeking jobs in software development and other occupations that are feeling a big impact from AI. But these professions represent just a sliver of the overall labor market. What’s more, it’s uncertain how much blame AI should get for the job woes. Similarly unknown is whether the loss of entry-level jobs in AI-exposed occupations is a harbinger of what’s coming for others or simply an isolated symptom of what economists refer to as a “low-fire, low-hire” labor market caused by a variety of macroeconomic forces.

Insights into these uncertainties will tell us much about our working fates in the transition to an AI economy. There are no shortage of confident assertions and predictions about what is about to happen; while some people forecast the end of work, others say economic history teaches us that technology advances always lead to more and better jobs eventually. 

The honest answer is that no one knows for sure what AI will bring and whether this time will be different. To help figure it out, we need better and far more comprehensive data.

The statistics gleaned from the federal government’s monthly survey of 60,000 households for the BLS provide a broad overview of the changes to the labor market, while academics and even some AI companies have begun trying to gain a more granular view of specific jobs that are being affected. But the existing data-gathering tools don’t adequately explain how AI is affecting the huge and diverse US labor market.

There’s a long list of questions that we don’t have the data to fully answer. How is AI being used in the workplace? Does the increased use of AI mean the technology will replace workers, or will it make them more productive and valuable? Which occupations and skills are most affected? Who is in most peril from the changes? As David Deming, a professor of economics at Harvard University, puts it: “We’re sort of flying blind.”

To gather more insight into some of these questions, Deming and his colleagues have been surveying several thousand people every three months since 2024, asking them basic questions: Do you use generative AI, and how often? Does it save you time at work? Tracking the answers over time gives the economists important clues (it’s used by a little over 40% of workers but adoption varies by sectors) and allows them to estimate productivity gains (they’ve found some, but nothing economy-shaking). It has also helps document how quickly AI has been adopted in the workplace and how it compares with earlier technologies such as the PC and the internet (the pace has been faster but roughly in the same ballpark).

It’s far from a complete picture of how AI is changing work. But it provides some intriguing results; for example, a fair number of workers in manufacturing and other industrial sectors have tried AI. Deming’s results show that while businesses in general might be relatively slow to formally adopt the technology, lots of their employees are using it.

The young are most vulnerable

But by themselves the exposure results are not a true predictor of which jobs will be lost to AI. That depends on the kinds of tasks done by the technology, the extent to which the AI is adopted, various business calculations about the value of workers, and even the costs of deploying AI. But the exposure findings are a valuable starting point. 

In a working paper called “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” researchers at the Stanford Digital Economy Lab looked at 950 jobs, placing the occupations into five categories from least exposed to most. Then they used a vast data set from ADP, the world’s largest payroll provider, to look at employment growth in each of the categories. Their exclusive access to the ADP data set, which is far larger than the one available through the BLS, allows the researchers to better spot impacts by demographic. When they examined what was happening to different age groups, says Erik Brynjolfsson, the director of the lab who led the effort, “it was extremely striking.”

They spotted the drop in head count for 22-to-25-year-olds in the most exposed occupations, such as software development and customer service, beginning in late 2022, when ChatGPT was first publicly released. Other researchers reported evidence that the decline in these jobs began well before ChatGPT and questioned whether the labor market could react so quickly to the introduction of AI technology. 

But while the Stanford researchers acknowledge that other factors in addition to AI probably contributed to the early declines, they say that after controlling for those factors, they saw convincing evidence of a significant effect from AI after 2024 and growing in 2025 to a 16% decline in entry-level jobs in AI-exposed occupations. In contrast, head count grew for older workers in the same occupations, as did the number of jobs in the less exposed occupations.

Digging deeper into the data, the researchers found another important clue, though one that wasn’t totally unexpected. The impact on head counts depended on how AI was being used. It was specifically the jobs where tasks could be automated (that is, AI could do them “with minimal human involvement”) that accounted for the decrease in employment—jobs for people like software developers. In jobs where AI was mainly used but to augment human work, head counts grew faster than the average for entry-level workers.

That’s consistent with one explanation for the woes of many young would-be workers. It could be, according to the Stanford paper, that entry-level jobs depend more on the types of knowledge that people acquire through education but that can readily be mimicked by AI; the authors call this codified knowledge. It might be particularly easy to automate such tasks as entry-level coding. In contrast, older workers have more so-called tacit knowledge, the type based on their experience. That type of wisdom is harder for AI to replace.

Despite the findings about AI’s impact on young workers, Bharat Chandar, an economist at Stanford and one of the authors (along with Brynjolfsson and Ruyu Chen), stresses that it’s still early when it comes to understanding how the technology will affect jobs in the future. It could be that the job loss will spread to older workers and to less AI-exposed occupations, he says. But Chandar says it is also possible that firms and workers will adjust to shifting labor demands, and the effects will level off or even disappear.

A recent paper by economists at the Federal Reserve Board found, not surprisingly, that annual employment growth for coders has slowed significantly—by about 3%—since the introduction of ChatGPT. But here’s a critical detail: Overall employment for coders continues to grow. Employment in coding jobs is still rising, they noted, just more slowly than before 2022. 

In short, coding jobs are not going away, at least not anytime soon. But it’s an occupation that is clearly being transformed by AI.

One of the somewhat surprising wrinkles uncovered by recent research is that wages in sectors highly exposed to AI have risen relatively fast since the introduction of ChatGPT. One explanation is that employers are still willing to pay for the kinds of knowledge and experience that are, at least for now, hard to replace with AI. If true, this suggests not the end of work in AI-exposed jobs but, more specifically, the demise of the typical career model in which young graduates are hired to do software tasks that can be automated and are slowly trained to gain that valuable tacit experience. The earn-while-you-learn model might finally be broken—at least for some occupations.

The simple truth could be that coding skills are no longer a guarantee of a job. That may help to explain the drop-off of computer science majors at schools around the country. Future canaries in the cubicles are sniffing out the dangers of looking for a job when their skills can be matched by AI.

But a closer look at the data shows that students are not necessarily turning away from AI-related careers. Rather, they appear to be tailoring their skills to the changes they see underway as AI becomes increasingly important for various disciplines. Interest is rising in AI-adjacent fields like data science and cybersecurity. One fast-growing major: artificial intelligence itself (a recent addition to many college offerings).

Is this time different?

Anxiety over the potential of AI to replace workers is nothing new. I wrote “How Technology Is Destroying Jobs” in 2013, describing how a slew of new digital technologies, including AI, were beginning to threaten white-collar work. I wasn’t alone. It was a popular theme at a time when the labor market was sluggish and jobs were scarce. 

In one of his last days in office in late 2016, President Obama issued a report written by his top economic and science advisors warning that AI was threatening workers. Among the findings was that automated vehicles—especially driverless trucks—could eliminate 2.2 million to 3.1 million existing US jobs.  Around the same time, one of the pioneers of AI, Geoffrey Hinton, said that “people should stop training radiologists” because it was “completely obvious” the occupation was soon to be replaced by AI.

Perhaps this time is different, and we can put aside the lessons of economic history. Certainly, AI has gained unimaginable powers to do humanlike tasks. Perhaps it will devour jobs in ways that we’ve never seen before. And perhaps that will happen abruptly, without a warning buried in the labor statistics. But the previous bouts of AI job anxiety still hold a prescient lesson: Our real focus needs to be less on the dystopian fears and more on the very real transitions in the workplace that will likely affect millions of people.

“Even if there is not mass or even increased unemployment, the transition could still be very difficult,” says Jed Kolko, senior fellow at the Peterson Institute for International Economics and former undersecretary of commerce in the Biden administration. “And what does a difficult transition period mean? It means people losing jobs, or people’s jobs being redefined in ways that make those jobs pay worse or be less meaningful. And some people whose jobs are threatened may not be able to adapt.”

The more we understand this transition, the better prepared we’ll be to deal with it.  And for that we’ll need better and more complete data.

For McEntarfer, the former commissioner of the BLS, the real question is the speed of any disruption. “If it happens at the normal pace of technological change, labor markets will have time to adapt. If there is a sudden and severe disruption, then that will be a big challenge for policymakers,” she says. “That’s really the most important question facing us right now: how rapid this transformation is going to be.” And, she adds, “we’ll know by watching the data.”

Two decades ago, the country was caught flat-footed by the so-called China shock as free-trade policies led to an influx of imports and the devastation of manufacturing jobs in many parts of the country. It took years for researchers to understand the data showing how the trade policies, generally welcomed by economists, were destroying communities. Today the threat of an economic transformation brought on by AI is far larger and points to potentially far more damage for huge groups of workers.

To head off another devastating labor transition, we will need well-timed government and business policies, especially programs to train and reskill workers. If McEntarfer and other labor economists are correct, we probably have time to design deliberate and effective strategies to manage the transition. But first we need to better understand what is going on—and how fast.

It’s hard to find an economist who is more enthusiastic about AI’s future than Stanford’s Brynjolfsson, who believes that we’re likely on the brink of a huge boost that will transform the economy. “Perhaps the best productivity growth of my lifetime is coming up,” he says.

But Brynjolfsson also warns that a lack of data is severely limiting our visibility into the economic and societal impacts that are coming. At a time when hundreds of billions are being spent on rolling out the technology, he says, “we’re not investing even 1% of that on understanding the transition.”

Artificial intelligence has not so far produced a clean story of mass unemployment. Aggregate employment in developed countries remains broadly stable, and recent assessments have found limited evidence that AI has shifted the headline numbers. But a troubling change may be hiding beneath the surface: the quiet weakening of the first rung of the career ladder.

The most worrisome evidence is showing up exactly where we should expect it first: in early-career hiring. A working paper released in November 2025 by the Stanford Digital Economy Lab found that workers aged 22 to 25 in the most AI-exposed occupations experienced a 16% relative decline in employment after the spread of generative AI, even after controlling for other factors that might affect firms’ employment decisions. An Anthropic report from March 2026 provides suggestive evidence that led to a similar conclusion.

More experienced workers in those same occupations did not suffer the same decline. Employment is not also declining in the entry-level jobs with low AI exposure. The concern is specific to early-career jobs that are exposed to AI.

That is not a minor signal. It suggests that firms may be using AI to substitute for the junior tasks through which people traditionally gain their first foothold—at least for those in jobs where generative AI is used extensively, like software developers, customer service representatives, computer programmers, and information systems managers.

The time is now to make changes in the way we train, prepare, and support young people who are about to enter the workforce. Educational institutions need to reorient for the era of an AI-augmented workforce. Governments must incentivize businesses to hire and train early-career workers. Businesses, in turn, need to recognize the importance of developing a long-term workforce experienced in AI—a process that begins with entry-level workers. And students themselves should take on the responsibility of not only becoming AI fluent but learning how to apply that knowledge in various fields.

In short, we must change the way we have traditionally thought of entry-level work.

This is especially true because the broader labor market for recent graduates is also softening. The Federal Reserve Bank of New York reported that in the fourth quarter of 2025, the unemployment rate for recent college graduates rose to 5.6%, while the underemployment rate (the share of graduates working in jobs that typically do not require a college degree) reached 42.5%, its highest level since the covid pandemic. No single statistic can prove that AI is the sole cause of that deterioration. Hiring in general is way down post-pandemic, and young people are particularly vulnerable to the slowdown. But it would be a mistake to ignore the possibility that AI is accelerating an already difficult transition from school to work.

Behind these statistics is a great deal of personal distress. Recent graduates today often submit hundreds of applications before they receive a single offer, and surveys consistently find elevated rates of anxiety, financial precarity, and burnout among young workers in extended job searches. If AI quietly closes the door on typical early jobs, people will pay the price in delayed independence, postponed family formation, and the sense that their first serious professional efforts have been refused.

It also matters because entry-level jobs are part of the economy’s training system. Junior analysts learn which numbers can be trusted. Young software developers learn how production systems fail. New marketers learn how customers behave outside the neat language of dashboards. Early-career legal and financial staff learn how rules, judgment, deadlines, and human relationships actually interact. If AI absorbs more of the drafting, triage, coding, summarizing, and administrative preparation that once helped train entry-level workers, firms may become more efficient in the short run while society becomes less capable in the longer run.

The right way to improve the skills of young workers is not to tell them, “Learn to code.” That advice, which shaped more than a decade of federal initiatives and university expansion, rested on the premise that coding was a stable, scalable skill almost anyone could learn and parlay into a middle-class job. The premise no longer holds. The layer of work AI handles well—translating a specification into routine code, reproducing standard patterns, debugging predictable errors—is precisely the layer that “learn to code” programs were built around.

Supervising AI systems in their work is now a much more relevant skill. So understanding the outputs AI systems produce will become very important.

To help people develop such skills, we should require universities, community colleges, and professional programs to embed AI literacy, data literacy, prompt-based workflow skills, verification skills, and domain judgment into ordinary degrees. Every graduate should know how to use AI tools, check their output, understand their limits, and combine them with human expertise. This matters even for graduates entering occupations that look relatively safe from AI, such as those in health care. Almost every job contains tasks—drafting, summarizing, scheduling, research, basic data work, routine communication—for which AI is already a substantial productivity tool.

The competition most young workers will experience is not human versus machine but colleague versus AI-augmented colleague. For most young workers, the realistic path to making themselves valuable is not to avoid AI but to become fluent in the technology and combine that with domain judgment, contextual reasoning, and human relationship skills. To this end, schools should emphasize paid co-ops, apprenticeships, and employer-linked projects so students build judgment in real workplaces before they graduate.

Governments should also create targeted tax credits, wage subsidies, and training grants for employers that hire early-career workers into structured, AI-augmented roles. The architecture for this kind of conditional, behavior-linked subsidy already exists in US tax policy. What is missing is a version of these instruments built specifically around early-career AI-augmented work.

Firms, for their part, should stop making hiring decisions based only on short-run cost savings from AI. Young workers are not valuable only for the tasks they perform this quarter. Their value lies in learning, skill formation, institutional memory, and future productivity. Entry-level hiring is not just an expense. It is an investment in the future stock of judgment inside the firm. The most effective AI-augmented senior workforce of the late 2030s will be drawn overwhelmingly from the junior cohort of today. Firms that automate away the learning stage may improve their immediate margins but find themselves, a decade from now, without anyone who understands how their own AI-driven workflows actually behave.

Students graduating this spring and next face a tough labor market in transition. AI fluency is becoming a commodity. Domain expertise without AI fluency is being outpaced. The combination is what is genuinely scarce. The mechanical engineer with knowledge of manufacturing and AI proficiency; the software programmer with knowledge of financial services who is also a whiz at AI—these are the types of people who will be in demand.

Georgios Petropoulos is an assistant professor at the USC Marshall School of Business. His research focuses on the implications of information technologies for innovation, competition policy, and labor markets.

During Tuesday’s Google I/O keynote, Demis Hassabis, the CEO of Google DeepMind, proclaimed that we are currently “standing in the foothills of the singularity.” It was a striking statement—the singularity is the theoretical future moment when AI rapidly exceeds human intelligence and dramatically transforms the world. But what struck me as I listened in the audience was the context in which he said those words. 

He was on stage to close out the session with a segment on scientific AI, the centerpiece of which was a video detailing how the company’s weather prediction software provided an advance alert about Hurricane Melissa’s catastrophic landfall in Jamaica last year—and potentially saved lives. If that software, called WeatherNext, helped anyone escape the storm or better fortify their home, that’s an enormous and meaningful achievement. But it’s hardly evidence of an impending singularity.

The juxtaposition of Hassabis’ lofty rhetoric with the real-world results of WeatherNext highlighted the tension between two very different approaches to AI for science. The first focuses on AI tools, like WeatherNext, that are designed and trained to solve specific scientific problems. The second is agentic, LLM-based systems that could one day execute cutting-edge research projects without human involvement.

This second vision powers a great deal of AI enthusiasm right now, including recent excitement around recursive self-improvement, or the idea that AI systems could eventually become the primary drivers of AI advancement—a process that would get faster and faster as the AI systems grow smarter. And agentic systems are now making real research contributions, sometimes with limited human guidance.

Just this week, Pushmeet Kohli, Google Cloud’s chief scientist, published a piece in a special AI and science issue of the journal Daedalus, writing: “We are moving toward AI that doesn’t just facilitate science but begins to do science.” With autonomous AI scientists on the horizon, it’s harder to justify massive efforts to develop super-specialized tools—even one like AlphaFold, for which DeepMind scientists won a Nobel Prize, or a potentially life-saving system like WeatherNext. It also heralds a far stranger future for science, in which humans and AI systems collaborate as peers—or AI even makes scientific progress on its own.

