Background: Artificial intelligence (AI) is increasingly used in digital health, particularly through large language models (LLMs), to support patient engagement and behavior change. One novel application is the delivery of motivational interviewing (MI), an evidence-based, patient-centered counseling technique designed to enhance motivation and resolve ambivalence around health behaviors. AI tools, including chatbots, mobile apps, and web-based agents, are being developed to simulate MI techniques at scale. While these innovations are promising, important questions remain about how faithfully AI systems can replicate MI principles or achieve meaningful behavioral impact. Objective: This scoping review aimed to summarize existing empirical studies evaluating AI-driven systems that apply MI techniques to support health behavior change. Specifically, we examined the feasibility of these systems; their fidelity to MI principles; and their reported behavioral, psychological, or engagement outcomes. Methods: We systematically searched PubMed, Embase, Scopus, Web of Science, and Cochrane Library for empirical studies published between January 1, 2018, and February 25, 2025. Eligible studies involved AI-driven systems using natural language generation, understanding, or computational logic to deliver MI techniques to users targeting a specific health behavior. We excluded studies using AI solely for training clinicians in MI. Three independent reviewers screened and extracted data on study design, AI modality and type, MI components, health behavior focus, MI fidelity assessment, and outcome domains. Results: Of the 1001 records identified, 15 (1.5%) met the inclusion criteria. Of these 15 studies, 6 (40%) were exploratory feasibility or pilot studies, and 3 (20%) were randomized controlled trials. AI modalities included rule-based chatbots (9/15, 60%), LLM-based systems (4/15, 27%), and virtual or mobile agents (2/15, 13%). Targeted behaviors included smoking cessation (6/15, 40%), substance use (3/15, 20%), COVID-19 vaccine hesitancy, type 2 diabetes self-management, stress, mental health service use, and opioid use during pregnancy. Of the 15 studies, 13 (87%) reported positive findings on feasibility or user acceptability, while 6 (40%) assessed MI fidelity using expert review or structured coding, with moderate to high alignment reported. Several studies found that users perceived the AI systems as judgment free, supportive, and easier to engage with than human counselors, particularly in stigmatized contexts. However, limitations in empathy, safety transparency, and emotional nuance were commonly noted. Only 3 (20%) of the 15 studies reported substantially significant behavioral changes. Conclusions: AI systems delivering MI show promise for enhancing patient engagement and scaling behavior change interventions. Early evidence supports their usability and partial fidelity to MI principles, especially in sensitive domains. However, most systems remain in early development, and few have been rigorously tested. Future research should prioritize randomized evaluations; standardized fidelity measures; and safeguards for LLM safety, empathy, and accuracy in health-related dialogue. Trial Registration: OSF Registries 10.17605/OSF.IO/G9N7E; https://osf.io/g9n7e

Background: Determining effective nurse staffing levels is crucial for ensuring quality patient care and operational efficiency within hospitals. Traditional workload prediction methods often rely on professional judgment or simple volume-based approaches, which can be inaccurate. Machine learning offers a promising avenue for more data-driven and precise predictions, by using historical nursing workload data to forecast future patient care requirements, which could help with staff planning while also improving patient outcomes and nurse well-being. Objective: This methodological study aims to use nursing activity data, specifically LEP (Leistungserfassung in der Pflege; “documentation of nursing activities”), to predict future workload requirements using machine learning techniques. Methods: We conducted a retrospective observational study at the University Hospital of Zürich, using nursing workload data for inpatients across eight wards, collected between 2017 and 2021. Data were transformed to represent nursing workload per ward and shift, with three shifts per day. Variables used in modeling included historical workload trends, patient characteristics, and upcoming operations. Machine learning models, including linear regression variants and tree-based methods (Random Forest and XGBoost), were trained and tested on this dataset to predict workload 72 hours in advance, on a shift-by-shift basis. Model performance was assessed using mean absolute error (MAE) and mean absolute percentage error (MAPE), and results were compared against a baseline of assuming no change in workload from the time of prediction. Prediction accuracy was further evaluated by categorizing future workload changes into decreased, similar, or increased workload relative to current shift levels. Results: Our findings demonstrate that machine learning models consistently outperform the baseline across all wards. The best-performing model was the Lasso Regression model, which achieved an average improvement in accuracy of 25.0% compared to the baseline. When used to predict upcoming changes in workload levels, the model achieved strong classification performance, giving an average AUROC of 0.79 and precision values between 66.2% and 75.3%. Crucially, the model severely misclassified—predicting an upcoming increase as a decrease, and vice versa—in just 0.17% of cases, highlighting potential reliability for using the model in practice. Key variables identified as important for predictions include historical shift workload averages and overall ward workload trends. Conclusions: This study suggests the potential of machine learning to enhance nurse workload prediction, while highlighting the need for refinement. Limitations due to potential discrepancies between recorded nursing activities and the actual workload highlight the need for further investigation into data quality. To maximize impact, future research should focus on: 1) utilizing more diverse data, 2) more advanced machine learning architecture that perform time-series modelling, 3) addressing data quality concerns, and 4) conducting controlled trials for real-world evaluation.

