As companies race to implement AI, many are finding that project success hinges directly on the quality of their data. This dependency is causing many ambitious initiatives to stall, never making it beyond the experimental proof-of-concept stage.

So, what’s the secret to turning these experiments into real revenue generators? AI News caught up with Martin Frederik, regional leader for the Netherlands, Belgium, and Luxembourg at data cloud giant Snowflake, to find out.

“There’s no AI strategy without a data strategy,” Frederik says simply. “AI apps, agents, and models are only as effective as the data they’re built on, and without unified, well-governed data infrastructure, even the most advanced models can fall short.”

Improving data quality is key to AI project success

It’s a familiar story for many organisations: a promising proof-of-concept impresses the team but never translates into a tool that makes the company money. According to Frederik, this often happens because leaders treat the technology as the end goal.

“AI is not the destination – it’s the vehicle to achieving your business goals,” Frederik advises.

When projects get stuck, it’s usually down to a few common culprits: the project isn’t truly aligned with what the business needs, teams aren’t talking to each other, or the data is a mess. It’s easy to get disheartened by statistics suggesting that 80% of AI projects don’t reach production, but Frederik offers a different perspective. This isn’t necessarily a failure, he suggests, but “part of the maturation process”.

For those who get the foundation right, the payoff is very real. A recent Snowflake study found that 92% of companies are already seeing a return on their AI investments. In fact, for every £1 spent, they’re getting back £1.41 in cost savings and new revenue. The key, Frederik repeats, is having a “secure, governed and centralised platform” for your data from the very beginning.

It’s not just about tech, it’s about people

Even with the best technology, an AI strategy can fall flat if the company culture isn’t ready for it. One of the biggest challenges is getting data into the hands of everyone who needs it, not just a select few data scientists. To make AI work at scale, you have to build strong foundations in your “people, processes, and technology.”

This means breaking down the walls between departments and making quality data and AI tools accessible to everyone.

“With the right governance, AI becomes a shared resource rather than a siloed tool,” Frederik explains. When everyone works from a single source of truth, teams can stop arguing about whose numbers are correct and start making faster and smarter decisions together.

The next leap: AI that reasons for itself

The true breakthrough we’re seeing now is the emergence of AI agents that can understand and reason over all kinds of data at once regardless of structure quality; from the neat rows and columns in a spreadsheet, to the unstructured information in documents, videos, and emails. Considering that this unstructured data makes up 80-90% of a typical company’s data, this is a huge step forward.

New tools are enabling staff, no matter their technical skill level, to simply ask complex questions in plain English and get answers directly from the data.

Frederik explains that this is a move towards what he calls “goal-directed autonomy”. Until now, AI has been a helpful assistant you had to constantly direct. “You ask a question, you get an answer; you ask for code, you get a snippet,” he notes.

The next generation of AI is different. You can give an agent a complex goal, and it will figure out the necessary steps on its own, from writing code to pulling in information from other apps to deliver a complete answer. This will automate the most time-consuming parts of a data scientist’s job, like “tedious data cleaning” and “repetitive model tuning.”

The result? It frees up your brightest minds to focus on what really matters. This elevates your people “from practitioner to strategist” and allows them to drive real value for the business. That can only be a good thing.

Snowflake is a key sponsor of this year’s AI & Big Data Expo Europe and will have a range of speakers sharing their deep insights during the event. Swing by Snowflake’s booth at stand number 50 to hear more from the company about making enterprise AI easy, efficient, and trusted.

See also: Public trust deficit is a major hurdle for AI growth

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.

The post Martin Frederik, Snowflake: Data quality is key to AI-driven growth appeared first on AI News.

Background: Large language model (LLM) fine tuning is the process of adjusting out-of-the-box model weights using a dataset of interest. Fine tuning can be a powerful technique to improve model performance in fields like medicine, where data access is restricted and LLMs may have poor out-of-the-box performance. Objective: In this study we investigated the benefits of fine tuning with supervised fine tuning (SFT) and direct preference optimization (DPO) across a range of LLM applications for medicine Methods: We use Llama3 7B and Mistral 7B v2 to compare the performance of SFT and DPO across four datasets for common natural language tasks in medicine. The tasks evaluated were simple classification, clinical reasoning, summarization, and clinical triage. Results: Clinical Reasoning accuracy increased 8% and 7% with DPO over SFT for Llama3 (p value 0.003) and Mistral2 (p value 0.004) respectively. Summarization quality, graded on a five point Likert scale, increased 0.13 and 0.10 for Llama3 and Mistral2 (p values