To be clear, Google does not appear to be abandoning its work on specialized AI for science tools. AlphaGenome and AlphaEarth Foundations, which are trained for genetics and Earth science applications respectively, were released last summer, and the newest version of WeatherNext came out in November.

What’s more, such tools remain extremely popular among scientists. Last year, for instance, Google reported that protein structure predictions from AlphaFold have been used by over three million researchers worldwide. And Isomorphic Labs, a Google subsidiary that aims to use AlphaFold and related technologies to develop new drugs, just raised a $2 billion Series B funding round.

But there are concrete signs of realignment, in both enthusiasm and resources. Last month, the Los Angeles Times reported that Google fellow John Jumper, who won the Nobel for AlphaFold, is now working on AI coding, not on science-specific AI tools. It’s not surprising that Google is assigning its best minds to the coding problem, as the company has recently taken a reputational hit because its coding tools don’t currently stand up to those offered by Anthropic and OpenAI. But it may also signal a prioritization of agentic science on Google’s part, as coding abilities are key to the success of some of those systems. 

Across the industry, agentic researcher systems are showing real potential. This week, OpenAI announced that one of their models had disproved an important mathematics conjecture—perhaps the most meaningful contribution that generative AI has made to mathematics so far, according to some mathematicians.

Importantly, the model used by OpenAI is not specialized for solving mathematical problems, or even for research; according to the company, it’s a general-purpose reasoning model in the vein of GPT-5.5. If general agents can make independent contributions to mathematical research, they might soon be able to do the same in science (though the fact that ideas in science must be verified experimentally makes it a tougher domain for AI).

Google is certainly devoting a lot of attention toward an agent-driven scientific future. The big scientific announcement at I/O was the new Gemini for Science package, which unites several of the company’s LLM-based scientific systems under one brand.

This includes the hypothesis-generating AI Co-Scientist and algorithm-optimizing AlphaEvolve, which are still not publicly available—but as Google is now allowing any researcher to apply for access to Gemini for Science, they may soon see wider adoption in the scientific community. Scientists who were involved in early testing are enthusiastic about their potential: Gary Peltz, a Stanford geneticist, compared using the AI Co-Scientist to “consulting the oracle of Delphi” in a Nature Medicine article.

Gemini for Science isn’t incompatible with specialized tools; to the contrary, agentic systems can be designed to call on such tools when they might be useful. And no agentic system can predict the structure that a protein will fold into without AlphaFold’s help (at least not yet). But the company seems to be shifting its public image—and at least some resources and personnel, such as Jumper—away from specifically developing those kinds of tools. Though it has only been five years since AlphaFold solved the protein-folding problem, both the technology and the discourse have quickly moved beyond that once-revolutionary achievement.

Google has been careful to position this new set of scientific agents as an accelerant for human scientists, rather than a replacement for them—the choice of the name AI Co-Scientist as opposed to AI Scientist, for instance, appears quite deliberate. Hassabis uses that same human-centric framing when he talks about changes in the landscape of scientific AI. “For the next decade or so, we should think about AI as this amazing tool to help scientists,” Hassabis said in an interview published in the Daedalus issue. “Beyond that timeframe, it is hard to say with any certainty, but perhaps these systems will become more like collaborators.”

But no one can be an effective scientific collaborator without also being a skilled scientist in their own right. And if Hassabis is anywhere near the mark when he talks about the “foothills of the singularity,” then AI scientists could eventually exceed the capabilities of their human counterparts.

In a discussion with the journalist Mike Allen at I/O, Hassabis spoke of how he was initially inspired to pursue AI when he observed how progress in physics had stagnated since the 1970s; he wondered whether the human mind had reached its limits in that domain, and if AI could help to overcome that barrier. Superhuman agentic scientists would certainly fit that bill. We might not ever get anywhere near there, but Google seems to be aiming itself toward that summit.

This Sunday, a group of 42 athletes will gather in Las Vegas to compete in a somewhat unusual sporting competition. Participants in the inaugural Enhanced Games are being encouraged to take performance-enhancing drugs. The goal is to “push the boundaries of human performance.”

The games’ organizers have said that competitors will only be taking substances that have been approved by the US Food and Drug Administration, and that they are all being medically monitored and supervised. But they have also said they expect to see world records broken—and are offering substantial prizes to athletes who succeed in doing so.

As you might expect, the event is generating a mix of curiosity, excitement, and condemnation from various quarters. To me, it feels like very much a reflection of where we are today—an era of peptide-crazed looksmaxxing in which consumers are being encouraged to get thinner than ever, optimize for longevity, and have their “best baby.” It’s 2026, and if you’re not enhancing, what are you even doing?

So, these games. They’ll feature competitions in four categories: swimming, track and field, weightlifting, and strongman (which also involves lifting weights). Many of the competitors already hold national and world records, and some are Olympic medalists. They’ve been paid a salary and will compete for prizes from a $25 million pot. The money has been a major draw for at least some of the athletes.

Another draw is the opportunity to openly experiment with drugs that might boost their performance. In the world of elite sport, every microsecond and every millimeter counts. Athletes—most of whom arguably have genetics on their side already—follow meticulous diet, training, and recovery protocols and wear specially designed gear that allows them to reach for those performance bests.

But within most sporting communities, there are limits. The World Anti-Doping Agency—an international outfit that fights the use of drugs in sports—maintains a lengthy list of “non-approved substances” that are banned in international sporting events. It features many anabolic steroids (which can build muscle), hormones (such as those that stimulate testosterone production or increase the ability of blood to carry oxygen), growth factors (which can stimulate muscle growth and repair, among other things), and more.

Some of these substances have been FDA approved to treat health disorders. And that means they can be used by participants in the Enhanced Games, according to the organization’s rules.

I’ll briefly point out the obvious here—just because a drug has been approved by the FDA doesn’t mean it’s totally safe for everyone and anyone. The risks associated with use of anabolic steroids, for example, include high blood pressure, acne, depression, and liver tumors. Growth hormone use can cause weak muscles, affect vision, and even lead to diabetes.

“Technological doping,” or using improved equipment to gain advantage, has also been supported by the games’ organizers. Last year, participating swimmer Kristian Gkolomeev was reported to have broken a record in a 50-meter freestyle time trial while wearing a polyurethane “super” swimsuit. Such suits have been banned for use in the Olympics since a slew of record-breaking performances in 2008 and 2009. Back then, the swimming governing body ruled that they gave athletes an unfair advantage. But hey, this is the Enhanced Games, where the word “unfair” seems to have a completely different meaning.

Can we expect more records to be broken on Sunday? Maybe. In addition to prize money for winning an event, any athlete who manages to beat a record stands to win up to $1 million, the sum also awarded to Gkolomeev last year following his time trial. But those performances won’t be recognized by official sporting bodies.

Plenty of concerns have been raised about these games. Some argue that they are unsafe and promote risky drug use. Others see them as a “clown show,” and a slap in the face to “clean” athletes who train hard without the use of prohibited drugs. World Athletics president Sebastian Coe has said that anyone who takes part is “moronic,” and World Aquatics, which oversees international competitions in water sports, has banned Enhanced Games participants from its events and activities.

But. The games—and the participating athletes—will still get a huge amount of attention. As a result, so will performance-enhancing drugs. Enhanced, the company behind the games, also runs an online store. There, you can buy a $52 T-shirt emblazoned with the message “I am Enhanced.”

There is also a range of prescription drugs on offer, including peptides “to support recovery, vitality, and longevity.” One of these is a growth hormone that the FDA approved in 1997 for the treatment of children with “growth failure.” The compounded version offered on the Enhanced website, which is not FDA approved, is marketed for longevity, supporting deep sleep and “overall wellness and vitality.” (“Marketed” is the key word here. The drug has, again, not been approved for that purpose.)

It all fits very well with the zeitgeist. Sure, we don’t yet have any drugs that are designed to extend human lifespan. But the search for anti-aging drugs is getting more attention—and funding—than ever. People, particularly women, are seemingly not allowed to visibly age anymore—we have filters and facelifts for that now. The idea that “death is wrong” is gaining acceptance.

And self-experimentation is rife. “Biohacking” was shortlisted for Collins Dictionary’s Word of the Year in 2025. Peptides are everywhere, despite all the unknowns surrounding their safety and effectiveness. So are longevity clinics, despite the fact that most are selling unproven treatments. US states like Montana are making it easier for people to get hold of unapproved “therapies.”

Companies are even offering would-be parents the option to choose the potential future children expected to live longest. Yep—you can supposedly optimize your embryos now, too.

In this climate, the Enhanced Games don’t feel so radical. They feel entirely fitting for our era of questionable optimization despite the risks —an era when, apparently, being human is no longer enough.

The vibes were strong at Code with Claude, Anthropic’s two-day event for software developers in London that kicked off on May 19, the same day as Google’s I/O in Palo Alto. (A coincidence, not a flex, Anthropic staffers assured me.)

“Who here has shipped a pull request in the last week that was completely written by Claude?” Jeremy Hadfield, an engineer at Anthropic, asked from the main stage. Almost half the people in the packed room—many sitting with laptops on their knees, coding or prompting as they watched the talks—raised their hands.

Pull requests are fixes or updates to existing software that are submitted for review before they go live. They are the bread and butter of software development, the chunks of code that most professional developers spend their lives writing—or did until now.

“Who here has shipped a pull request that was completely written by Claude where they did not read the code at all?” Hadfield asked next. Nervous laughter. Most of the hands stayed up.

It’s not news that LLM-powered tools like Anthropic’s Claude Code and OpenAI’s Codex have upended the way software gets made. Top tech companies now like to boast of how little code their developers write by hand. (“Most software at Anthropic is now written by Claude,” Hadfield said. “Claude has written most of the code in Claude Code.”) OpenAI, Google, and Microsoft make similar claims. Many others wish they could.

Even so, it is striking how normal this new paradigm already seems, and how fast it has set in. This was the second year that Anthropic has put on developer events, which also run in San Francisco and Tokyo. This time last year, the company had just released Claude 4. It could code, kind of. But with Anthropic’s latest string of updates—especially Claude 4.6 and then 4.7, released in February and April—Claude Code is a tool that more and more developers seem happy to hand their work off to.   

ANTHROPIC (GRAPHIC) / WILL DOUGLAS HEAVEN (PHOTO)

Anthropic says its goal is to push automation as far as it will go. Instead of using AI to generate code and then having humans clean it up and fix the mistakes, it wants Claude to check and correct its own work. “The default isn’t ‘I’m going to prompt Claude’—the default is now ‘I’m going to have Claude prompt itself,’” Boris Cherny, who heads Claude Code, said in the opening keynote.

If all goes well, human developers shouldn’t even see the error messages when something doesn’t work. That will all be handled by Claude, which will test and tweak, test and tweak, until everything runs as it should. As Ravi Trivedi, an engineer at Anthropic, put it in another talk: “The key principle is getting out of Claude’s way. We like to say: ‘Let it cook.’”

Trivedi presented a new feature in Claude Managed Agents, Anthropic’s cloud-based setup for building and running multi-agent systems, announced two weeks ago, which the company calls dreaming. Claude agents write notes to themselves, recording and saving useful information about specific tasks. When another coding agent, say, starts to work on the same code that others have worked on, it can use the notes they left behind to get up to speed faster and learn from any errors those previous agents may have made.

Dreaming is a system that Claude agents can use to read through the notes and consolidate the information they contain, spotting patterns and common issues across different tasks. In theory, dreaming should help coding agents learn about a particular code base and get better and better at working on it.

Success stories

Code with Claude is an event aimed at developers. As well as product showcases and hands-on workshops from Anthropic, there were how-tos from a range of companies that have reshaped their software development teams around Claude Code, including Spotify and Delivery Hero as well as Lovable, Base44, and Monday.com—three startups vibe-coding apps that help people vibe-code apps.

There were no signs of unease at Code with Claude. Everybody I met wanted in.

And yet outside the conference there have been a number of reports that many coders are starting to question this bright new future. Some gripe in online forums like Reddit and Hacker News that AI coding tools are being pushed by managers chasing productivity gains, when in practice the technology makes software development harder because of all the extra code developers now have to review. “The only people I’ve heard saying that generated code is fine are those who don’t read it,” a user called pron posted on Hacker News last week. 

Others claim that their coding abilities have fallen off as they hand more tasks to AI. And researchers have warned that AI tools can produce unsafe code that will make software more vulnerable to attacks.  

I sat down with Claude engineering lead Katelyn Lesse and Claude product lead Angela Jiang and asked them what they made of the concerns that a sudden flood of code generated (and shipped) without proper human oversight was kicking serious security and maintenance problems down the road.

“All of the old software development best practices still apply. They’ve applied this entire time,” said Lesse. “I think there are a lot of people and teams that may have lost sight of them in this moment.” 

And yet as Anthropic and others push for greater automation and tools like Claude Code improve, the temptation increases to offload more and more tasks, including oversight. Lesse told me that some of the technical managers at Anthropic are exhausted by keeping up with all the code their teams now produce. “Part of things happening so much more quickly is just managing your time,” she said.

“I think that right now Claude is probably as good as a midlevel engineer at writing code,” she added. You still need expert engineers to design a system and troubleshoot harder problems, she said. “But over time we want Claude to get better and better at all different types of engineering.”

Jiang agreed: “I think the absolute end state we’re trying to get to is Claude basically being able to build itself.”

Correction: Dreaming is a feature of Claude Managed Agents not Claude Code. The article has been updated.

Is it the Department of Defense or the Department of War? The Gulf of Mexico or the Gulf of America? A vaccine—or an “individualized neoantigen treatment”?

That’s the Trump-era vocabulary paradox facing Moderna, the covid-19 shot maker whose plans for next-generation mRNA vaccines against flus and emerging pathogens have been dashed by vaccine skeptics in the federal government. Canceled contracts and unfriendly regulators have pushed the Massachusetts-based biotech firm to a breaking point. Last year, Robert F. Kennedy Jr., head of the Department of Health and Human Services, zeroed in on mRNA, unwinding support for dozens of projects—including a $776 million award to Moderna for a bird flu vaccine. By January, the company was warning it might have to stop late-stage programs to develop vaccines against infections altogether.

That raises the stakes for a second area of Moderna’s research. In a partnership with Merck, it’s been using its mRNA technology to destroy tumors through a very, very promising technique known as a cancer vacc—

“It’s not a vaccine,” a spokesperson for Merck jumped in before the V-word could leave my mouth. “It’s an individualized neoantigen therapy.”

Oh, but it is a vaccine. And here’s how it works. Moderna sequences a patient’s cancer cells to find the ugliest, most peculiar molecules on their surface. Then it packages the genetic code for those same molecules, called neoantigens, into a shot. The patient’s immune system has its orders: Kill any cells with those yucky surface markers.

Mechanistically, it’s similar to the covid-19 vaccines. What’s different, of course, is that the patient is being immunized against a cancer, not a virus.

And it looks like a possible breakthrough. This year, Moderna and Merck showed that such shots halved the chance that patients with the deadliest form of skin cancer would die from a recurrence after surgery.

In its formal communications, like regulatory filings, Moderna hasn’t called the shot a cancer vaccine since 2023. That’s when it partnered up with Merck and rebranded the tech as individualized neoantigen therapy, or INT. Moderna’s CEO said at the time that the renaming was to “better describe the goal of the program.” (BioNTech, the European vaccine maker that’s also working in cancer, has shifted its language too, moving from “neoantigen vaccine” in 2021 to “mRNA cancer immunotherapies” in its latest report.)

The logic of casting it as a therapy is that patients already have cancer—so it’s a treatment as opposed to a preventive measure. But it’s no secret what the other goal is: to distance important innovation from vaccine fearmongering, which has been inflamed by high-ranking US officials. “Vaccines are maybe a dirty word nowadays, but we still believe in the science and harnessing our immune system to not only fight infections, but hopefully to also fight … cancers,” Kyle Holen, head of Moderna’s cancer program, said last summer during BIO 2025, a big biotech event in Boston.

Not everyone is happy with the word games. Take Ryan Sullivan, a physician at Massachusetts General Hospital who has enrolled patients in Moderna’s trials. He says the change raises questions over whether trial volunteers are being properly informed. “There is some concern that there will be patients who decline to treat their cancer because it is a vaccine,” Sullivan told me. “But I also felt it was important, as many of my colleagues did, that you have to call it what it is.”

But is it worth going to the mat for a word? Lillian Siu, a medical oncologist at the Princess Margaret Cancer Centre, in Toronto, who has played a role in safety testing for the new shots, watches US politics from a distance. She believes name change is acceptable “if it allows the research to continue.”

Holen told me the doctors complaining to Moderna were basically motivated by a desire to defend vaccines—which are, of course, among the greatest public health interventions of all time. They wanted the company to stand strong. 