Background: Adolescent idiopathic scoliosis is a common spinal deformity typically treated with exercise therapy. Despite the increasing use of digital technologies in interventions, there remains a gap in understanding how to effectively integrate behavior change techniques (BCTs) and behavior theories within these digital solutions. Objective: This review aims to identify the digital characteristics of interventions and the BCTs used, and to analyze potential theoretical mechanisms with the Theoretical Domains Framework and the capability, opportunity, motivation, and behavior model. Methods: We conducted a scoping review according to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines. A total of 5 databases, including PubMed, Web of Science, Embase, Cochrane Library, and CINAHL, were selected for screening eligible studies up to April 4, 2024. We included studies of any design type that involved patients with adolescent idiopathic scoliosis using digital interventions for exercise rehabilitation, including qualitative, quantitative, or mixed methods studies, and study protocols with detailed descriptions of digital interventions. Two researchers independently screened studies and extracted data into tables for descriptive analysis. The Mixed Methods Appraisal Tool was used to assess the quality of studies. Results: Out of the 3267 identified papers, 21 (0.64%) studies were included. The most frequently used technologies were videoconferencing (n=7) and instructional videos (n=5). The three most common BCT clusters were “Shaping Knowledge” (n=19), “Social Support” (n=16), and “Antecedents” (n=16). “Knowledge” was the most used mechanism of action (n=21), followed by “Skills” (n=16), “Environmental Context and Resources” (n=16), and “Social Influences” (n=16). The studies primarily addressed “Capability” and “Opportunity,” with less emphasis on “Motivation,” particularly “Automatic Motivation.” Conclusions: This review identified common digital technologies and their characteristics, analyzed potential mechanisms of behavior change in interventions, and provided recommendations for technology utilization. Future research should further evaluate the effectiveness of digital technologies while enhancing patient motivation and user experience. Trial Registration: PROSPERO CRD42024530851; https://www.crd.york.ac.uk/PROSPERO/view/CRD42024530851

Nature Biotechnology, Published online: 15 September 2025; doi:10.1038/s41587-025-02822-6

AI designs de novo antibiotics

Called the Agent Payments Protocol (AP2), the system is meant to be interoperable between AI platforms, payment systems and vendors.

Researchers and health care institutions have increasingly applied structured futures methods—such as the futures wheel, scenario analysis, forecasting, and horizon scanning—to systematically explore, generate, and prepare for multiple possible futures. However, discussions around the future of medicine, specialties, or therapeutic areas have often relied on the subjective opinions or perspectives of key opinion leaders rather than on future strategies, policies, visions, and scenarios that are grounded in rigorous and established methods. This underscores the need for futures methods to be widely adopted and effectively incorporated into both medical practice and health care policymaking. Integrating structured foresight techniques into strategic planning enables clinicians and policymakers to transition from reactive decision-making to proactive, plausible approaches that shape a more resilient and adaptive health care system. Our goal with this paper is to provide a methodological guide that is supported by case studies, demonstrating how futures methods can be systematically applied in health care. By offering practical examples, we intend to empower medical professionals, health care leaders, researchers, patients, and policymakers with the tools to anticipate and navigate future challenges and opportunities more effectively.

Nature, Published online: 15 September 2025; doi:10.1038/d41586-025-02940-w

Developers of griefbots say that they help people by allowing them to commune with recreations of the dead, but others say that the technology is fraught with danger.