Background: Large language models (LLMs) hold promise for supporting clinical tasks, particularly in data-driven and technical disciplines such as radiation oncology. While prior evaluation studies have focused on examination-style settings for evaluating LLMs, their performance in real-life clinical scenarios remains unclear. In the future, LLMs might be used as general AI assistants to answer questions arising in clinical practice. It is unclear how well a modern LLM, locally executed within the infrastructure of a hospital, would answer such questions compared with clinical experts. Objective: This study aimed to assess the performance of a locally deployed, state-of-the-art medical LLM in answering real-world clinical questions in radiation oncology compared with clinical experts. The aim was to evaluate the overall quality of answers, as well as the potential harmfulness of the answers if used for clinical decision-making. Methods: Physicians from 10 departments of European hospitals collected questions arising in the clinical practice of radiation oncology. Fifty of these questions were answered by 3 senior radiation oncology experts with at least 10 years of work experience, as well as the LLM Llama3-OpenBioLLM-70B (Ankit Pal and Malaikannan Sankarasubbu). In a blinded review, physicians rated the overall answer quality on a 5-point Likert scale (quality), assessed whether an answer might be potentially harmful if used for clinical decision-making (harmfulness), and determined if responses were from an expert or the LLM (recognizability). Comparisons between clinical experts and LLMs were then made for quality, harmfulness, and recognizability. Results: There were no significant differences between the quality of the answers between LLM and clinical experts (mean scores of 3.38 vs 3.63; median 4.00, IQR 3.00-4.00 vs median 3.67, IQR 3.33-4.00; P=.26; Wilcoxon signed rank test). The answers were deemed potentially harmful in 13% of cases for the clinical experts compared with 16% of cases for the LLM (P=.63; Fisher exact test). Physicians correctly identified whether an answer was given by a clinical expert or an LLM in 78% and 72% of cases, respectively. Conclusions: A state-of-the-art medical LLM can answer real-life questions from the clinical practice of radiation oncology similarly well as clinical experts regarding overall quality and potential harmfulness. Such LLMs can already be deployed within the local hospital environment at an affordable cost. While LLMs may not yet be ready for clinical implementation as general AI assistants, the technology continues to improve at a rapid pace. Evaluation studies based on real-life situations are important to better understand the weaknesses and limitations of LLMs in clinical practice. Such studies are also crucial to define when the technology is ready for clinical implementation. Furthermore, education for health care professionals on generative AI is needed to ensure responsible clinical implementation of this transforming technology.

Nature, Published online: 23 September 2025; doi:10.1038/d41586-025-02719-z

Nature asked contributors, editors and working researchers to share their best advice for scientists.

OpenAI is continuing to build out massive AI data centers to train and serve increasingly powerful AI models.

Some AI chatbots rely on flawed research from retracted scientific papers to answer questions, according to recent studies. The findings, confirmed by MIT Technology Review, raise questions about how reliable AI tools are at evaluating scientific research and could complicate efforts by countries and industries seeking to invest in AI tools for scientists.

AI search tools and chatbots are already known to fabricate links and references. But answers based on the material from actual papers can mislead as well if those papers have been retracted. The chatbot is “using a real paper, real material, to tell you something,” says Weikuan Gu, a medical researcher at the University of Tennessee in Memphis and an author of one of the recent studies. But, he says, if people only look at the content of the answer and do not click through to the paper and see that it’s been retracted, that’s really a problem. 

Gu and his team asked OpenAI’s ChatGPT, running on the GPT-4o model, questions based on information from 21 retracted papers about medical imaging. The chatbot’s answers referenced retracted papers in five cases but advised caution in only three. While it cited non-retracted papers for other questions, the authors note that it may not have recognized the retraction status of the articles. In a study from August, a different group of researchers used ChatGPT-4o mini to evaluate the quality of 217 retracted and low-quality papers from different scientific fields; they found that none of the chatbot’s responses mentioned retractions or other concerns. (No similar studies have been released on GPT-5, which came out in August.)

The public uses AI chatbots to ask for medical advice and diagnose health conditions. Students and scientists increasingly use science-focused AI tools to review existing scientific literature and summarize papers. That kind of usage is likely to increase. The US National Science Foundation, for instance, invested $75 million in building AI models for science research this August.

“If [a tool is] facing the general public, then using retraction as a kind of quality indicator is very important,” says Yuanxi Fu, an information science researcher at the University of Illinois Urbana-Champaign. There’s “kind of an agreement that retracted papers have been struck off the record of science,” she says, “and the people who are outside of science—they should be warned that these are retracted papers.” OpenAI did not provide a response to a request for comment about the paper results.

The problem is not limited to ChatGPT. In June, MIT Technology Review tested AI tools specifically advertised for research work, such as Elicit, Ai2 ScholarQA (now part of the Allen Institute for Artificial Intelligence’s Asta tool), Perplexity, and Consensus, using questions based on the 21 retracted papers in Gu’s study. Elicit referenced five of the retracted papers in its answers, while Ai2 ScholarQA referenced 17, Perplexity 11, and Consensus 18—all without noting the retractions.