But that’s not what’s happening. When Moderna’s latest results were published in February, the paper’s main text didn’t use the word “vaccine” at all. It was only in the footnotes that you could see the term—in the titles of old papers and patents.

All this could be a sign that Kennedy’s strategy is working. His agencies often appear to make mRNA vaccines a focus of people’s worries, impede their reach, devalue them for companies, and sideline their defenders. 

Still, Moderna’s strategy may be working too. So far, at least, the government hasn’t had much to say about the company’s cancer vacc— I mean, its individualized neoantigen therapy.

This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

I.

We had crash-landed on the planet. We were far from home. The spaceship could not be repaired, and the rescue beacon had failed. Besides me, only the astrogator, part of the captain, and the ship’s AI mind were left. 

Outside, the atmosphere registered as hostile to most organisms. We huddled in the lifeboat, which was inoperable but still held air. Vast storms buffeted our cockleshell shelter, although we knew from prior readings that other areas remained calm. All that remained to us was to explore, if we wanted to live. The captain gave me the sole weapon. She tasked the astrogator with carrying some tools that would not unduly weigh him down.

Little existed on the planet except deserts of snow. But alien artifacts lay in an area near us. We were an exploration team, so this discovery had oddly comforted us, even though we had been on our way elsewhere. The massive systems failure had no discernible source, and the planet had been our only choice for landfall.

The artifacts took the form of 13 domes, spread out over that hostile terrain. The domes had been linked by cables just below shoulder level, threaded through the tops of metal posts at irregular intervals. Whether intended or not, these cables and rods formed a series of paths between the domes. 

Before our instruments failed, the AI had reported that the domes appeared to have a heat signature. The cables pulsed under our grip in a way that teased promised warmth far ahead. It took some time to get used to the feeling.

The shortest path between domes was a thousand miles long. The longest path was 10 thousand miles long. Our suit technology was good: A suit could recycle water, generate food, create oxygen. It could push us into various states of near hibernation while motors in the legs drove us forward. For the captain, the suit would compensate for having lost her legs and ease her pain. We estimated we could reach the nearest path and follow it to the nearest dome … and that was it. If the dome had life support capabilities, or even just a way to replenish our suits, we would live. Otherwise, we would probably die.

We revised the estimate of our survival downward when we reached the path and soon encountered the skeletons of dead astronauts littering the way. In all shapes and sizes, cocooned within their suits. Their huddled forms under the snow displayed a serenity at odds with their fate. But when I wiped the frost from face plates, we saw the extremity of their suffering.

It is difficult to explain how we felt walking among so many fatalities. So many dead first contacts. 

We no longer had to puzzle over the systems failure. Spaceships came here to crash, and intelligent entities came here to die, for whatever reason. We could not presume our fate would be any different, and adjusted our expectations accordingly. The AI’s platitudes about courage did not raise morale. There were too many lost there in the frozen wastes. 

Here were the ghastly emissaries of hundreds of spacefaring species we had never before encountered.

The number of the bodies and their haphazard positioning hampered our ability to make progress to the dome. The AI estimated our chances of survival at below 50% for the first time. We would starve in our suits as the motors propelled us forward. We would become desiccated and exist in an elongation of our thoughts that made us weak and stupid until the light winked out. But still, we had no choice. So even in places where the dead in their suits were piled high, we would simply plunge forward, over and through them, headed for the dome. 

What we would find there, as I have said, we did not know. But we were in an area of the galaxy where ancient civilizations had died out millions of years ago. We had been on our way to a major site, an ancient city on a moon with no atmosphere in a wilderness of stars. 

Although our emotions fluctuated, a professional awe and curiosity about the dead eventually came over us. This created much debate over the comms. We had made a discovery for the ages, but our satisfaction was bittersweet. Even if we lived longer than expected, we would never return home, never see our friends or family again. The AI might continue on after we were dead, but I doubt it envied being the one to report on our discovery centuries hence. And to who?

Here were the ghastly emissaries of hundreds of spacefaring species we had never before encountered. Their suits displayed an extraordinary range, although our examination was cursory. Some even appeared to be made out of scales and other biological substances from their home worlds, giving us further clues as to their origins. 

The burial of the suits by snow and the lack of access to anything other than a screaming face or faces, often distorted by time and ice, worked against recording much usable data. This issue was compounded in those cases where the suit was part of the organism and they had not needed any “artificial skin,” as the AI put it, to survive harsh conditions. That many had died despite appearing well-­prepared for the planet’s environment sobered us up even before our own suits dispensed drugs to help our mental states. 

After a time, each face seemed to express some aspect of our own stress and terror at the seriousness of our situation. After a time, the sheer welter of detail defeated us and caused us extreme distress. The captain made the observation that even one instance of alien contact might cause physiological and mental conditions, including anxiety, stress, fatigue. Here, we were constantly encountering the alien dead of what seemed at times an infinite number of civilizations. 

We stopped recording. We recommitted ourselves to the slog toward the nearest dome. 

The captain’s drugs unit had failed, but the AI found a way to help her by turning off the heating element in select panels of her suit. Some parts of her would soon be lost to the cold, but the system would allow her to live on with some measure of comfort.

I must admit, we were just glad the screaming had stopped and welcomed her counsel.

II.

For a long time, as we labored in our spacesuits on that planet—following the path, beleaguered by snowstorms—we could not understand why we found so many dead astronauts, of so many unknown alien types, and yet no spaceships. During good visibility, our line of sight reached, unbroken, for 500 miles. Where were the crash sites? 

But one day we chanced upon an antenna sticking up out of the ground. Clumsy attempts at excavation soon revealed that below this antenna lay a vast dead spaceship of a kind we had never seen before. The gash that had opened it to the elements had laid bare its unique architecture, but also gave the illusion that the snow had spilled out of it to create the world around us rather than having infiltrated and accumulated inside over time.

Aspects of the spaceship’s texture gave the startling suggestion that it had been made of some ultra-hard wood or wood equivalent. Clambering partway up to stare at the inner compartments, we all felt the strangeness of the dimensions and proportions of the living quarters. There was no sign of the occupants. Perhaps, I suggested, they had headed for the domes. Perhaps they had even made it to the domes. I tried and failed to keep hope from my voice.

But the captain had ordered the AI to perform a materials analysis. The “snow” in this region had been contaminated by ash and tiny particles of bone. The AI estimated that more than 70% of the white surrounding us was made of the remains of vertebrate sentient life and the remnants of suits. Of invertebrates there was no telling. A thaw might bring not just the drip, drip of water but a shushing sound indicative of bone particulate in the mixture. I imagined there might even be the clink of small objects not rendered down by whatever intense heat had created the ash.

The astrogator had insisted on digging deeper into the ship, with the idea that some recognizable commonality between technologies might yield a part or parts with which he could fix our ship. The rest of us allowed this delusion for the obvious reasons. But upon his return, he held in his hands ovals of snow not much larger than the space formed by the circle between a thumb and finger. Many of them had soft indentations, as one might find in the afterbirth of reptiles from eggs. A kind of ghostly cilia-like tread appeared along the bottoms of these objects.

The astrogator did not find any technology of use to us. Instead, he discovered that the species piloting the spaceship had been so different from us as to be safely encapsuled in suits the size of eggs. Much of what had spilled into or spilled out of the gash constituted the bodies of the crew, in their hundreds of thousands. Their suits had been inadequate to the conditions. They had died en masse attempting to escape their own ship.

The AI speculated that it had been a generation ship, perhaps fleeing a planet with a dying star. If we wondered how the AI had reached this conclusion, it was because we did not want it to be true.

The captain became silent upon receiving this further news and did not speak to us for more than 100 miles of further progress. 

As we left that site, unsure exactly what we stepped upon, we also knew that since the spaceship was entirely covered by snow, it had been falling into the sediment for days or months or years. We knew then that our ship might not be visible against the horizon should we retrace our steps. The already bleak probability of rescue through visual identification of a crash site from above would be lost to us in time, even as the line of cables remained perpetually visible to the horizon. We now thought of the planet as a trap. But of what sort? 

III.

We could not be sure, but in the absence of the captain’s voice, it may have been the AI that put forward the idea of the planet’s being “duplicitous.” The phrasing concerned us, for there was a duplicity in using the planet as the subject of the spoken sentence. A sphere rotating around a sun in deep space could not exhibit forethought or premeditation or other qualities of sentience. 

The AI meant whoever or whatever had created the conditions on the planet that allowed spacecraft to be trapped and then the occupants placed in a perilous situation with no recourse. But I distinctly recall the AI using the words “the planet.” In addition to being inaccurate, this also let us know that the AI did not have any analysis available that might help us understand the agency and motivations acting upon us. 

But in a sense, the AI only voiced something I had felt for several miles: that there existed an overlay to the planet’s surface, an area or space or different landscape unavailable to us. This overlay had also not been available to any of the prior astronauts who had died here. In this area or space or different landscape existed a wealth of the usual hoped-for things: a breathable atmosphere and abundant food and water. 

While we struggled with the line through the snow and through the storms that welled up, others could see us but chose to ignore us for reasons or perhaps just for their own well-being. For hundreds, possibly thousands of years, as explorers had died here in merciless and terrible ways, there raged a sumptuous feast for the senses, as excessive as it was ancient and unending.

I cannot tell you how powerfully the AI’s words struck us, so that our mouths watered at the thought of real food and of clean, unrecycled water, of a freedom unencumbered by suits and breathing apparatus. Even at our intended destination, we would have spent most of our days aboard a small space station. This tedium would have been broken only by the arduous process of reaching the unbreathable surface and its ancient ruins of jagged black stone. 

This vision that overtook us functioned not just as tantalizing delusion. It scared us so much that we could not compartmentalize it in our thoughts. It continued to overwhelm us like a wave.

We fought for the first time, with the astrogator expressing the wish to return to the ruined spacecraft and explore nearby areas for parts, while the captain broke silence to order us to continue to make progress toward the nearest dome. The AI, which had brought us to this point, stole the captain’s silence and said no more.

For each of us, those endless white plains with no real elevation, just the metal rope and the metal posts, had become a kind of repetition that hurt the brain, and the mind with it.

IV.

About halfway to the nearest dome, amid a storm that reduced our progress incrementally and our line of sight to nothing, we came upon a peculiar tableau. 

Six astronaut suits had fallen across and around the metal rope. With the flurries of snow, it took us, even with our powerful headlamps, some minutes to determine the nature of the obstruction. The six suits had been created for a humanoid species that must have had torsos like nine-foot-long slabs, attached to six limbs, three for walking. Their heads had flared out like thick fans. All the helmets were cracked open, and curled inside were the skeletons of some other intelligent species no larger than 40 or 50 pounds, possibly warm-blooded. With no sign of the original occupants. 

After a brief analysis cut short by the conditions, we postulated that the warm-blooded species had worn breathable skin suits that, as they failed, required these intruders to seek shelter. All they could find were these six dead astronauts. Because we could discover no trace of the original occupants, the AI put forward the theory that this smaller species had eaten every scrap of the remains within the suits. 

Then they too had perished, and in time, the AI suggested, something smaller would take up residence inside those bodies, then smaller still within those, and smaller still—

At this point, the captain attempted a soft reboot of the AI using a coded question. We could hear the concern in her voice.

Yet the AI continued undeterred, suggesting that we might find this to be a common situation. It might be replicated across the planet, depending on a system’s ability to break down and process meat that had not evolved alongside the devourer for millions of years. In all likelihood, most who attempted to eat in this way died soon after, poisoned by alien flesh.

V.

The AI began to talk to us in strange alien voices at mile 700, as we labored through the snowstorm to hold onto the cables and thus the path. The AI warbled and chirped and howled and hummed and clucked. The AI spoke in voices like fossilized choruses of beasts, vast and harmonious. And in voices like dry grass spun to fire by the sun. And in voices like the dissolution of all things, darkness in the blinding white that scared me. 

At first we thought the AI was deranged. Then that the AI channeled voices from the dome 300 miles ahead. But finally, the AI managed to make known to us that these were the voices of the dead astronauts we had come across from time to time. Huddled frozen. The suits in so many shapes and sizes. That the voices of the dead were channeled through the AI, and nothing could stop them.

VI.

The snow began to betray us, as the storms created different forms of ice. Often, our arms became weary, our legs cramping, and we had to rest with greater frequency. We came to accept the solid crunch that could support our weight. We came to reject the feather-light freshness that felt effortless underfoot but could give way just as easily as if it were air. In some places, slick purple-hued ice welled up in sluggish layers as if something half-alive. In others, we discovered strange islands of elevation, with brutal curls and curves that suggested two continental shelves had clashed in that space.

As we adapted to these conditions, and as conditions worsened and still we adapted, we came to feel an illusion of competency, one that made even the astrogator temporarily cheerful. The sounds through the comms of our efforts, the deeper breathing, the occasional muffled curse, seduced us in this regard. We felt that we were becoming adroit at handling the snow. We began to believe if we could only make it to the dome, we would be saved.

Yet this uptick in morale ran parallel to, rather than intersected with, the idea of our ultimate survival.

VII.

We lost track of the distance left to us without the AI to tell us. Or the captain, in her pain, no longer thought to issue updates. But across the distance left to us came sights beyond reckoning: three giant astronauts spaced 50 miles apart. Larger than most starships, each body lay sprawled across an area larger than several fields and in very different conditions.

The first had been badly burned and was thus unrecoverable, even in terms of salvage. The astronaut had crawled or pulled itself along for some distance. It had left a long smudge of black and red across that expanse. The alien species was, as ever, unknown to us, but the five arms were sunk in the ground as if in agony. The skull had once held three eyes, and the face plate had been cracked by force so strong it resembled a meteor strike. The body was bloated, the fabric of the suit gray with a shimmer of green that came and went, linked to photosensitive skin cells. The way the flesh took up space, and how it exhibited aspects more plant than animal, made it impossible to study further.

The second was a sprawl of limbs, with the suggestion of a defensive posture. The debris of conflict flared out to the side in an incomprehensible display. The suit had an intactness that surprised us, but a similar crack in the face plate without any trace of body within. The rest of the suit had become inhabited by a wealth of other dead astronauts of varying sizes and shapes, who had sought shelter or sustenance and then become trapped or simply … given up. As the AI had predicted, we had once again encountered bodies providing other bodies with temporary sustenance and shelter.

I felt like a parasite who beheld a god. Or was the scale even more ludicrous?

But this condition was not at first evident to us, becoming apparent only after we had clambered for an hour to reach the cracked face plate and the entry hole extended like a broken archway before us.

Despite the number of remains within, and the difficulty in moving through them to explore, the captain ordered an exhaustive recon. Her pulse in the readings had a thready quality. Sometimes I felt, and the astrogator too when we took private comms, that the captain had begun to say things similar to the AI’s delusions. Yet we obeyed the order, on the chance that some internal calculation on the captain’s part meant she believed this was the only way we would survive. 

What did we expect to find in the dead body of a once-­intelligent giant? Food? Oxygen? Some cause of death? To put off the thought of our own death by seeking shelter with a death so large we could not comprehend it?

I felt like a parasite who beheld a god. Or was the scale even more ludicrous? I had trouble envisioning the way the body must have twisted as it pitched forward into that icy ground. I had trouble holding onto my own thoughts.

More and more pressure moved through my skull as I contemplated that scene. We were in the midst of something none of my kind had ever known. We might be the only ones, ever. I better understood the unraveling of the AI and of the captain. My sharpness had dulled, taking my calm with it.

It was impossible to tell how long the astronaut had taken to die. Unless somewhere within that fallen figure some hint of life hid that we would never find.

The storms fell away, rose, then fell away again. 

VIII.

The third huge astronaut was full of light and life and shone out across the wasteland of snow like a beacon. For a moment, I thought we had pierced the invisible layer and could see what lay beyond the veil. We would have comforts beyond anything found on our ruined spaceship even when it had been fit to cross galactic space. There would not be recycled urine for our water. There would not be the faint stink of sweat creeping into our suits as the ventilation system began to fail. Our liquid food would not taste stale and moldy. 

As we approached, the suit extended almost to the horizon in that foreshortened perspective created by the left foot. We noted through our remaining instrumentation that the suit remained intact. The pressure told us a kind of air circulated within its sealed surfaces. 

We climbed with a renewed energy, the promise of sanctuary so close making us giddy. We each exhorted the others on with such exuberance that it made me a little afraid. What lay on the other side of this state of mind but a fall?

