Meta’s AI helped write 10 health papers in 4 weeks
Reviewers flag limits in the patient data; the findings offer research leads rather than reasons to change care.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:004:55
Read transcript
Meta’s AI-assisted health-research experiment produced ten manuscripts in four weeks. But the more important result may be the boundary its reviewers drew: these papers offer leads to investigate, not evidence on their own that patient care should change. Medical researchers used Meta’s Muse Spark 1.3 and Muse Wonder models to generate hypotheses, write and run code, and help draft manuscripts. They worked inside an internal software framework, with access to public information and a licensed database of five million de-identified electronic health records. Those records were collected retrospectively, not in a new clinical trial. Researchers remained responsible for the consequential scientific choices: which patients to study, which variables and methods to use, and how to interpret the results. The manuscripts received feedback from peers in the research group, analysis review by a biostatistician, and summary reviews from independent experts. That scrutiny surfaced limits the workflow and dataset could not resolve. One study reported a one-year heart-rhythm diagnosis risk of seven percent after infective endocarditis, compared with four percent after bloodstream infections without endocarditis. Penn State cardiologist Michael Farbaniec praised the methods, but cautioned that more intensive heart monitoring could explain the difference by detecting more cases. He called the result hypothesis-generating, not practice-changing. Other examples make the same point. A behavioral-health emergency study found worse outcomes among patients who left without being seen, but lacked details about illness severity, diagnoses and why they left. A diabetes study found documented remission more often ended when patients restarted treatment than when blood sugar rose. It could not establish why treatment resumed; its narrower recommendation was to report those outcomes separately. The practical takeaway is that AI can help accelerate analysis and drafting, while the quality of the underlying records and the researchers’ judgment still determine what a result means. On the software side, ArtCraft is testing how much of that same acceleration can be turned into a free alternative to established tools. Engineer Brandon Thomas has released 13 open-source apps for creative work, documents, presentations and spreadsheets. PhotoCraft targets Photoshop; WordCraft, DeckCraft and GridCraft aim at Word, PowerPoint and Excel. Thomas says he used Anthropic’s Claude Opus 5.5 and Rust, building from public file-format specifications rather than Adobe’s proprietary source code. That is his clean-room account, not a verified legal finding. Intellectual-property scholar Andres Guadamuz has questioned whether the process can be independently established. The apps are free, but feature parity, security, privacy and long-term maintenance remain unproven. They may suit technically comfortable users willing to download and configure open-source software. For professional or business use, the trade is not simply subscription cost versus zero cost; it also includes setup work and responsibility for evaluating the software. The cost question looks different inside the AI industry itself. OpenAI UK’s annual accounts show that 110 employees shared £255 million in compensation in 2025, an average of about £2.318 million per person. That is not an average cash salary: £168 million was in company shares, compared with £87 million in salaries. The Guardian reported that compensation varied substantially, with the largest packages going to AI researchers who help build OpenAI’s models. The UK subsidiary reported £293 million in turnover, more than triple the prior year, and £20.3 million in profit. These are figures for OpenAI UK Ltd, not worldwide totals. Its London operation includes research as well as government affairs, sales and marketing, so the accounts capture more than a research lab. And as AI systems take on more work, another practical question is what information they carry forward. Sakana AI’s My Fish feature lets Sakana Chat users inspect saved memories on web and mobile, trace facts to the conversations they came from, and see when memories were read or written. A connections map shows which memories were used together in a reply; its lines represent shared use in an answer, not simply related topics. A timeline also records memory activity, and a chip in the chat can link users to a memory being read or saved. Memory remains on by default, though users can turn it off in settings. Sakana describes more accurate memory selection and updating as future goals; this release makes existing behavior visible. Across these stories, the useful thing to watch is not just what AI can produce or remember, but what people can inspect and verify: the evidence behind a health claim, the provenance and safety of software, the composition of compensation, and the source of a chatbot’s recall.