Background: Even though Digital Health Technology (DHT) is widely utilized in the United States (U.S.) at both hospital provider and individual levels, it is beset with several challenges that have contributed to inequities in the health service delivery. Previous studies have shown that health inequities observed may be amplified many by DHT requirements. Objective: The objectives of this scoping review are aimed at synthesizing information on DHT inequities by exploring evidence that describes DHT infrastructure needs focused on promoting health equity in the U.S. and identifying key challenges at both the individual/patient level and at the health service provider's level. Methods: We adapted Arksey and O'Malley's scoping review guidelines in our review. We searched PubMed, Web of Science, CINAHL, and PsycINFO were searched. We also conducted supplementary searches on Google Scholar. The inclusion criteria were peer-reviewed publications that broadly conceptualize or analyze DHT infrastructure from a health equity perspective and the challenges of DHT requirements between 2020 and 2024. Following a full-text screening using eligibility criteria such as studies were included if they examined DHT infrastructure in the U.S. from a health equity perspective, discussed health disparities resulting from DHT interventions, or investigated the variables influencing health inequities connected to DHT. Two researchers evaluated each citation’s individually at the title and abstract levels. Thematic approach and qualitative analysis determined this scoping review’s outcome. Results: Of the 628 research articles from the search, 27 were included in the analysis based on the inclusion criteria. In this review, we discussed factors such as elderly population, education, race, ethnicity, and socioeconomic status leading to health inequities in DHT. Patients and Service providers challenges that exist in health inequities related to DHT. The most common challenges for service providers were infrastructure and technical issues such as inadequate integration with existing workflows, user-unfriendly health information exchange (HIE) interfaces, and lack of skilled staff, while for individuals or patients, this included limited broadband internet access, cultural or linguistic appropriateness, and access to digital tools. Conclusions: The study identified that in the U.S., DHT is an essential part of the delivery of health services, yet it is saddled with key challenges leading to health inequities. Finding pragmatic solutions to these challenges can improve health equity in DHT.

Large language models (LLMs), such as OpenAI’s GPT series and Google’s PaLM, are transforming healthcare by improving clinical decision-making, enhancing patient communication, and simplifying administrative tasks. However, their performance relies heavily on prompt design, where small changes in wording or structure can greatly impact output quality. This poses a challenge for clinicians who are not experts in natural language processing (NLP). This tutorial combines prompt engineering techniques tailored for clinical use, covering methods like zero-shot, few-shot, chain-of-thought, and meta-prompting. We examine four critical dimensions (accuracy, bias mitigation, privacy protection, and workflow integration) through clinical case studies grounded in real-world practice. We provide actionable guidance on defining objectives, applying core principles, iteratively refining prompts, and integrating them into interoperable electronic health record (EHR) systems. This framework helps clinicians leverage LLMs to improve decision-making, streamline documentation, and enhance patient communication while maintaining ethical standards and ensuring patient safety.

Hugging Face has unveiled FinePDFs, the largest publicly available corpus built entirely from PDFs. The dataset spans 475 million documents in 1,733 languages, totaling roughly 3 trillion tokens. At 3.65 terabytes in size, FinePDFs introduces a new dimension to open training datasets by tapping into a resource long considered too complex and expensive to process.

By Robert Krzaczyński

Nature Medicine, Published online: 15 September 2025; doi:10.1038/s41591-025-03965-4

A brief, self-administered digital cognitive test, in combination with a blood test, accurately detects clinical Alzheimer’s disease in primary care.

Nature Medicine, Published online: 11 September 2025; doi:10.1038/s41591-025-03936-9

In an irrevocably changed landscape, reform of the global health system needs to answer key questions on functions, what should be delivered in different contexts and at different levels, and how the system should operate.