When we reached the helmet plate, we could see inside not a face or a skull, but instead such a richness of healthy growth that we fell silent before it. None of us could, I believe, understand exactly what we saw, except that it equaled ecosystem—resplendent with vibrant greens and blues, stippled with other colors. There might be some parallel to a terrarium full of moss and exotic plants. There might be some sense of life moving amongst those plants, as of jewel-like amphibians or even tiny shy sapphire birds. We could not smell or taste or hear what lay behind the face plate. We could not experience it in that way, but somehow we each imagined enough to be calmed and comforted by it. 

The astrogator said he might be able to create a hole in the plate or elsewhere on the body to let us in, and then patch the surface such that not too much air or vitality would spill out. This workaround might take an hour or two, due to the delicate nature of what we saw within. But it was possible.

The captain considered the astrogator’s proposal and then agreed. The weather had begun to turn dangerous again. That we should begin immediately did not need to be said. With the proper pressure brought to bear, we would have some measure of sanctuary from which to recover for a final push to the dome. It could be the difference between life and death, the astrogator said. If the atmosphere was breathable, we might even be able to give the captain some better solution to her pain.

I unclipped the astrogator’s equipment from his waist and threw it off the mountain that was the astronaut and watched it sail through the air and into the snow. Then I used my weapon to fry it where it lay. Then I threw my weapon into the snow, too, in a place where the featheriness would cover it and hide it forever. 

We were a team and I had helped my team while showing them I posed no threat—although I knew the astrogator and the captain would not see it that way. I stood there on the face plate that we could no longer open with the diminished tools at our disposal as they both yelled at me through the comms. It’s unimportant what they said to me. They were admonishing me for something that had already happened and that they had no power to stop. I did not bother to explain, but began to make the descent to the ground so we could once again take up the metal rope and make for the dome.

That in that frozen hellscape, the persistence of life in that manner, an oasis in the midst of nothing, could be categorized as a miracle.

ROGAN BROWN

This reply incensed him and he castigated me in words not used among members of a team that respect each other. Once more, I ignored him, but told him if our oxygen got low, he could have mine if we calculated he could make it to the base. I meant this, as I knew the odds were low anyway. I had hurt my knee taking the equipment from the astrogator and then making my way so rapidly down from the dead astronaut.

The astrogator did not reply, by which I knew he did not accept my answer.

The reason I took the tools and destroyed them is because the wind had told me something it had not whispered to the captain or the astrogator. The wind had not spoken to me before, so I believed what it told me. That the astronaut within the suit lived on, if unable to move. That what we saw on the outside and registered as ecosystem, as separate “plants” and “animals,” instead formed a composite life-form and that to crack open the suit or cut through the suit at a leg would have been a violation.

That in that frozen hellscape, the persistence of life in that manner, an oasis in the midst of nothing, could be categorized as a miracle. 

I would not snuff that out. I could not allow that to be snuffed out. But I remembered too how I felt looking at that vast and alien country behind the face plate. So calm, so comforted, overcome by the depths of an emotion I could not place. Would I replace that feeling with the feeling of seeing all those explorers dead within the other vast suit? Even as I become one of them? 

Because the planet had already told us the rules, the consequences, and the ultimate outcome. There are no odds so terrible that they could not be experienced, and in dozens of ways, in this place. 

So I trudged on and the astrogator cursed me and cursed me and called out my childhood and how badly I must have been brought up and how I must have cheated to pass the psych exams, and yet I had thought the same of him at various points during our journey.

See how beautiful the snow is, falling now, I said to him over the comms. See how precise and geometric this line we follow across this expanse. 

He did not reply, but a little later he told me he no longer believed in the line at all, and by his calculations he would get to the dome faster if he abandoned it and struck out on his own.

I could not stop the astrogator and did not want to, so I watched him become a smaller and smaller figure against the white until the white ate him up and I was alone.

X.

I have been walking a long time, visiting with the dead. Here, against an arch of heaven that appears no different than what I see directly in front of me. 

Jeff VanderMeer is the author of the critically acclaimed, bestselling Southern Reach series, translated into 38 languages. His short fiction has appeared in Vulture, Slate, New York Magazine, Black Clock, Interzone, American Fantastic Tales (Library of America), and many others.

A rare warm spell in January melted enough snow to uncover Cornell University’s newest athletic field, built for field hockey. Months before, it was a meadow teeming with birds and bugs; now it’s more than an acre of synthetic turf roughly the color of the felt on a pool table, almost digital in its saturation. The day I walked up the hill from a nearby creek to take a look, the metal fence around the field was locked, but someone had left a hallway-size piece of the new simulated grass outside the perimeter. It was bristly and tough, but springy and squeaky under my booted feet. I could imagine running around on it, but it would definitely take some getting used to.

My companion on this walk seemed even less favorably disposed to the thought. Yayoi Koizumi, a local environmental advocate, has been fighting synthetic-turf projects at Cornell since 2023. A petite woman dressed that day in a faded plum coat over a teal vest, with a scarf the colors of salmon, slate, and sunflowers, Koizumi compulsively picked up plastic trash as we walked: a red Solo cup, a polyethylene Dunkin’ container, a five-foot vinyl panel. She couldn’t bear to leave this stuff behind to fragment into microplastic bits—as she believes the new field will. “They’ve covered the living ground in plastic,” she said. “It’s really maddening.” 

The new pitch is one part of a $70 million plan to build more recreational space at the university. As of this spring, Cornell plans to install something like a quarter million square feet of synthetic grass—what people have colloquially called “astroturf” since the middle of the last century. University PR says it will be an important part of a “health-promoting campus” that is “supportive of holistic individual, social, and ecological well-being.” Koizumi runs an anti-plastic environmental group called Zero Waste Ithaca, which says that’s mostly nonsense.

This fight is more than just the usual town-versus-gown tension. Synthetic turf used to be the stuff of professional sports arenas and maybe a suburban yard or two; today communities across the United States are debating whether to lay it down on playgrounds, parks, and dog runs. Proponents say it’s cheaper and hardier than grass, requiring less water, fertilizer, and maintenance—and that it offers a uniform surface for more hours and more days of the year than grass fields, a competitive advantage for athletes and schools hoping for a more robust athletic program.

But while new generations of synthetic turf look and feel better than that mid-century stuff, it’s still just plastic. Some evidence suggests it sheds bits that endanger users and the environment, and that it contains PFAS “forever chemicals”—per- and polyfluoroalkyl substances, which are linked to a host of health issues. The padding within the plastic grass is usually made from shredded tires, which might also pose health risks. And plastic fields need to be replaced about once a decade, creating lots of waste.

Yet people are buying a lot of the stuff. In 2001, Americans installed just over 7 million square meters of synthetic turf, just shy of 11,000 metric tons. By 2024, that number was 79 million square meters—enough to carpet all of Manhattan and then some, almost 120,000 metric tons. Synthetic turf covers 20,000 athletic fields and tens of thousands of parks, playgrounds, and backyards. And the US is just 20% of the global market. 

Where real estate is limited and demand for athletic facilities is high, artificial turf is tempting. “It all comes down to land and demand.”

Frank Rossi, professor of turf science, Cornell

Those increases worry folks who study microplastics and environmental pollution. Any actual risk is hard to parse; the plastic-making industry insists that synthetic fields are safe if properly installed, but lots of researchers think that isn’t so. “They’re very expensive, they contain toxic chemicals, and they put kids at unnecessary risk,” says Philip Landrigan, a Boston College epidemiologist who has studied environmental toxins like lead and microplastics.

But at Cornell, where real estate is limited and demand for athletic facilities is high, synthetic turf was a tempting option. As Frank Rossi, a professor of turf science at Cornell, told me: “It all comes down to land and demand.”

In 1965, Houston’s new, domed base­ball stadium was an icon of space-age design. But the Astrodome had a problem: the sun. Deep in the heart of Texas, it shined brightly through the Astrodome’s skylights—so much so that players kept missing fly balls. So the club painted over the skylights. Denied sunlight, the grass in the outfield withered and died.

A replacement was already in the works. In the late 1950s a Ford Foundation–funded educational laboratory determined that a soft, grasslike surface material would give city kids more places to play outside and had prevailed upon the Monsanto corporation to invent one. The result was clipped blades of nylon stuck to a rubber base, which the company called ChemGrass. Down it went into Houston’s outfield, where it got a new, buzzier name: AstroTurf.

AP PHOTO/ED KOLENOVSKY, FILE

That first generation of simulated lawn was brittle and hard, but quality has improved. Today, there are a few competing products, but they’re all made by extruding a petroleum-based polymer—that’s plastic—through tiny holes and then stitching or fusing the resulting fibers to a carpetlike bottom. That gets attached to some kind of padding, also plastic. In the 1970s the industry started layering that over infill, usually sand; by the 1990s, “third generation” synthetic turf had switched to softer fibers made of polyethylene. Beneath that, they added infill that combined sand and a soft, cheap shredded rubber made from discarded automobile tires, which pile up by the hundreds of millions every year. This “crumb rubber” provides padding and fills spaces between the blades and the backing.

In the early 1980s, nearly half the professional baseball and football fields in the US had synthetic turf. But many players didn’t like it. It got hotter than real grass, gave the ball different action, and seemed to be increasing the rate of injuries among athletes. Since the 1990s, most pro sports have shifted back toward grass—water and maintenance costs pale in comparison to the importance of keeping players happy or sparing them the risk of injury. 

But at the same time, more universities and high schools are buying the artificial stuff. The advantages are clear, especially in places where it rains either too much or not enough. A natural-grass field is usable for a little more than 800 hours a year at the most, spread across just eight months in the cooler, wetter northern US. An artificial-turf field can see 3,000 hours of activity per year. For sports like lacrosse, which begins in late winter, this makes artificial turf more appealing. Most lacrosse pitches are now synthetic. So are almost all field hockey pitches; players like the way the even, springy turf makes the ball bounce.

Furthermore, supporters say synthetic turf needs less maintenance than grass, saving money and resources. That’s not always true; workers still have to decompact the playing surface and hose it off to remove bird poop or cool it down. Sometimes the infill needs topping up. But real grass allows less playing time, and because grass athletic fields often need to be rotated to avoid damage, synthetic ground cover can require less space. Hence the market’s explosive growth in the 21st century.

The city and town of Ithaca—two separate political entities with overlapping jurisdiction over Cornell construction projects—held multiple public meetings about the university’s new synthetic fields: the field hockey pitch and a complex called the Meinig Fieldhouse. Koizumi’s group turned up in force, and a few folks who worked at Cornell came to oppose the idea too—submitting pages of citations and studies on the risks of synthetic grass.

At two of those meetings, dozens of Cornell athletes turned out to support the turf. Representatives of the university and the athletic department declined to speak with me for this story, citing an ongoing lawsuit from Zero Waste Ithaca. But before that, Nicki Moore, Cornell’s director of athletics, told a local newspaper that demand from campus groups and sports teams meant the fields were constantly overcrowded. “Activities get bumped later and later, and sometimes varsity teams won’t start practicing until 10 at night, you know?” Moore told the paper. “Availability of all-weather space should normalize scheduling a great deal.”

That argument wasn’t universally convincing. “It’s a bad idea, but that’s from the environmental perspective,” says Marianne Krasny, director of Cornell’s Civic Ecology Lab and one of the speakers at those hearings. “Obviously the athletic department thinks it’s a great idea.”

GETTY IMAGES

Members of Cornell on Fire, a climate action group with members from both the university and the town, joined in opposing the use of artificial turf, citing the fossil-fuel origins of the stuff. They described the nominal support of the project from student athletes as inauthentic, representing not grassroots support but, yes, an astroturf campaign. 

Sorting out the actual science here isn’t simple. Over time, the plastic that synthetic turf is made of sheds bits of itself into the environment. In one study, published in 2023 in the journal Environmental Pollution, researchers found that 15% of the medium-­size and microplastic particles in a river and the Mediterranean Sea outside Barcelona, Spain, came from artificial turf, mostly in the form of tiny green fibers. Back in 2020, the European Chemicals Agency estimated that infill material from artificial-­turf fields in the European Union was contributing 16,000 metric tons of microplastics to the environment each year—38% of all annual microplastic pollution. Most of that came from the crumb rubber infill, which Europe now plans to ban by 2031. 

This pollution worries the Cornell activists. Ithaca is famous for scenic gorges and waterways. The new field hockey pitch is uphill from a local creek that empties into Cayuga Lake, the longest of the Finger Lakes and the source of drinking water for over 40,000 people.

And it’s not just the plastic bits. When newer generations of synthetic turf switched to durable high-density polyethylene, the new material gunked up the extruders used in the manufacturing process. So turf makers started adding fluorinated polymers—a type of PFAS. Some of these environmentally persistent “forever chemicals” cause cancer, disrupt the endocrine system, or lead to other health problems. Research in several different labs has found PFAS in many types of plastic grass.

But the key to assessing the threat here is exposure. Heather Whitehead, an analytical chemist then at the University of Notre Dame, found PFAS in synthetic turf at levels around five parts per billion—but estimated it’d be in water running off the fields at three parts per trillion; for context, the US Environmental Protection Agency’s legal drinking-water limit on one of the most widespread and dangerous PFAS chemicals is four parts per trillion. “These chemicals will wash off in small amounts for long periods of time,” says Graham Peaslee, Whitehead’s advisor and an emeritus nuclear physicist who studies PFAS concentrations. “I think it’s reason enough not to have artificial turf.”

This gets confusing, though. There are over 16,000 different types of PFAS, few have been well studied, and different ­companies use different manufacturing techniques. Companies represented by the Synthetic Turf Council now “use zero intentionally added PFAS,” says Melanie Taylor, the group’s president. “This means that as the field rolls off the assembly line, there are zero PFAS-formulated materials present.”

Some researchers are skeptical of the industry’s assurances. They’re hard to confirm, especially because there are a lot of ways to test for PFAS. The type of synthetic turf going onto the new field hockey pitch at Cornell is called GreenFields TX; the university had a sample tested using an EPA method that looks for 40 different PFAS compounds. It came back negative for all of them. The local activists countered that the test doesn’t detect the specific types they’re most concerned about, and in 2025 they paid for three more tests on newly purchased synthetic turf. Two clearly found fluorine—the F in “PFAS”—and one identified two distinct PFAS compounds. (The company that makes GreenFields TX, TenCate, declined to comment, citing ongoing litigation.)

PFAS isn’t the only potential problem. There’s also the crumb rubber made from tires. A billion tires get thrown out every year worldwide, and if they aren’t recycled they sit in giant piles that make great habitats for rats and mosquitoes; they also occasionally catch fire. Lots of the tires that go into turf are made of styrene-­butadiene rubber, or SBR. In bulk, that’s bad. Butadiene is a carcinogen that causes leukemia, and fumes from styrene can cause nervous system damage. SBR also contains high levels of lead.

But how much of that comes out of synthetic-­turf infill? Again, that’s hotly debated. Researchers around the world have published suggestive studies finding potentially dangerous levels of heavy metals like zinc and lead in synthetic turf, with possible health risks to people using the fields. But a review of many of the relevant studies on turf and crumb rubber from Canada’s National Collaborating Centre for Environmental Health determined that most well-conducted health risk assessments over the last decade found exposures below levels of concern for cancer and certain other diseases. A 2017 report by the European Chemicals Agency—the same people who found all those microplastics in the environment—“found no reason to advise people against playing sports on synthetic turf containing recycled rubber granules as infill material.” And a multiyear study from the EPA, published in 2024, found much the same thing—although the researchers said that levels of certain synthetic chemicals were elevated inside places that used indoor artificial turf. They also stressed that the paper was not a risk assessment. 

The problem is, the kinds of cancers these chemicals can cause may take decades to show up. Long-term studies haven’t been done yet. All the evidence available so far is anecdotal—like a series for the Philadelphia Inquirer that linked the deaths of six former Phillies players from a rare type of brain cancer called glioblastoma to years spent playing on PFAS-containing artificial turf. That’d be about three times the usual rate of glioblastoma among adult men, but the report comes with a lot of cautions—small sample size, lots of other potential causes, no way to establish causation.

Synthetic turf has one negative that no one really disputes: It gets very hot in the sun—as hot as 150 °F (66 °C). This can actually burn players, so they often want to avoid using a field on very hot days.

GETTY IMAGES

Athletes playing on artificial turf also have a higher rate of foot and ankle injuries, and elite-level football players seem to be more predisposed to knee injuries on those surfaces. But other studies have found rates of knee and hip injury to be roughly comparable on artificial and natural turf—a point the landscape architect working on the Cornell project made in the information packet the university sent to the city. Athletic departments and city parks departments say that the material’s upsides make it worthwhile, given that there’s no conclusive proof of harm.

Back in Ithaca, Cornell hired an environmental consulting firm called Haley & Aldrich to assess the evidence. The company concluded that none of the university’s proposed installations of artificial turf would have a negative environmental impact. People from Cornell on Fire and Zero Waste Ithaca told me they didn’t trust the firm’s findings; representatives from Haley & Aldrich declined to comment.