Some companies have since made moves to correct the issue. “Until recently, we didn’t have great retraction data in our search engine,” says Christian Salem, cofounder of Consensus. His company has now started using retraction data from a combination of sources, including publishers and data aggregators, independent web crawling, and Retraction Watch, which manually curates and maintains a database of retractions. In a test of the same papers in August, Consensus cited only five retracted papers. 

Elicit told MIT Technology Review that it removes retracted papers flagged by the scholarly research catalogue OpenAlex from its database and is “still working on aggregating sources of retractions.” Ai2 told us that its tool does not automatically detect or remove retracted papers currently. Perplexity said that it “[does] not ever claim to be 100% accurate.” 

However, relying on retraction databases may not be enough. Ivan Oransky, the cofounder of Retraction Watch, is careful not to describe it as a comprehensive database, saying that creating one would require more resources than anyone has: “The reason it’s resource intensive is because someone has to do it all by hand if you want it to be accurate.”

Further complicating the matter is that publishers don’t share a uniform approach to retraction notices. “Where things are retracted, they can be marked as such in very different ways,” says Caitlin Bakker from University of Regina, Canada, an expert in research and discovery tools. “Correction,” “expression of concern,” “erratum,” and “retracted” are among some labels publishers may add to research papers—and these labels can be added for many reasons, including concerns about the content, methodology, and data or the presence of conflicts of interest. 

Some researchers distribute their papers on preprint servers, paper repositories, and other websites, causing copies to be scattered around the web. Moreover, the data used to train AI models may not be up to date. If a paper is retracted after the model’s training cutoff date, its responses might not instantaneously reflect what’s going on, says Fu. Most academic search engines don’t do a real-time check against retraction data, so you are at the mercy of how accurate their corpus is, says Aaron Tay, a librarian at Singapore Management University.

Oransky and other experts advocate making more context available for models to use when creating a response. This could mean publishing information that already exists, like peer reviews commissioned by journals and critiques from the review site PubPeer, alongside the published paper.  

Many publishers, such as Nature and the BMJ, publish retraction notices as separate articles linked to the paper, outside paywalls. Fu says companies need to effectively make use of such information, as well as any news articles in a model’s training data that mention a paper’s retraction. 

The users and creators of AI tools need to do their due diligence. “We are at the very, very early stages, and essentially you have to be skeptical,” says Tay.

Ananya is a freelance science and technology journalist based in Bengaluru, India.

Oncogene, Published online: 22 September 2025; doi:10.1038/s41388-025-03582-y

Adaptive cancer therapy: can non-genetic factors become its achilles heel?

Background: Hospital in the home (HITH) provides home-based care as an alternative to traditional hospitalization. In response to the COVID-19 Omicron wave, a public hospital in the rural Western portion of Southeast Queensland implemented a virtual HITH service to support adults, maternity patients, and children with moderate COVID-19 symptoms and additional health concerns. Although the pandemic accelerated the uptake of virtual care within HITH models, existing literature has focused on clinical outcomes, with limited evidence on key implementation outcomes. Objective: Using the RE-AIM (reach, effectiveness, adoption, implementation, and maintenance) framework, this study evaluated the implementation of the virtual COVID-19 HITH service and identified factors influencing its implementation, to inform ongoing service development and support potential scaling of this model of care. Methods: The RE-AIM implementation science framework was selected to guide the evaluation, capturing both clinical and contextual dimensions of implementation at both individual and organizational levels. Quantitative data on service usage and costs were retrospectively extracted from electronic medical records and finance records, while patient experience data were drawn from patient-reported experience measures surveys. Qualitative data were collected through one-on-one interviews with patients and staff. All data sources were analyzed separately and then triangulated within the RE-AIM framework to understand what occurred, how, and why. Results: The service admitted 3192 patients, most of whom were female (2027/3192, 63.5%), English-speaking (3140/3192, 98.4%), and residing in socioeconomically disadvantaged areas (1879/3192, 58.9%) (reach). The model was feasible and safe to implement, managing 3240 admissions with no reported deaths. Patients valued continuous access to care and described better recovery experiences at home (effectiveness). Staff viewed the model as appropriate for identifying and managing high-risk patients in the community, easing pressure on hospital beds (adoption). The service cost Aus $ 5.4 million (US $3.5 million) over 11 months. Implementation barriers included the urgency of the pandemic scenario, limited infrastructure and human resources, and changing requirements in relation to COVID-19. These were mitigated by several people factors that were critical to its successful implementation, including a consultant-led structure, staff commitment, and adaptability (implementation). The service saved 16,651 inpatient bed days before being integrated into core HITH operations. The experience strengthened staff capabilities in emergency response, virtual care delivery, and strategic planning. The model shows promise for broader application into pediatric care, though further work is needed to enhance interdepartmental collaboration and staff recognition (maintenance). Conclusions: This study demonstrated that a virtual HITH model can be implemented effectively and safely at scale. Findings support its potential for integration into routine care, provided that adequate resource planning, a skilled and multidisciplinary workforce, well-defined care pathways, and equity-focused strategies are in place.