Background: In modern, high-speed work settings, the significance of mental health disorders is increasingly acknowledged as a pressing health issue, with potential adverse consequences for organizations, including reduced productivity and increased absenteeism. Over the past few years, various mental health management solutions, such as biofeedback applications, have surfaced as promising avenues to improve employees’ mental well-being. However, most studies on these interventions have been conducted in controlled laboratory settings. Objective: This review aimed to systematically identify and analyze studies that implemented biofeedback-based interventions in real-world occupational settings, focusing on their effectiveness in improving psychological well-being and mental health. Methods: A systematic review was conducted following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. We searched PubMed and EBSCO databases for studies published between 2012 and 2024. Inclusion criteria were original peer-reviewed studies that focused on employees and used biofeedback interventions to improve mental health or prevent mental illness. Exclusion criteria included nonemployee samples, lack of a description of the intervention, and low methodological quality (assessed using the Physiotherapy Evidence Database [PEDro] checklist). Data were extracted on study characteristics, intervention type, physiological and self-reported outcomes, and follow-up measures. Risk of bias was assessed, and VOSviewer was used to visualize the distribution of research topics. Results: A total of 9 studies met the inclusion criteria. The interventions used a range of delivery methods, including traditional biofeedback, mobile apps, mindfulness techniques, virtual reality, and cerebral blood flow monitoring. Most studies focused on breathing techniques to regulate physiological responses (eg, heart rate variability and respiratory sinus arrhythmia) and showed reductions in stress, anxiety, and depressive symptoms. Mobile and app-directed interventions appeared particularly promising for improving resilience and facilitating recovery after stress. Of the 9 studies, 8 (89%) reported positive outcomes, with 1 (11%) study showing initial increases in stress due to logistical limitations in biofeedback access. Sample sizes were generally small, and long-term follow-up data were limited. Conclusions: Biofeedback interventions in workplace settings show promising short-term results in reducing stress and improving mental health, particularly when incorporating breathing techniques and user-friendly delivery methods such as mobile apps. However, the field remains underexplored in occupational contexts. Future research should address adherence challenges, scalability, cost-effectiveness, and long-term outcomes to support broader implementation of biofeedback as a sustainable workplace mental health strategy.

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.

It’s been a big year for video generation. In the last nine months OpenAI made Sora public, Google DeepMind launched Veo 3, the video startup Runway launched Gen-4. All can produce video clips that are (almost) impossible to distinguish from actual filmed footage or CGI animation. This year also saw Netflix debut an AI visual effect in its show The Eternaut, the first time video generation has been used to make mass-market TV.

Sure, the clips you see in demo reels are cherry-picked to showcase a company’s models at the top of their game. But with the technology in the hands of more users than ever before—Sora and Veo 3 are available in the ChatGPT and Gemini apps for paying subscribers—even the most casual filmmaker can now knock out something remarkable. 

The downside is that creators are competing with AI slop, and social media feeds are filling up with faked news footage. Video generation also uses up a huge amount of energy, many times more than text or image generation. 

With AI-generated videos everywhere, let’s take a moment to talk about the tech that makes them work.

How do you generate a video?

Let’s assume you’re a casual user. There are now a range of high-end tools that allow pro video makers to insert video generation models into their workflows. But most people will use this technology in an app or via a website. You know the drill: “Hey, Gemini, make me a video of a unicorn eating spaghetti. Now make its horn take off like a rocket.” What you get back will be hit or miss, and you’ll typically need to ask the model to take another pass or 10 before you get more or less what you wanted. 

So what’s going on under the hood? Why is it hit or miss—and why does it take so much energy? The latest wave of video generation models are what’s known as latent diffusion transformers. Yes, that’s quite a mouthful. Let’s unpack each part in turn, starting with diffusion. 

What’s a diffusion model?

Imagine taking an image and adding a random spattering of pixels to it. Take that pixel-spattered image and spatter it again and then again. Do that enough times and you will have turned the initial image into a random mess of pixels, like static on an old TV set. 

A diffusion model is a neural network trained to reverse that process, turning random static into images. During training, it gets shown millions of images in various stages of pixelation. It learns how those images change each time new pixels are thrown at them and, thus, how to undo those changes. 

The upshot is that when you ask a diffusion model to generate an image, it will start off with a random mess of pixels and step by step turn that mess into an image that is more or less similar to images in its training set. 

But you don’t want any image—you want the image you specified, typically with a text prompt. And so the diffusion model is paired with a second model—such as a large language model (LLM) trained to match images with text descriptions—that guides each step of the cleanup process, pushing the diffusion model toward images that the large language model considers a good match to the prompt. 

An aside: This LLM isn’t pulling the links between text and images out of thin air. Most text-to-image and text-to-video models today are trained on large data sets that contain billions of pairings of text and images or text and video scraped from the internet (a practice many creators are very unhappy about). This means that what you get from such models is a distillation of the world as it’s represented online, distorted by prejudice (and pornography).

It’s easiest to imagine diffusion models working with images. But the technique can be used with many kinds of data, including audio and video. To generate movie clips, a diffusion model must clean up sequences of images—the consecutive frames of a video—instead of just one image. 

What’s a latent diffusion model? 