Longtime activists say that as global consumption of fossil fuels declines, petrochemical companies are desperate to find other markets. That means plastics. “There’s a big push to shift more petrochemicals into plastic products for an end market,” says Jeff Gearhart, a consumer product researcher at the Ecology Center. “Industry people, with a vested interest in petrochemicals, are looking to expand and build out alternative markets for this stuff.”

All that and more went before the decision-­makers in Ithaca. In September 2024, the City of Ithaca Planning Board unanimously issued a judgment that the Meinig Fieldhouse would not have a significant environmental impact and thus would not need to complete a full environmental impact assessment. Six months later, the town made the same determination for the field hockey pitch.

Zero Waste Ithaca sued in New York’s supreme court, which ruled against the group. Koizumi and lawyers from Pace University’s Environmental Litigation Clinic have appealed. She says she’s still hopeful the court might agree that Ithaca authorities made a mistake by not requiring an environmental impact statement from the college. “We have the science on our side,” she says.

Ithaca is a pretty rarefied place, an Ivy League university town. But these same tensions—potential long-term environmental and public health consequences versus the financial and maintenance concerns of the now—are pitting worried citizens against their representatives and city agencies around the country. 

New York City has 286 municipal synthetic-­turf fields, with more under construction. In Inwood, the northernmost neighborhood in Manhattan, two fields were approved via Zoom meetings during the pandemic, and Massimo Strino, a local artist who makes kaleidoscopes, says he found out only when he saw signs announcing the work on one of his daily walks in Inwood Hill Park, along the Hudson River. He joined a campaign against the plan, gathering more than 4,300 signatures. “I was canvassing every weekend,” Strino says. “You can count on one hand, literally, the number of people who said they were in favor.” 

But that doesn’t include the group that pushed for one of those fields in the first place: Uptown Soccer, which offers free and low-cost lessons and games to 1,000 kids a year, mostly from underserved immigrant families. “It was turning an unused community space into a usable space,” says David Sykes, the group’s executive director. “That trumped the sort of abstract concerns about the environmental impacts. I’m not an expert in artificial turf, but the parks department assured me that there was no risk of health effects.”

Artificial turf doesn’t go away. “You’re going to be paying to get rid of it. Somebody will have to take it to a dump, where it will sit for a thousand years.”

Graham Peaslee, emeritus nuclear physicist studying PFAS concentrations, University of Notre Dame

New York City councilmember Christopher Marte disagrees. He has introduced a bill to ban new artificial turf from being installed in parks, and he hopes the proposal will be taken up by the Parks Committee this spring. Last session, the bill had 10 cosponsors—that’s a lot. Marte says he expects resistance from lobbyists, but there’s precedent. The city of Boston banned artificial turf in 2022.  

Upstate, in a Rochester suburb called Brighton, the school district included synthetic-­turf baseball and softball diamonds in a wide-ranging February 2024 capital improvement proposition. The measure passed. In a public meeting in November 2025, the school board acknowledged the intent to use synthetic grass—or, as concerned parents had it, “to rip up a quarter ­million square feet of this open space and replace it with artificial turf,” says David Masur, executive director of the environmental group PennEnvironment, whose kids attend school in Brighton. Parents and community members mobilized against the plan, further angered when contractors also cut down a beloved 200-year-old tree. School superintendent Kevin McGowan says it’s too late to change course. Masur has been working to oppose the plan nevertheless—he says school boards are making consequential decisions about turf without sharing information or getting input, even though these fields can cost millions of dollars of taxpayer money.

In short, the fights can get tense. On Martha’s Vineyard, in Massachusetts, a meeting about plans to install an artificial field at a local high school had to be ended early amid verbal abuse. A staffer for the local board of health who voiced concern about PFAS in the turf quit the board after discovering bullet casings in her tote bag, she said, which she perceived as a death threat. After an eight-year fight, the board eventually banned artificial turf altogether. 

What happens next? Well, outdoor artificial turf lasts only eight to 12 years before it needs to be taken up and replaced. The Synthetic Turf Council says it’s at least partially recyclable and cites a company called BestPLUS Plastic Lumber as a purveyor of products made from recycled turf. The company says one of its products, a liner called GreenBoard that artificial turf can be nailed into, is at least 40% recycled from fake grass. Joseph Sadlier, vice president and general manager of plastics recycling at BestPLUS, says the company recycles over 10 million pounds annually. 

Yet the material is piling up. In 2021, a Danish company called Re-Match announced plans to open a recycling plant in Pennsylvania and began amassing thousands of tons of used plastic turf in three locations. The company filed for bankruptcy in 2025.

In Ithaca, university representatives told planning boards that it would be possible to recycle the old artificial turf they ripped out to make way for the Meinig Fieldhouse. That didn’t happen. An anonymous local activist tracked the old rolls to a hauling company a half-hour’s drive south of campus and shared pictures of them sitting on the lot, where they stayed for months. It’s unclear what their ultimate fate will be.

That’s the real problem: Artificial turf just doesn’t go away. “You’re going to be paying to get rid of it,” says Peaslee, the PFAS expert. “Somebody will have to take it to a dump, where it will sit for a thousand years.” At minimum, real grass is a net carbon sink, even including installation and maintenance. Synthetic turf releases greenhouse gases. One life-­cycle analysis of a 2.2-acre synthetic field in Toronto determined that it would emit 55 metric tons of carbon dioxide over a decade. Plastic fields need less water to maintain, but it takes water to make plastic, and natural grass lets rainwater seep into the ground. Synthetic turf sends most of it away as runoff.

It’s a boggling set of issues to factor into a decision. Rossi, the Cornell turf scientist, says he can understand why a school in the northern United States might go plastic, even when it cares about its students’ health. “It was the best bad option,” he says. Concerns about microplastics and PFAS are “significant issues we have not fully addressed.” And they need to be. 

Douglas Main is a journalist and former senior editor and writer at National Geographic.

When I started digging into desalination technology for a new story, I couldn’t help but obsess over the numbers.

I’d known on some level that desalination—pulling salt out of seawater to produce fresh water—was an increasingly important technology, especially in water-stressed regions including the Middle East. But just how much some countries rely on desalination, and how big a business it is, still surprised me.

For more on how this crucial water infrastructure is increasingly vulnerable during the war in Iran, check out my latest story. Here, though, let’s look at the state of desalination technology, by the numbers.

Desalination produces 77% of all fresh water and 99% of drinking water in Qatar.

Globally, we rely on desalination for just 1% of fresh-water withdrawals. But for some countries in the Middle East, and particularly for the Gulf Cooperation Council countries (Bahrain, Qatar, Kuwait, the United Arab Emirates, Saudi Arabia, and Oman), it’s crucial.

Qatar, home to over 3 million people, is one of the most staggering examples, with nearly all its drinking water supplies coming from desalination. But many major cities in the region couldn’t exist without the technology. There are no permanent rivers on the Arabian Peninsula, and supplies of fresh water are incredibly limited, so countries rely on facilities that can take in seawater and pull out the salt and other impurities.

The Middle East is home to just 6% of the world’s population and over 27% of its desalination facilities.

The region has historically been water-scarce, and that trend is only continuing as climate change pushes temperatures higher and changes rainfall patterns.

Of the 17,910 desalination facilities that are operational globally, 4,897 are located in the Middle East, according to a 2026 study in npj Clean Water. The technology supplies not only municipal water used by homes and businesses, but also industries including agriculture, manufacturing, and increasingly data centers.

One massive desalination plant in Saudi Arabia produces over 1 million cubic meters of fresh water per day.

The Ras Al-Khair water and power plant in Eastern Province, Saudi Arabia, is one of a growing number of gigantic plants that output upwards of a million cubic meters of water each day. That amount of water can meet the needs of millions of people in Riyadh City. Producing it takes a lot of power—the attached power plant has a capacity of 2.4 gigawatts.

While this plant is just one of thousands across the region, it’s an example of a growing trend: The average size of a desalination plant is about 10 times what it was 15 years ago, according to data from the International Energy Agency. Communities are increasingly turning to larger plants, which can produce water more efficiently than smaller ones.

Between 2024 and 2028, the Middle East’s desalination capacity could grow by over 40%.

Desalination is only going to be more crucial for life in the Middle East. The region is expected to spend over $25 billion on capital expenses for desalination facilities between 2024 and 2028, according to the 2026 npj Clean Water study. More massive plants are expected to come online in Saudi Arabia, Iraq, and Egypt during that time.

All this growth could consume a lot of electricity. Between growth of the technology generally and the move toward plants that use electricity rather than fossil fuels, desalination could add 190 terawatt-hours of electricity demand globally by 2035, according to IEA data. That’s the equivalent of about 60 million households.

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

We evolved for a linear world. If you walk for an hour, you cover a certain distance. Walk for two hours and you cover double that distance. This intuition served us well on the savannah. But it catastrophically fails when confronting AI and the core exponential trends at its heart.

Three advances are now converging to enable this. First, the basic calculators got faster. Nvidia’s chips have delivered an over sevenfold increase in raw performance in just six years, from 312 teraflops in 2020 to 2,250 teraflops today. Our own Maia 200 chip, launched this January, delivers 30% better performance per dollar than any other hardware in our fleet. Second, the numbers arrive faster thanks to a technology called HBM, or high bandwidth memory, which stacks chips vertically like tiny skyscrapers; the latest generation, HBM3, triples the bandwidth of its predecessor, feeding data to processors fast enough to keep them busy all the time. Third, the room of people with calculators became an office and then a whole campus or city. Technologies like NVLink and InfiniBand connect hundreds of thousands of GPUs into warehouse-size supercomputers that function as single cognitive entities. A few years ago this was impossible.

Mustafa Suleyman is CEO of Microsoft AI.

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here.

As the conflict in Iran has escalated, a crucial resource is under fire: the desalination technology that supplies water across much of the region.

In early March, Iran’s foreign minister accused the US of attacking a desalination plant on Qeshm Island in the Strait of Hormuz and disrupting the water supply to nearly 30 villages. (The US denied responsibility.) In the weeks since, both Bahrain and Kuwait have reported damage to desalination plants and blamed Iran, though Iran also denied responsibility.

In late March, President Donald Trump threatened the destruction of “possibly all desalinization plants” in Iran if the Strait of Hormuz was not reopened. Since then, he’s escalated his threats against Iran, warning of plans to attack other crucial civilian infrastructure like power plants and bridges.

Countries in the Middle East, particularly the Gulf states, rely on the technology to turn salt water into fresh water for farming, industry, and—crucially—drinking. The mounting attacks and threats to date highlight just how vital the industry is to the region—a situation made even more precarious by rising temperatures and extreme weather driven by climate change.

Right now, 83% of the Middle East is under extremely high water stress, says Liz Saccoccia, a water security associate at the World Resources Institute. Future projections suggest that’s going to increase to about 100% by 2050, she adds: “This is a continuing trend, and it’s getting worse, not better.”

Here’s a look at desalination technology in the Middle East and what wartime threats to the critical infrastructure could mean for people in the region. 

A vital resource

Desalination technology has helped provide water supplies in the Middle East since the early 20th century and became widespread in the 1960s and 1970s.

There are two major categories of desalination plants. Thermal plants use heat to evaporate water, leaving salt and other impurities behind. The vapor can then be condensed into usable fresh water. The alternative is membrane-based technology like reverse osmosis, which pushes water through membranes that have tiny pores—so small that salt can’t get through.

Early desalination plants in the Middle East were the first type, burning fossil fuels to evaporate water, leaving the salt behind. This technique is incredibly energy-intensive, and over time, processes that rely on filters became the dominant choice.

Membrane technologies have made up essentially all new desalination capacity in recent years; the last major thermal plant built in the Gulf came online in 2018. Many reverse osmosis plants still rely on fossil fuels, but they’re more efficient. Since then, membrane technologies have added more than 15 million cubic meters of daily capacity—enough to supply water to millions of people.

Capacity has expanded quickly in recent years; between 2006 and 2024, countries across the Middle East collectively spent over $50 billion building and upgrading desalination facilities, and nearly that much operating them.

Today, there are nearly 5,000 desalination plants operational across the Middle East.

And looking ahead, growth is continuing. Between 2024 and 2028, daily capacity is expected to grow from about 29 million cubic meters to 41 million cubic meters.

Uneven vulnerabilities

Some countries rely on the technology more than others. Iran, for example, uses desalination for about 3% of its municipal fresh water. The country has access to groundwater and some surface water, including rivers, though these resources are being stretched thin by agriculture and extreme drought.

Other nations in the region, particularly the Gulf countries (Bahrain, Qatar, Kuwait, the United Arab Emirates, Saudi Arabia, and Oman), have much more limited water resources and rely heavily on desalination. Across these six nations, all but the UAE get more than half their drinking water from desalination, and for Bahrain, Qatar, and Kuwait the figure is more than 90%.

“The Gulf countries are much, much more vulnerable to attacks on their desalination plants than Iran is,” says David Michel, a senior associate in the global food and water security program at the Center for Strategic and International Studies.

There are thousands of desalination facilities across the region, so the system wouldn’t collapse if a small number were taken offline, Michel says. However, in recent years there’s been a trend toward larger, more centralized plants.

The average desalination plant is about 10 times larger than it was 15 years ago, according to data from the International Energy Agency. The largest desalination plants today can produce 1 million cubic meters of water daily, enough for hundreds of thousands of people. Taking one or more of these massive facilities offline could have a significant effect on the system, Michel says.

Escalating threats

Desalination facilities are quite linear, meaning there are multiple steps and pieces of equipment that work in sequence—and the failure of a component in that chain can take an entire facility down. Attacks on water inlets, transportation networks, and power supplies can also disrupt the system, Michel says. 

During the Gulf War in 1991, Iraqi forces pumped oil into the gulf, contaminating the water and shutting down desalination plants in Kuwait. 

The facilities are also generally located close to other targets in this conflict. Desalination is incredibly energy intensive, so about three-quarters of facilities in the region are next to power plants. Trump has repeatedly threatened power plants in Iran. In response, Iran’s military has said that if civilian targets are hit, the country will respond with strikes that are “much more devastating and widespread.” Other governments and organizations, including the United Nations, the European Union, and the Red Cross, have broadly condemned threats to infrastructure as illegal. 

But war isn’t the only danger facing these plants, even if it is the most immediate. Some studies have suggested that global warming could strengthen cyclones in the region, and these extreme weather events could force shutdowns or damage equipment.

Water pollution could also cause shutdowns. Oil spills, whether accidental or intentional, as in the case of the Gulf War, can  wreak havoc. And in 2009, a red algae bloom closed desalination plants in Oman and the United Arab Emirates for weeks. The algae fouled membranes and blocked the plants from being able to take water in from the Persian Gulf and the Gulf of Oman.

Desalination facilities could become more resilient to threats in the future, and they may need to as their importance continues to grow. 

There’s increasing interest in running desalination facilities at least partially on solar power, which could help reduce dependence on the oil that powers most facilities today. The Hassyan seawater desalination project in the UAE, currently under construction, would be the largest reverse osmosis plant in the world to operate solely with renewable energy. 

Another way to increase resilience is for countries to build up more strategic water storage to meet demand. Qatar recently issued new policies that aim to improve management and storage of desalinated water, for example. Countries could also work together to invest in shared infrastructure and policies that help strengthen the water supply through the region. 

Preparedness, resilience, and cooperation will be key for the Middle East broadly as critical infrastructure, including the water supply, is increasingly under threat. 

“The longer the conflict goes on, the more likely we’ll see significant water infrastructure damage,” says Ginger Matchett, an assistant director at the Atlantic Council. “What worries me is that after this war ends, some of the lessons will show how water can be weaponized more strategically than previously imagined.” 

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

Within Silicon Valley’s orbit, an AI-fueled jobs apocalypse is spoken about as a given. The mood is so grim that a societal impacts researcher at Anthropic, responding Wednesday to a call for more optimistic visions of AI’s future, said there might be a recession in the near term and a “breakdown of the early-career ladder.” Her less-measured colleague Dario Amodei, the company’s CEO, has called AI “a general labor substitute for humans” that could do all jobs in less than five years. And those ideas are not just coming from Anthropic, of course. 

These conversations have unsurprisingly left many workers in a panic (and are probably contributing to support for efforts to entirely pause the construction of data centers, some of which gained steam last week). The panic isn’t being helped by lawmakers, none of whom have articulated a coherent plan for what comes next.

Even economists who have cautioned that AI has not yet cut jobs and may not result in a cliff ahead are coming around to the idea that it could have a unique and unprecedented impact on how we work. 