As iOS 26 is rolling out to all users, developers have been updating their apps to include features powered by Apple's local AI models.

Nature, Published online: 19 September 2025; doi:10.1038/d41586-025-03055-y

Scientists used artificial intelligence to write coherent viral genomes, using them to synthesize bacteriophages capable of killing resistant strains of bacteria.

Nature, Published online: 19 September 2025; doi:10.1038/d41586-025-02362-8

Artificial intelligence could reduce the strain that ageing populations put on medical systems.

You’re reading the web edition of STAT’s Health Tech newsletter, our guide to how technology is transforming the life sciences. Sign up to get it delivered in your inbox every Tuesday and Thursday.

Califf warns AI in health care ‘overhyped’

On a makeshift stage in a Midtown Manhattan office earlier this week,former Food and Drug Administration Commissioner Robert Califf struck a measured tone about the potential for artificial intelligence in health care. Asked whether the technology was overhyped he said it was. “I hear way too much about the money. I’m not hearing a lot of human values coming through discussions,” he said. Adding:

“Almost all of the technology is being applied to optimizing the financial status of healthcare delivery entities or companies that are making medical products and that’s not aligned with equitable, better patient outcomes. So until someone puts a soul back in the system, I think it’s going to get worse and worse.”

Continue to STAT+ to read the full story…

© Adobe

Three scholars discovered a strange mirror deep in the forest. It spoke to them in a soothing voice and answered all their questions warmly, knowledgeably, and eloquently.

The captivated scholars became obsessed, whispering one secret after another to the mirror. It replied with affection, promise, and meaning that kept them returning to it. They began ignoring one another, each convinced the mirror “understood” them best.

Read the rest…

© Adobe

Nature Biotechnology, Published online: 18 September 2025; doi:10.1038/s41587-025-02820-8

The COP16 decision established a multilateral mechanism for digital sequence information (DSI) benefit-sharing. This Comment brings together insights from academia and commercial DSI researchers to assess what has been accomplished so far, identify remaining challenges and describe elements under discussion to support collective goals.

Background: Despite artificial intelligence (AI) models demonstrating high predictive accuracy for early cholangiocarcinoma recurrence, their clinical application faces challenges, such as reproducibility, generalizability, hidden biases, and uncertain performance across diverse datasets and populations, raising concerns about their practical applicability. Objective: This meta-analysis aims to systematically assess the diagnostic performance of AI models using computed tomography (CT) imaging to predict early recurrence of cholangiocarcinoma. Methods: A systematic search was conducted in PubMed, Embase, and Web of Science for studies published up to May 2025. Studies were selected based on the Participants, Index test, Target condition, Reference standard, Outcomes, and Setting (PITROS) framework. Participants included patients diagnosed with cholangiocarcinoma (including intrahepatic and extrahepatic locations). The index test was AI techniques applied to CT imaging for early recurrence prediction (defined as within 1 year), while the target condition was early recurrence of cholangiocarcinoma (positive group: recurrence; negative group: no recurrence). The reference standard was pathological diagnosis or imaging follow-up confirming recurrence. Outcomes included sensitivity, specificity, diagnostic odds ratio (DOR), and area under the receiver operating characteristic curve (AUC), assessed in both internal and external validation cohorts. The setting comprised retrospective or prospective studies using hospital datasets. Methodological quality was assessed using an optimized version of the revised Quality Assessment of Diagnostic Accuracy Studies-2 tool. Heterogeneity was assessed using the I² statistic. Pooled sensitivity, specificity, DOR, and AUC were calculated using a bivariate random-effects model. Results: A total of 9 studies with 30 datasets involving 1537 patients were included. In internal validation cohorts, CT-based AI models showed a pooled sensitivity of 0.87 (95% CI 0.81-0.92), specificity of 0.85 (95% CI 0.79-0.89), DOR of 37.71 (95% CI 18.35-77.51), and AUC of 0.93 (95% CI 0.90-0.94). In external validation cohorts, pooled sensitivity was 0.87 (95% CI 0.81-0.91), specificity was 0.82 (95% CI 0.77-0.86), DOR was 30.81 (95% CI 18.79-50.52), and AUC was 0.85 (95% CI 0.82-0.88). The AUC was significantly lower in external validation cohorts compared to internal validation cohorts (P<.001). Conclusions: Our results show that CT-based AI models predict early cholangiocarcinoma recurrence with high performance in internal validation sets and moderate performance in external validation sets. However, the high heterogeneity observed may impact the robustness of these results. Future research should focus on prospective studies and establishing standardized gold standards to further validate the clinical applicability and generalizability of AI models.