All this takes a huge amount of compute (read: energy). That’s why most diffusion models used for video generation use a technique called latent diffusion. Instead of processing raw data—the millions of pixels in each video frame—the model works in what’s known as a latent space, in which the video frames (and text prompt) are compressed into a mathematical code that captures just the essential features of the data and throws out the rest. 

A similar thing happens whenever you stream a video over the internet: A video is sent from a server to your screen in a compressed format to make it get to you faster, and when it arrives, your computer or TV will convert it back into a watchable video. 

And so the final step is to decompress what the latent diffusion process has come up with. Once the compressed frames of random static have been turned into the compressed frames of a video that the LLM guide considers a good match for the user’s prompt, the compressed video gets converted into something you can watch.  

With latent diffusion, the diffusion process works more or less the way it would for an image. The difference is that the pixelated video frames are now mathematical encodings of those frames rather than the frames themselves. This makes latent diffusion far more efficient than a typical diffusion model. (Even so, video generation still uses more energy than image or text generation. There’s just an eye-popping amount of computation involved.) 

What’s a latent diffusion transformer?

Still with me? There’s one more piece to the puzzle—and that’s how to make sure the diffusion process produces a sequence of frames that are consistent, maintaining objects and lighting and so on from one frame to the next. OpenAI did this with Sora by combining its diffusion model with another kind of model called a transformer. This has now become standard in generative video. 

Transformers are great at processing long sequences of data, like words. That has made them the special sauce inside large language models such as OpenAI’s GPT-5 and Google DeepMind’s Gemini, which can generate long sequences of words that make sense, maintaining consistency across many dozens of sentences. 

But videos are not made of words. Instead, videos get cut into chunks that can be treated as if they were. The approach that OpenAI came up with was to dice videos up across both space and time. “It’s like if you were to have a stack of all the video frames and you cut little cubes from it,” says Tim Brooks, a lead researcher on Sora.

Using transformers alongside diffusion models brings several advantages. Because they are designed to process sequences of data, transformers also help the diffusion model maintain consistency across frames as it generates them. This makes it possible to produce videos in which objects don’t pop in and out of existence, for example. 

And because the videos are diced up, their size and orientation do not matter. This means that the latest wave of video generation models can be trained on a wide range of example videos, from short vertical clips shot with a phone to wide-screen cinematic films. The greater variety of training data has made video generation far better than it was just two years ago. It also means that video generation models can now be asked to produce videos in a variety of formats. 

What about the audio? 

A big advance with Veo 3 is that it generates video with audio, from lip-synched dialogue to sound effects to background noise. That’s a first for video generation models. As Google DeepMind CEO Demis Hassabis put it at this year’s Google I/O: “We’re emerging from the silent era of video generation.” 

The challenge was to find a way to line up video and audio data so that the diffusion process would work on both at the same time. Google DeepMind’s breakthrough was a new way to compress audio and video into a single piece of data inside the diffusion model. When Veo 3 generates a video, its diffusion model produces audio and video together in a lockstep process, ensuring that the sound and images are synched.  

You said that diffusion models can generate different kinds of data. Is this how LLMs work too? 

No—or at least not yet. Diffusion models are most often used to generate images, video, and audio. Large language models—which generate text (including computer code)—are built using transformers. But the lines are blurring. We’ve seen how transformers are now being combined with diffusion models to generate videos. And this summer Google DeepMind revealed that it was building an experimental large language model that used a diffusion model instead of a transformer to generate text. 

Here’s where things start to get confusing: Though video generation (which uses diffusion models) consumes a lot of energy, diffusion models themselves are in fact more efficient than transformers. Thus, by using a diffusion model instead of a transformer to generate text, Google DeepMind’s new LLM could be a lot more efficient than existing LLMs. Expect to see more from diffusion models in the near future!

The $300B deal is a reminder that despite Oracle’s legacy status, it shouldn’t be overlooked when it comes to AI infrastructure. But key questions around power and how OpenAI will pay for this remain.

Nature, Published online: 11 September 2025; doi:10.1038/d41586-025-02936-6

Authors and peer reviewers are failing to disclose the use of LLMs despite journal policies limiting their use.

Nature, Published online: 10 September 2025; doi:10.1038/s41586-025-09374-4

Cancer evolutionary dynamics are quantitatively inferred using a method, EVOFLUx, applied to fluctuating DNA methylation.