Alex Imas, based at the University of Chicago, is one of those economists. He shared two things with me when we spoke on Friday morning: a blunt assessment that our tools for predicting what this will look like are pretty abysmal, and a “call to arms” for economists to start collecting the one type of data that could make a plan to address AI in the workforce possible at all. 

On our abysmal tools: consider the fact that any job is made up of individual tasks. One part of a real estate agent’s job, for example, is to ask clients what sort of property they want to buy. The US government chronicled thousands of these tasks in a massive catalogue first launched in 1998 and updated regularly since then. This was the data that researchers at OpenAI used in December to judge how “exposed” a job is to AI (they found a real estate agent to be 28% exposed, for example). Then in February, Anthropic used this data in its analysis of millions of Claude conversations to see which tasks people are actually using its AI to complete and where the two lists overlapped.

But knowing the AI exposure of tasks leads to an illusory understanding of how much a given job is at risk, Imas says. “Exposure alone is a completely meaningless tool for predicting displacement,” he told me.

Sure, it is illustrative in the gloomiest case—for a job in which literally every task could be done by AI with no human direction. If it costs less for an AI model to do all those tasks than what you’re paid—which is not a given, since reasoning models and agentic AI can rack up quite a bill—and it can do them well, the job likely disappears, Imas says. This is the oft-mentioned case of the elevator operator from decades ago; maybe today’s parallel is a customer service agent solely doing phone call triage. 

But for the vast majority of jobs, the case is not so simple. And the specifics matter, too: Some jobs are likely to have dark days ahead, but knowing how and when this will play out is hard to answer when only looking at exposure.

Take writing code, for example. Someone who builds premium dating apps, let’s say, might use AI coding tools to create in one day what used to take three days. That means the worker is more productive. The worker’s employer, spending the same amount of money, can now get more output. So then will the employer want more employees or fewer? 

This is the question that Imas says should keep any policymaker up at night, because the answer will change depending on the industry. And we are operating in the dark. 

In this coder’s case, these efficiencies make it possible for dating apps to lower prices. (A skeptic might expect companies to simply pocket the gains, but in a competitive market, they risk being undercut if they do.) These lower prices will always drive some increase in demand for the apps. But how much? If millions more people want it, the company might grow and ultimately hire more engineers to meet this demand. But if demand barely ticks up—maybe the people who don’t use premium dating apps still won’t want them even at a lower price—fewer coders are needed, and layoffs will happen.

Repeat this hypothetical across every job with tasks that AI can do, and you have the most pressing economic question of our time: the specifics of price elasticity, or how much demand for something changes when its price changes. And this is the second part of what Imas emphasized last week: We don’t currently have this data across the economy. But we could. 

We do have the numbers for grocery items like cereal and milk, Imas says, because the University of Chicago partners with supermarkets to get data from their price scanners. But we don’t have such figures for tutors or web developers or dietitians (all jobs found to have “exposure” to AI, by the way). Or at least not in a way that’s been widely compiled or made accessible to researchers; sometimes it’s scattered across private companies or consultancies. 

“We need, like, a Manhattan Project to collect this,” Imas says. And we don’t need it just for jobs that could obviously be affected by AI now: “Fields that are not exposed now will become exposed in the future, so you just want to track these statistics across the entire economy.”

Getting all this information would take time and money, but Imas makes the case that it’s worth it; it would give economists the first realistic look at how our AI-enabled future could unfold and give policymakers a shot at making a plan for it.

For years Mike McClary sold the Guardian LTE Flashlight, a heavy-duty black model, online through his small outdoor brand. The product, designed for brightness and durability, became one of his most popular items ever. Even after he stopped offering it around 2017, customers kept sending him emails asking where they could buy it. 

When McClary decided to revisit the Guardian flashlight in 2025, he didn’t begin the way he might have in the past, by combing through supplier listings and sending inquiries to factories. Instead, he opened Accio, an AI sourcing and researching tool on Alibaba.com.

For small entrepreneurs in the US, deciding what to sell and where to make it has traditionally been a slow, labor-intensive process that can take months. Now that work is increasingly being done by AI tools like Accio, which help connect businesses with manufacturers in countries including China and India. Business owners and e-commerce experts told MIT Technology Review that these AI tools are making sourcing more accessible and significantly shortening the time it takes to go from product idea to launch. 

McClary, 51, who runs his business from his Illinois living room, has sold products ranging from leather conditioner to camping lights, including one rechargeable lantern that brought in half a million dollars. Like many small online merchants, he built his business by being extremely scrappy—spotting demand for a product, tweaking existing designs, finding a factory, doing modest marketing, and getting the goods in front of customers fast. 

This time, though, he began by telling Accio about the flashlight’s original design, production cost, and profit margin. Then Accio suggested several changes, making it smaller and slightly less bright and switching its charging method to battery power. It also identified a manufacturer in Ningbo, China, that McClary said could cut the manufacturing cost from $17 to about $2.50 per unit.

McClary took the process from there, contacting the supplier himself to discuss the revised design. Within a month, the new version of the Guardian flashlight was back up for sale on Amazon and on his brand’s website.

The new factory hunt

Although Alibaba is better known for owning Taobao, the biggest shopping site in China, its first business was Alibaba.com, the primary website that lists Chinese factories open for bulk orders. Placing an order with a manufacturer usually requires far more than clicking “Buy.” Sellers often spend days or weeks browsing listings, comparing suppliers’ reviews and manufacturing capacities, asking about minimum order quantities, requesting samples, and negotiating timelines and customization options. 

But Accio has gained significant momentum by changing how that sourcing gets done. Launched in 2024, Accio exceeded 10 million monthly active users in March 2026, according to the company. That means about one in five Alibaba users consults with AI about product sourcing.

Accio’s interface looks a lot like ChatGPT or Claude: Users type a question into an empty box and choose between “fast” and “thinking” modes. But when asked about products, the tool returns more than text, offering charts, links, and visuals and asking follow-up questions to clarify the buyer’s needs. It then narrows the field to one or a handful of suppliers that appear capable of delivering. After that, the human work begins: Users still have to reach out to suppliers themselves and negotiate the details.

Zhang Kuo, the president of Alibaba.com, told MIT Technology Review that the tool is built on multiple frontier models, including the company’s own Qwen series, a popular family of open-source large language models. The system is able to pull from the site’s millions of supplier profiles and is trained on 26 years of proprietary transaction data.

For tasks like product research and sourcing analysis, the tool “blows it away” compared with general AI tools like ChatGPT, says Richard Kostick, CEO of the beauty brand 100% Pure.

Many websites have tried using AI to assist shopping, but Alibaba has been one of the most aggressive. In March, Eddie Wu, CEO of the site’s parent company Alibaba Group, told managers that integrating the company’s core services with Qwen’s AI capabilities is a top priority. During a Chinese New Year promotion of Qwen’s personal shopping AI agent, where the company gave away cash, customers placed 200 million orders, the firm says.

Vincenzo Toscano, an e-commerce seller and consultant, recommended Accio to his clients before deciding to try it himself for a new sunglasses brand. He came in with a rough vision: a brand shaped by his Italian heritage, his personal style, and a boutique aesthetic. He says the AI helped turn that concept into something more concrete, suggesting materials, refining the look, and pointing to design ideas that felt current.

But the tool has clear limits. McClary, who uses AI tools regularly, says Accio is strongest when it comes to product ideation, but less helpful on marketing questions such as advertising and social media outreach. To use it well, he says, buyers still need to challenge its recommendations, since some can be generic.

The rest of the business

As platforms become more AI-driven, manufacturers are adjusting too. Sally Li, a representative at a makeup packaging company in Wuhan, China, says her firm has started writing more detailed product descriptions and adding information about its equipment and manufacturing experience on Alibaba.com because it suspects those details make its listings more likely to be surfaced by AI.

Yan says manufacturers cannot tell whether an inquiry from a customer was generated or guided by AI, and that her firm is not using AI to negotiate pricing or product details.

“AI agents are increasingly used by people to assist purchase decisions and even directly making transactions, and with clear guardrails, they can become extremely useful,” says Jiaxin Pei, a research scientist at the Stanford Institute for Human-Centered AI, “but agents need to act transparently, securely, and in the customer’s best interest.” Pei says developers of these tools should disclose the data they collect and the incentives built into them to ensure that the marketplace remains fair.

Zhang, of Alibaba.com, says Accio currently does not include advertising. Suppliers can pay for higher placement in Alibaba.com’s regular search results, but Zhang says Accio is “not integrated” with that system. “We haven’t had a clear answer in terms of how to monetize this tool,” he says. For now, users can pay for additional tokens to continue chatting with the agent after their free queries run out.

Sellers say that while AI tools have made it easier to come up with ideas and get a business off the ground, they do not replace the core skills that make someone good at e-commerce. McClary believes that even when sellers have access to the same market information, some are still better at making decisions, acting quickly, and actually delivering on orders. Those differences, he says, still go a long way.

Toscano, the brand founder and e-commerce consultant, feels good about officially launching his new brand of sunglasses in just a few months: “We [small business owners] always have to bootstrap a lot of decisions. Deciding what to sell often comes down to an educated guess,” he says, “And we’re now in an era when making those decisions is easier than ever.”

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here.

In January, Elon Musk’s SpaceX filed an application with the US Federal Communications Commission to launch up to one million data centers into Earth’s orbit. The goal? To fully unleash the potential of AI without triggering an environmental crisis on Earth. But could it work?

SpaceX is the latest in a string of high-tech companies extolling the potential of orbital computing infrastructure. Last year, Amazon founder Jeff Bezos said that the tech industry will move toward large-scale computing in space. Google has plans to loft data-crunching satellites, aiming to launch a test constellation of 80 as early as next year. And last November Starcloud, a startup based in Washington State, launched a satellite fitted with a high-performance Nvidia H100 GPU, marking the first orbital test of an advanced AI chip. The company envisions orbiting data centers as large as those on Earth by 2030.

Proponents believe that putting data centers in space makes sense. The current AI boom is straining energy grids and adding to the demand for water, which is needed to cool the computers. Communities in the vicinity of large-scale data centers worry about increasing prices for those resources as a result of the growing demand, among other issues.

In space, advocates say, the water and energy problems would be solved. In constantly illuminated sun-synchronous orbits, space-borne data centers would have uninterrupted access to solar power. At the same time, the excess heat they produce would be easily expelled into the cold vacuum of space. And with the cost of space launches decreasing, and mega-rockets such as SpaceX’s Starship promising to push prices even lower, there could be a point at which moving the world’s data centers into space makes sound business sense. Detractors, on the other hand, tell a different story and point to a variety of technological hurdles, though some say it’s possible they may be surmountable in the not-so-distant future. Here are four of the must-haves we’d need to make space-based data centers a reality. 

A way to carry away heat 

AI data centers produce a lot of heat. Space might seem like a great place to dispel that heat without using up massive amounts of water. But it’s not so simple. To get the power needed to run 24-7, a space-based data center would have to be in a constantly illuminated orbit, circling the planet from pole to pole, and never hide in Earth’s shadow. And in that orbit, the temperature of the equipment would never drop below 80 °C, which is way too hot for electronics to operate safely in the long term. 

Getting the heat out of such a system is surprisingly challenging. “Thermal management and cooling in space is generally a huge problem,” says Lilly Eichinger, CEO of the Austrian space tech startup Satellives.

On Earth, heat dissipates mostly through the natural process of convection, which relies on the movement of gases and liquids like air and water. In the vacuum of space, heat has to be removed through the far less efficient process of radiation. Safely removing the heat produced by the computers, as well as what’s absorbed from the sun, requires large radiative surfaces. The bulkier the satellite, the harder it is to send all the heat inside it out into space.

But Yves Durand, former director of technology at the European aerospace giant Thales Alenia Space, says that technology already exists to tackle the problem.

The company previously developed a system for large telecommunications satellites that can pipe refrigerant fluid through a network of tubing using a mechanical pump, ultimately transferring heat from within a spacecraft to radiators on the exterior. Durand led a 2024 feasibility study on space-based data centers, which found that although challenges exist, it should be possible for Europe to put gigawatt-scale data centers (on par with the largest Earthbound facilities) into orbit before 2050. These would be considerably larger than those envisioned by SpaceX, featuring solar arrays hundreds of meters in size—larger than the International Space Station.

Computer chips that can withstand a radiation onslaught

The space around Earth is constantly battered by cosmic particles and lashed by solar radiation. On Earth’s surface, humans and their electronic devices are protected from this corrosive soup of charged particles by the planet’s atmosphere and magnetosphere. But the farther away from Earth you venture, the weaker that protection becomes. Studies show that aircraft crews have a higher risk of developing cancer because of their frequent exposure to high radiation at cruising altitude, where the atmosphere is thin and less protective.

Electronics in space are at risk of three types of problems caused by high radiation levels, says Ken Mai, a principal systems scientist in electrical and computer engineering at Carnegie Mellon University. Phenomena known as single-event upsets can cause bit flips and corrupt stored data when charged particles hit chips and memory devices. Over time, electronics in space accumulate damage from ionizing radiation that degrades their performance. And sometimes a charged particle can strike the component in a way that physically displaces atoms on the chip, creating permanent damage, Mai explains.

Traditionally, computers launched to space had to undergo years of testing and were specifically designed to withstand the intense radiation present in Earth’s orbit. These space-hardened electronics are much more expensive, though, and their performance is also years behind the state-of-the-art devices for Earth-based computing. Launching conventional chips is a gamble. But Durand says cutting-edge computer chips use technologies that are by default more resistant to radiation than past systems. And in mid-March, Nvidia touted hardware, including a new GPU, that is “bringing AI compute to orbital data centers.” 

Nvidia’s head of edge AI marketing, Chen Su, told MIT Technology Review, that “Nvidia systems are inherently commercial off the shelf, with radiation resilience achieved at the system level rather than through radiation‑hardened silicon alone.” He added that satellite makers increase the chips’ resiliency with the help of shielding, advanced software for error detection, and architectures that combine the consumer-grade devices with bespoke, hardened technologies.

Still, Mai says that the data-crunching chips are only one issue. The data centers would also need memory and storage devices, both of which are vulnerable to damage by excessive radiation. And operators would need the ability to swap things out or adapt when issues arise. The feasibility and affordability of using robots or astronaut missions for maintenance is a major question mark hanging over the idea of large-scale orbiting data centers.

“You not only need to throw up a data center to space that meets your current needs; you need redundancy, extra parts, and reconfigurability, so when stuff breaks, you can just change your configuration and continue working,” says Mai. “It’s a very challenging problem because on one hand you have free energy and power in space, but there are a lot of disadvantages. It’s quite possible that those problems will outweigh the advantages that you get from putting a data center into space.”

In addition to the need for regular maintenance, there’s also the potential for catastrophic loss. During periods of intense space weather, satellites can be flooded with enough radiation to kill all their electronics. The sun has just passed the most active phase of its 11-year cycle with relatively little impact on satellites. Still, experts warn that since the space age began, the planet has not experienced the worst the sun is capable of. Many doubt whether the low-cost new space systems that dominate Earth’s orbits today are prepared for that.

A plan to dodge space debris

Both large-scale orbiting data centers such as those envisioned by Thales Alenia Space and the mega-constellations of smaller satellites as proposed by SpaceX give a headache to space sustainability experts. The space around Earth is already quite crowded with satellites. Starlink satellites alone perform hundreds of thousands of collision avoidance maneuvers every year to dodge debris and other spacecraft. The more stuff in space, the higher the likelihood of a devastating collision that would clutter the orbit with thousands of dangerous fragments.

Large structures with hundreds of square meters of solar arrays would quickly suffer damage from small pieces of space debris and meteorites, which would over time degrade the performance of their solar panels and create more debris in orbit. Operating one million satellites in low Earth orbit, the region of space at the altitude of up to 2,000 kilometers, might be impossible to do safely unless all satellites in that area are part of the same network so they can communicate effectively to maneuver around each other, Greg Vialle, the founder of the orbital recycling startup Lunexus Space, told MIT Technology Review.

“You can fit roughly four to five thousand satellites in one orbital shell,” Vialle says. “If you count all the shells in low Earth orbit, you get to a number of around 240,000 satellites maximum.”

And spacecraft must be able to pass each other at a safe distance to avoid collisions, he says. 

“You also need to be able to get stuff up to higher orbits and back down to de-orbit,” he adds. “So you need to have gaps of at least 10 kilometers between the satellites to do that safely. Mega-constellations like Starlink can be packed more tightly because the satellites communicate with each other. But you can’t have one million satellites around Earth unless it’s a monopoly.”

On top of that, Starlink would likely want to regularly upgrade its orbiting data centers with more modern technology. Replacing a million satellites perhaps every five years would mean even more orbital traffic—and it could increase the rate of debris reentry into Earth’s atmosphere from around three or four pieces of junk a day to about one every three minutes, according to a group of astronomers who filed objections against SpaceX’s FCC application. Some scientists are concerned that reentering debris could damage the ozone layer and alter Earth’s thermal balance. 

Economical launch and assembly

The longer hardware survives in orbit, the better the return on investment. But for orbital data centers to make economic sense, companies will have to find a relatively cheap way to get that hardware in orbit. SpaceX is betting on its upcoming Starship mega-rocket, which will be able to carry up to six times as much payload as the current workhorse, Falcon 9. The Thales Alenia Space study concluded that if Europe were to build its own orbital data centers, it would have to develop a similarly potent launcher. 

But launch is only part of the equation. A large-scale orbital data center won’t fit in a rocket—even a mega-rocket. It will need to be assembled in orbit. And that will likely require advanced robotic systems that do not exist yet. Various companies have conducted Earth-based tests with precursors of such systems, but they are still far from real-world use.

Durand says that in the short term, smaller-scale data centers are likely to establish themselves as an integral part of the orbital infrastructure, by processing images from Earth-observing satellites directly in space without having to send them to Earth. That would be a huge help for companies selling insights from space, as many of these data sets are extremely large, and competition for opportunities to downlink them to Earth for processing via ground stations is growing.

“The good thing with orbital data centers is that you can start with small servers and gradually increase and build up larger data centers,” says Durand. “You can use modularity. You can learn little by little and gradually develop industrial capacity in space. We have all the technology, and the demand for space-based data processing infrastructure is huge, so it makes sense to think about it.”

Smaller facilities probably won’t do much to offset the strain that terrestrial data centers are placing on the planet’s water and electricity, though. That vision of the future might take decades to come to fruition, some critics think—if it even gets off the ground at all. 

As the war in Iran continues to engulf the Middle East and the Strait of Hormuz stays closed, one of the most visible global economic ripple effects has been fossil-fuel prices. In particular, you can’t get away from news about the price of gasoline, which just topped an average of $4 a gallon in the US, its highest level since 2022.

But looking ahead, further consequences for the global economy could be looming in plastics. Plastics are made using petrochemicals, and the supply chain impacts of the oil bottleneck near Iran are starting to build up. 

Plastic production accounts for roughly 5% of global carbon dioxide emissions today. And our current moment shows just how embedded oil and gas products are in our lives. It goes far beyond their use for energy. 

As I write this, I’m wearing clothes that contain plastic fibers, typing on a plastic keyboard, and looking through the plastic lenses of my glasses. It’s hard to imagine what our world looks like without plastic. And in some ways, moving away from fossil-derived plastic could prove even more complicated than decarbonizing our energy system. 

Crude oil prices have been on a roller-coaster in recent weeks, and prices have recently topped $100 a barrel.

Crude oil contains a huge range of hydrocarbons, and it’s typically refined by putting it through a distillation unit that separates the raw material into different fractions according to their boiling point. Those fractions then go on to be further processed into everything from jet fuel to asphalt binder. We’ve already seen the price spikes for some materials pulled out of crude oil, like gasoline and jet fuel.

Let’s zoom in on another component, naphtha. It can be added to gasoline and jet fuel to improve performance. It can also be used as a solvent or as a raw material to make plastics.

The Middle East currently accounts for about 20% of global naphtha production­ and supplies about 40% of the market in Asia, where prices are already up by 50% over the last month.

We’re starting to see these effects trickle down already. The price of polypropylene (which is made from naphtha and used for food containers, bottle caps, and even automotive parts) is climbing, especially in Asia.  

Typically, manufacturers have a bit of stock built up, but that’ll be exhausted soon, likely in the coming weeks. The largest supplier of water bottles in India recently announced that it would raise prices by 11% after its packaging costs went up by over 70%, according to reporting from Reuters. Toys could be more expensive this holiday season as manufacturers grapple with supply chain concerns.

Americans will likely feel these ripples especially hard if disruptions continue. The average US resident used over 250 kilograms of new plastics in 2019, according to a 2022 report from the Organization for Economic Cooperation and Development. That’s an absolutely massive number—the global average is just 60 kilograms.

The effects of higher prices for both fuels and feedstocks could compound and multiply, and alternatives aren’t widely available. Bio-based plastics made with materials like plant sugars exist, but they still make up a vanishingly tiny portion of the market. As of 2025, global plastics production totaled over 431 million metric tons per year. Bio-based and bio-degradable plastics made up about 0.5% of that, a share that could reach 1% by 2030.

Bio-based plastics are much more expensive than their fossil-derived counterparts. And many are made using agricultural raw materials, so scaling them up too much could be harmful for the environment and might compete with other industries like food production.

Recycling isn’t the easy answer either. Mechanical recycling is the current standard method used for materials like the plastics that make up water bottles and disposable coffee cups. But that degrades the materials over time, so they can’t be used infinitely. Chemical recycling has its own host of issues—the facilities that do it can be highly polluting, and today plastics that go into advanced recycling plants largely don’t actually go into new plastics.

There’s been a lot of talk in recent weeks about how this energy crisis is going to push the world more toward renewable energy. Solar panels, electric vehicles, and batteries could suddenly become more attractive as we face the drastic consequences of a disruption in the global fossil-fuel supply.

But when it comes to plastic, the future looks far more complicated. Even though the plastics industry is facing much the same disruptions as the energy sector, there aren’t the same obvious alternatives available for a transition. Our lives are tied up in plastic, with uses ranging from the essential (like medical equipment) to the mundane (my to-go coffee cup). Soon, our economy could feel the effects of just how much we rely on fossil-derived plastics, and how hard it’s going to be to replace them. 

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

When Zeus, a medical student living in a hilltop city in central Nigeria, returns to his studio apartment from a long day at the hospital, he turns on his ring light, straps his iPhone to his forehead, and starts recording himself. He raises his hands in front of him like a sleepwalker and puts a sheet on his bed. He moves slowly and carefully to make sure his hands stay within the camera frame. 

Zeus is a data recorder for Micro1, a US company based in Palo Alto, California that collects real-world data to sell to robotics companies. As companies like Tesla, Figure AI, and Agility Robotics race to build humanoids—robots designed to resemble and move like humans in factories and homes—videos recorded by gig workers like Zeus are becoming the hottest new way to train them. 

Micro1 has hired thousands of contract workers in more than 50 countries, including India, Nigeria, and Argentina, where swathes of tech-savvy young people are looking for jobs. They’re mounting iPhones on their heads and recording themselves folding laundry, washing dishes, and cooking. The job pays well by local standards and is boosting local economies, but it raises thorny questions around privacy and informed consent. And the work can be challenging at times—and weird.

Zeus found the job in November, when people started talking about it everywhere on LinkedIn and YouTube. “This would be a real nice opportunity to set a mark and give data that will be used to train robots in the future,” he thought. 

Zeus is paid $15 an hour, which is good income in Nigeria’s strained economy with high unemployment rates. But as a bright-eyed student dreaming of becoming a doctor, he finds ironing his clothes for hours every day boring. 

“I really [do] not like it so much,” he says. “I’m the kind of person that requires … a technical job that requires me to think.” 

Zeus, and all the workers interviewed by MIT Technology Review, asked to be referred to only by pseudonyms because they were not authorized to talk about their work.

Humanoid robots are notoriously hard to build because manipulating physical objects is a difficult skill to master. But the rise of large language models underlying chatbots like ChatGPT has inspired a paradigm shift in robotics. Just as large language models learned to generate words by being trained on vast troves of text scraped from the internet, many researchers believe that humanoid robots can learn to interact with the world by being trained on massive amounts of movement data. 

Editor’s note: In a recent poll, MIT Technology Review readers selected humanoid robots as the 11th breakthrough for our 2026 list of 10 Breakthrough Technologies.

Robotics requires far more complex data about the physical world, though, and that is much harder to find. Virtual simulations can train robots to perform acrobatics, but not how to grasp and move objects, because simulations struggle to model physics with perfect accuracy. For robots to work in factories and serve as housekeepers, real-world data, however time-consuming and expensive to collect, may be what we need. 

Investors are pouring money feverishly into solving this challenge, spending over $6 billion on humanoid robots in 2025. And at-home data recording is becoming a booming gig economy around the world. Data companies like Scale AI and Encord are recruiting their own armies of data recorders, while DoorDash pays delivery drivers to film themselves doing chores. And in China, workers in dozens of state-owned robot training centers wear virtual-reality headsets and exoskeletons to teach humanoid robots how to open a microwave and wipe down the table. 

“There is a lot of demand, and it’s increasing really fast,” says Ali Ansari, CEO of Micro1. He estimates that robotics companies are now spending more than $100 million each year to buy real-world data from his company and others like it.

A day in the life

Workers at Micro1 are vetted by an AI agent named Zara that conducts interviews and reviews samples of chore videos. Every week, they submit videos of themselves doing chores around their homes, following a list of instructions about things like keeping their hands visible and moving at natural speed. The videos are reviewed by both AI and a human and are either accepted or rejected. They’re then annotated by AI and a team of hundreds of humans who label the actions in the footage.

Because this approach to training robots is in its infancy, it’s not clear yet what makes good training data. Still, “you need to give lots and lots of variations for the robot to generalize well for basic navigation and manipulation of the world,” says Ansari.

But many workers say that creating a variety of “chore content” in their tiny homes is a challenge. Zeus, a scrappy student living in a humble studio, struggles to record anything beyond ironing his clothes every day. Arjun, a tutor in Delhi, India, takes an hour to make a 15-minute video because he spends so much time brainstorming new chores.

“How much content [can be made] in the home? How much content?” he says. 

There’s also the sticky question of privacy. Micro1 asks workers not to show their faces to the camera or reveal personal information such as names, phone numbers, and birth dates. Then it uses AI and human reviewers to remove anything that slips through. 

But even without faces, the videos capture an intimate slice of workers’ lives: the interiors of their homes, their possessions, their routines. And understanding what kind of personal information they might be recording while they’re busy doing chores on camera can be tricky. Reviews of such footage might not filter out sensitive information beyond the most obvious identifiers.

For workers with families, keeping private life off camera is a constant negotiation. Arjun, a father of two daughters, has to wrangle his chaotic two-year-old out of frame. “Sometimes it’s very difficult to work because my daughter is small,” he says. 

Sasha, a banker turned data recorder in Nigeria, tiptoes around when she hangs her laundry outside in a shared residential compound so she won’t record her neighbors, who watch her in bewilderment.

While the workers interviewed by MIT Technology Review understand that their data is being used to train robots, none of them know how exactly their data will be used, stored, and shared with third parties, including the robotics companies that Micro1 is selling the data to. For confidentiality reasons, says Ansari, Micro1 doesn’t name its clients or disclose to workers the specific nature of the projects they are contributing to.

“It is important that if workers are engaging in this, that they are informed by the companies themselves of the intention … where this kind of technology might go and how that might affect them longer term,” says Yasmine Kotturi, a professor of human-centered computing at the University of Maryland.

Occasionally, some workers say, they’ve seen other workers asking on the company Slack channel if the company could delete their data. Micro1 declined to comment on whether such data is deleted.

“People are opting into doing this,” says Ansari. “They could stop the work at any time.”

Hungry for data

With thousands of workers doing their chores differently in different homes, some roboticists wonder if the data collected from them is reliable enough to train robots safely. 

“How we conduct our lives in our homes is not always right from a safety point of view,” says Aaron Prather, a roboticist at ASTM International. “If those folks are teaching those bad habits that could lead to an incident, then that’s not good data.” And the sheer volume of data being collected makes reviewing it for quality control challenging. But Ansari says the company rejects videos showing unsafe ways of performing a task, while clumsy movements can be useful to teach robots what not to do.

Then there’s the question of how much of this data we need. Micro1 says it has tens of thousands of hours of footage, while Scale AI announced it had gathered more than 100,000 hours.

“It’s going to take a long time to get there,” says Ken Goldberg, a roboticist at the University of California, Berkeley. Large language models were trained on text and images that would take a human 100,000 years to read, and humanoid robots may need even more data, because controlling robotic joints is even more complicated than generating text. “It’s going to take longer than people think,” he says.

When Dattu, an engineering student living in a bustling tech hub in India, comes home after a full day of classes at his university, he skips dinner and dashes to his tiny balcony, cramped with potted plants and dumbbells. He straps his iPhone to his forehead and records himself folding the same set of clothes over and over again. 

His family stares at him quizzically. “It’s like some space technology for them,” he says. When he tells his friends about his job, “they just get astounded by the idea that they can get paid by recording chores.”

Juggling his university studies with data recording, as well as other data annotation gigs, takes a toll on him. Still, “it feels like you’re doing something different than the whole world,” he says. 

For decades, artificial intelligence has been evaluated through the question of whether machines outperform humans. From chess to advanced math, from coding to essay writing, the performance of AI models and applications is tested against that of individual humans completing tasks. 

This framing is seductive: An AI vs. human comparison on isolated problems with clear right or wrong answers is easy to standardize, compare, and optimize. It generates rankings and headlines. 

But there’s a problem: AI is almost never used in the way it is benchmarked. Although   researchers and industry have started to improve benchmarking by moving beyond static tests to more dynamic evaluation methods, these  innovations resolve only part of the issue. That’s because they still evaluate AI’s performance outside the human teams and organizational workflows where its real-world performance ultimately unfolds. 

While AI is evaluated at the task level in a vacuum, it is used in messy, complex environments where it usually interacts with more than one person. Its performance (or lack thereof) emerges only over extended periods of use. This misalignment leaves us misunderstanding AI’s capabilities, overlooking systemic risks, and misjudging its economic and social consequences.

To mitigate this, it’s time to shift from narrow methods to benchmarks that assess how AI systems perform over longer time horizons within human teams, workflows, and organizations. I have studied real-world AI deployment since 2022 in small businesses and health, humanitarian, nonprofit, and higher-education organizations in the UK, the United States, and Asia, as well as within leading AI design ecosystems in London and Silicon Valley. I propose a different approach, which I call HAIC benchmarks—Human–AI, Context-Specific Evaluation.

What happens when AI fails 

For governments and businesses, AI benchmark scores appear more objective than vendor claims. They’re a critical part of determining whether an AI model or application is “good enough” for real-world deployment. Imagine an AI model that achieves impressive technical scores on the most cutting-edge benchmarks—98% accuracy, groundbreaking speed, compelling outputs. On the strength of these results, organizations may decide to adopt the model, committing sizable financial and technical resources to purchasing and integrating it. 

But then, once it’s adopted, the gap between benchmark and real-world performance quickly becomes visible. For example, take the swathe of FDA-approved AI models that can read medical scans faster and more accurately than an expert radiologist. In the radiology units of hospitals from the heart of California to the outskirts of London, I witnessed staff using highly ranked radiology AI applications. Repeatedly, it took them extra time to interpret AI’s outputs alongside hospital-specific reporting standards and nation-specific regulatory requirements. What appeared as a productivity-enhancing AI tool when tested in a vacuum introduced delays in practice. 

It soon became clear that the benchmark tests on which medical AI models are assessed do not capture how medical decisions are actually made. Hospitals rely on multidisciplinary teams—radiologists, oncologists, physicists, nurses—who jointly review patients. Treatment planning rarely hinges on a static decision; it evolves as new information emerges over days or weeks. Decisions often arise through constructive debate and trade-offs between professional standards, patient preferences, and the shared goal of long-term patient well-being. No wonder even highly scored AI models struggle to deliver the promised performance once they encounter the complex, collaborative processes of real clinical care.

The same pattern emerges in my research across other sectors: When embedded within real-world work environments, even AI models that perform brilliantly on standardized tests don’t perform as promised. 

When high benchmark scores fail to translate into real-world performance, even the most highly scored AI is soon abandoned to what I call the “AI graveyard.” The costs are significant: Time, effort and money end up being wasted. And over time, repeated experiences like this erode organizational confidence in AI and—in critical settings such as health—may erode broader public trust in the technology as well. 

When current benchmarks provide only a partial and potentially misleading signal of an AI model’s readiness for real-world use, this creates regulatory blind spots: Oversight is shaped by metrics that do not reflect reality. It also leaves organizations and governments to shoulder the risks of testing AI in sensitive real-world settings, often with limited resources and support. 

How to build better tests 

To close the gap between benchmark and real-world performance, we must pay attention to the actual conditions in which AI models will be used. The critical questions: Can AI function as a productive participant within human teams? And can it generate sustained, collective value? 

Through my research on AI deployment across multiple sectors, I have seen a number of organizations already moving—deliberately and experimentally—toward the HAIC benchmarks I favor. 

HAIC benchmarks reframe current benchmarking in four ways: 

1.     From individual and single-task performance to team and workflow performance (shifting the unit of analysis)

2.     From one-off testing with right/wrong answers to long-term impacts (expanding the time horizon)

3.     From correctness and speed to organizational outcomes, coordination quality, and error detectability (expanding outcome measures)

4.     From isolated outputs to upstream and downstream consequences (system effects)

Across the organizations where this approach has emerged and started to be applied, the first step is shifting the unit of analysis. 

For example, in one UK hospital system in the period 2021–2024, the question expanded from whether a medical AI application improves diagnostic accuracy to how the presence of AI within the hospital’s multidisciplinary teams affects not only accuracy but also coordination and deliberation. The hospital specifically assessed coordination and deliberation in human teams using and not using AI. Multiple stakeholders (within and outside the hospital) decided on metrics like how AI influences collective reasoning, whether it surfaces overlooked considerations, whether it strengthens or weakens coordination, and whether it changes established risk and compliance practices. 

This shift is fundamental. It matters a lot in high-stakes contexts where system-level effects matter more than task-level accuracy. It also matters for the economy. It may help recalibrate inflated expectations of sweeping productivity gains that are so far predicated largely on the promise of improving individual task performance. 

Once that foundation is set, HAIC benchmarking can begin to take on the element of time. 

Today’s benchmarks resemble school exams—one-off, standardized tests of accuracy. But real professional competence is assessed differently. Junior doctors and lawyers are evaluated continuously inside real workflows, under supervision, with feedback loops and accountability structures. Performance is judged over time and in a specific context, because competence is relational. If AI systems are meant to operate alongside professionals, their impact should be judged longitudinally, reflecting how performance unfolds over repeated interactions. 

I saw this aspect of HAIC applied in one of my humanitarian-sector case studies. Over 18 months, an AI system was evaluated within real workflows, with particular attention to how detectable its errors were—that is, how easily human teams could identify and correct them. This long-term “record of error detectability” meant the organizations involved could design and test context-specific guardrails to promote trust in the system, despite the inevitability of occasional AI mistakes.

A longer time horizon also makes visible the system-level consequences that short-term benchmarks miss. An AI application may outperform a single doctor on a narrow diagnostic task yet fail to improve multidisciplinary decision-making. Worse, it may introduce systemic distortions: anchoring teams too early in plausible but incomplete answers, adding to people’s  cognitive workloads, or generating downstream inefficiencies that offset any speed or efficiency gains at the point of the AI’s use. These knock-on effects—often invisible to current benchmarks—are central to understanding real impact. 

The HAIC approach, admittedly promises to make benchmarking more complex, resource-intensive, and harder to standardize. But continuing to evaluate AI in sanitized conditions detached from the world of work will leave us misunderstanding what it truly can and cannot do for us. To deploy AI responsibly in real-world settings, we must measure what actually matters: not just what a model can do alone, but what it enables—or undermines—when humans and teams in the real world work with it.

 Angela Aristidou is a professor at University College London and a faculty fellow at the Stanford Digital Economy Lab and the Stanford Human-Centered AI Institute. She speaks, writes, and advises about the real-life deployment of artificial-intelligence tools for public good.

Earlier this month, Microsoft launched Copilot Health, a new space within its Copilot app where users will be able to connect their medical records and ask specific questions about their health. A couple of days earlier, Amazon had announced that Health AI, an LLM-based tool previously restricted to members of its One Medical service, would now be widely available. These products join the ranks of ChatGPT Health, which OpenAI released back in January, and Anthropic’s Claude, which can access user health records if granted permission. Health AI for the masses is officially a trend. 

There’s a clear demand for chatbots that provide health advice, given how hard it is for many people to access it through existing medical systems. And some research suggests that current LLMs are capable of making safe and useful recommendations. But researchers say that these tools should be more rigorously evaluated by independent experts, ideally before they are widely released. 

In a high-stakes area like health, trusting companies to evaluate their own products could prove unwise, especially if those evaluations aren’t made available for external expert review. And even if the companies are doing quality, rigorous research—which some, including OpenAI, do seem to be—they might still have blind spots that the broader research community could help to fill.

“To the extent that you always are going to need more health care, I think we should definitely be chasing every route that works,” says Andrew Bean, a doctoral candidate at the Oxford Internet Institute. “It’s entirely plausible to me that these models have reached a point where they’re actually worth rolling out.”

“But,” he adds, “the evidence base really needs to be there.”

Tipping points 

To hear developers tell it, these health products are now being released because large language models have indeed reached a point where they can effectively provide medical advice. Dominic King, the vice president of health at Microsoft AI and a former surgeon, cites AI advancement as a core reason why the company’s health team was formed, and why Copilot Health now exists. “We’ve seen this enormous progress in the capabilities of generative AI to be able to answer health questions and give good responses,” he says.

But that’s only half the story, according to King. The other key factor is demand. Shortly before Copilot Health was launched, Microsoft published a report, and an accompanying blog post, detailing how people used Copilot for health advice. The company says it receives 50 million health questions each day, and health is the most popular discussion topic on the Copilot mobile app.

Other AI companies have noticed, and responded to, this trend. “Even before our health products, we were seeing just a rapid, rapid increase in the rate of people using ChatGPT for health-related questions,” says Karan Singhal, who leads OpenAI’s Health AI team. (OpenAI and Microsoft have a long-standing partnership, and Copilot is powered by OpenAI’s models.)

It’s possible that people simply prefer posing their health problems to a nonjudgmental bot that’s available to them 24-7. But many experts interpret this pattern in light of the current state of the health-care system. “There is a reason that these tools exist and they have a position in the overall landscape,” says Girish Nadkarni, chief AI officer​ at the Mount Sinai Health System. “That’s because access to health care is hard, and it’s particularly hard for certain populations.”

The virtuous vision of consumer-facing LLM health chatbots hinges on the possibility that they could improve user health while reducing pressure on the health-care system. That might involve helping users decide whether or not they need medical attention, a task known as triage. If chatbot triage works, then patients who need emergency care might seek it out earlier than they would have otherwise, and patients with more mild concerns might feel comfortable managing their symptoms at home with the chatbot’s advice rather than unnecessarily busying emergency rooms and doctor’s offices.

But a recent, widely discussed study from Nadkarni and other researchers at Mount Sinai found that ChatGPT Health sometimes recommends too much care for mild conditions and fails to identify emergencies. Though Singhal and  some other experts have suggested that its methodology might not provide a complete picture of ChatGPT Health’s capabilities, the study has surfaced concerns about how little external evaluation these tools see before being released to the public.

Most of the academic experts interviewed for this piece agreed that LLM health chatbots could have real upsides, given how little access to health care some people have. But all six of them expressed concerns that these tools are being launched without testing from independent researchers to assess whether they are safe. While some advertised uses of these tools, such as recommending exercise plans or suggesting questions that a user might ask a doctor, are relatively harmless, others carry clear risks. Triage is one; another is asking a chatbot to provide a diagnosis or a treatment plan. 

The ChatGPT Health interface includes a prominent disclaimer stating that it is not intended for diagnosis or treatment, and the announcements for Copilot Health and Amazon’s Health AI include similar warnings. But those warnings are easy to ignore. “We all know that people are going to use it for diagnosis and management,” says Adam Rodman, an internal medicine physician and researcher at Beth Israel Deaconess Medical Center and a visiting researcher at Google.

Medical testing

Companies say they are testing the chatbots to ensure that they provide safe responses the vast majority of the time. OpenAI has designed and released HealthBench, a benchmark that scores LLMs on how they respond in realistic health-related conversations—though the conversations themselves are LLM-generated. When GPT-5, which powers both ChatGPT Health and Copilot Health, was released last year, OpenAI reported the model’s HealthBench scores: It did substantially better than previous OpenAI models, though its overall performance was far from perfect. 

But evaluations like HealthBench have limitations. In a study published last month, Bean—the Oxford doctoral candidate—and his colleagues found that even if an LLM can accurately identify a medical condition from a fictional written scenario on its own, a non-expert user who is given the scenario and asked to determine the condition with LLM assistance might figure it out only a third of the time. If they lack medical expertise, users might not know which parts of a scenario—or their real-life experience—are important to include in their prompt, or they might misinterpret the information that an LLM gives them.

Singhal, the OpenAI health lead, notes that the company’s current GPT-5 series of models, which had not yet been released when the original HealthBench study was conducted, do a much better job of soliciting additional information than their predecessors. However, OpenAI has reported that GPT-5.4, the current flagship, is actually worse at seeking context than GPT-5.2, an earlier version.

Ideally, Bean says, health chatbots would be subjected to controlled tests with human users, as they were in his study, before being released to the public. That might be a heavy lift, particularly given how fast the AI world moves and how long human studies can take. Bean’s own study used GPT-4o, which came out almost a year ago and is now outdated. 

Earlier this month, Google released a study that meets Bean’s standards. In the study, patients discussed medical concerns with the company’s Articulate Medical Intelligence Explorer (AMIE), a medical LLM chatbot that is not yet available to the public, before meeting with a human physician. Overall, AMIE’s diagnoses were just as accurate as physicians’, and none of the conversations raised major safety concerns for researchers. 

Despite the encouraging results, Google isn’t planning to release AMIE anytime soon. “While the research has advanced, there are significant limitations that must be addressed before real-world translation of systems for diagnosis and treatment, including further research into equity, fairness, and safety testing,” wrote Alan Karthikesalingam, a research scientist at Google DeepMind, in an email. Google did recently reveal that Health100, a health platform it is building in partnership with CVS, will include an AI assistant powered by its flagship Gemini models, though that tool will presumably not be intended for diagnosis or treatment.

Rodman, who led the AMIE study with Karthikesalingam, doesn’t think such extensive, multiyear studies are necessarily the right approach for chatbots like ChatGPT Health and Copilot Health. “There’s lots of reasons that the clinical trial paradigm doesn’t always work in generative AI,” he says. “And that’s where this benchmarking conversation comes in. Are there benchmarks [from] a trusted third party that we can agree are meaningful, that the labs can hold themselves to?”

They key there is “third party.” No matter how extensively companies evaluate their own products, it’s tough to trust their conclusions completely. Not only does a third-party evaluation bring impartiality, but if there are many third parties involved, it also helps protect against blind spots.

OpenAI’s Singhal says he’s strongly in favor of external evaluation. “We try our best to support the community,” he says. “Part of why we put out HealthBench was actually to give the community and other model developers an example of what a very good evaluation looks like.” 

Given how expensive it is to produce a high-quality evaluation, he says, he’s skeptical that any individual academic laboratory would be able to produce what he calls “the one evaluation to rule them all.” But he does speak highly of efforts that academic groups have made to bring preexisting and novel evaluations together into comprehensive evaluations suites—such as Stanford’s MedHELM framework, which tests models on a wide variety of medical tasks. Currently, OpenAI’s GPT-5 holds the highest MedHELM score.

Nigam Shah, a professor of medicine at Stanford University who led the MedHELM project, says it has limitations. In particular, it only evaluates individual chatbot responses, but someone who’s seeking medical advice from a chatbot tool might engage it in a multi-turn, back-and-forth conversation. He says that he and some collaborators are gearing up to build an evaluation that can score those complex conversations, but that it will take time, and money. “You and I have zero ability to stop these companies from releasing [health-oriented products], so they’re going to do whatever they damn please,” he says. “The only thing people like us can do is find a way to fund the benchmark.”

No one interviewed for this article argued that health LLMs need to perform perfectly on third-party evaluations in order to be released. Doctors themselves make mistakes—and for someone who has only occasional access to a doctor, a consistently accessible LLM that sometimes messes up could still be a huge improvement over the status quo, as long as its errors aren’t too grave. 

With the current state of the evidence, however, it’s impossible to know for sure whether the currently available tools do in fact constitute an improvement, or whether their risks outweigh their benefits.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

Last Thursday, a California judge temporarily blocked the Pentagon from labeling Anthropic a supply chain risk and ordering government agencies to stop using its AI. It’s the latest development in the month-long feud. And the matter still isn’t settled: The government was given seven days to appeal, and Anthropic has a second case against the designation that has yet to be decided. Until then, the company remains persona non grata with the government. 

The stakes in the case—how much the government can punish a company for not playing ball—were apparent from the start. Anthropic drew lots of senior supporters with unlikely bedfellows among them, including former authors of President Trump’s AI policy.

But Judge Rita Lin’s 43-page opinion suggests that what is really a contract dispute never needed to reach such a frenzy. It did so because the government disregarded the existing process for how such disputes are governed and fueled the fire with social media posts from officials that would eventually contradict the positions it took in court. The Pentagon, in other words, wanted a culture war (on top of the actual war in Iran that began hours later). 

The government used Anthropic’s Claude for much of 2025 without complaint, according to court documents, while the company walked a branding tightrope as a safety-focused AI company that also won defense contracts. Defense employees accessing it through Palantir were required to accept terms of a government-specific usage policy that Anthropic cofounder Jared Kaplan said “prohibited mass surveillance of Americans and lethal autonomous warfare” (Kaplan’s declaration to the court didn’t include details of the policy). Only when the government aimed to contract with Anthropic directly did the disagreements begin. 

What drew the ire of the judge is that when these disagreements became public, they had more to do with punishment than just cutting ties with Anthropic. And they had a pattern: Tweet first, lawyer later. 

President Trump’s post on Truth Social on February 27 referenced “Leftwing nutjobs” at Anthropic and directed every federal agency to stop using the company’s AI. This was echoed soon after by Defense Secretary Pete Hegseth, who said he’d direct the Pentagon to label Anthropic a supply chain risk. 

Doing so necessitates that the secretary take a specific set of actions, which the judge found Hegseth did not complete. Letters sent to congressional committees, for example, said that less drastic steps were evaluated and deemed not possible, without providing any further details. The government also said the designation as a supply chain risk was necessary because Anthropic could implement a “kill switch,” but its lawyers later had to admit it had no evidence of that, the judge wrote.

Hegseth’s post also stated that “No contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.” But the government’s own lawyers admitted on Tuesday that the Secretary doesn’t have the power to do that, and agreed with the judge that the statement had “absolutely no legal effect at all.”

The aggressive posts also led the judge to also conclude that Anthropic was on solid ground in complaining that its First Amendment rights were violated. The government, the judge wrote while citing the posts, “set out to publicly punish Anthropic for its ‘ideology’ and ‘rhetoric,’ as well as its ‘arrogance’ for being unwilling to compromise those beliefs.”

Labeling Anthropic a supply chain risk would essentially be identifying it as a “saboteur” of the government, for which the judge did not see sufficient evidence. She issued an order last Thursday halting the designation, preventing the Pentagon from enforcing it and forbidding the government from fulfilling the promises made by Hegseth and Trump. Dean Ball, who worked on AI policy for the Trump administration but wrote a brief supporting Anthropic, described the judge’s order on Thursday as “a devastating ruling for the government, finding Anthropic likely to prevail on essentially all of its theories for why the government’s actions were unlawful and unconstitutional.”

The government is expected to appeal the decision. But Anthropic’s separate case, filed in DC, makes similar allegations. It just references a different segment of the law governing supply chain risks. 

The court documents paint a pretty clear pattern. Public statements made by officials and the President did not at all align with what the law says should happen in a contract dispute like this, and the government’s lawyers have consistently had to create justifications for social media lambasting of the company after the fact.

Pentagon and White House leadership knew that pursuing the nuclear option would spark a court battle; Anthropic vowed on February 27 to fight the supply chain risk designation days before the government formally filed it on March 3. Pursuing it anyway meant senior leadership was, to say the least, distracted during the first five days of the Iran war, launching strikes while also compiling evidence that Anthropic was a saboteur to the government, all while it could have cut ties with Anthropic by simpler means. 

But even if Anthropic ultimately wins, the government has other means to shun the company from government work. Defense contractors who want to stay on good terms with the Pentagon, for example, now have little reason to work with Anthropic even if it’s not flagged as a supply chain risk. 

“I think it’s safe to say that there are mechanisms the government can use to apply some degree of pressure without breaking the law,” says Charlie Bullock, a senior research fellow at the Institute for Law and AI. “It kind of depends how invested the government is in punishing Anthropic.”

From the evidence thus far, the administration is committing top-level time and attention to winning an AI culture war. At the same time, Claude is apparently so important to its operations that even President Trump said the Pentagon needed six months to stop using it. The White House demands political loyalty and ideological alignment from top AI companies, But the case against Anthropic, at least for now, exposes the limits of its leverage.

If you have information about the military’s use of AI, you can share it securely via Signal (username jamesodonnell.22).