In 2026 The Lancet published the first population-scale audit of reference integrity across the biomedical literature. Here is what it found, what it means for medical affairs, and the standard to hold every AI vendor to, including us.
The Lancet audit covered 2.5 million open-access biomedical papers and 97.1 million verified references. The rate was low and flat through 2023, then rose more than twelvefold in three years. The authors place the turn in mid-2024, about when papers drafted with large language models would have started to appear in print.
Figures from Topaz M, Roguin N, Gupta P, Zhang Z, Peltonen L-M. Fabricated citations: an audit across 2·5 million biomedical papers. The Lancet 2026;407:1779–80, 9 May 2026. doi:10.1016/S0140-6736(26)00603-3. The ~4 and 51.3 per-10,000 figures, the one-in-N rates and the mid-2024 placing are as stated in the text. The authors note that increased paper-mill activity and changes in journal indexing may also have contributed to the rise; the timing annotations follow their text.
These references are dangerous because they are invisible: correctly formatted, attributed to real researchers, plausibly dated. Review does not catch them. Of 2,810 affected papers, 98.4% had received no publisher action at the time of the audit. The authors state their limits: the design estimates precision, not recall, and the 2026 data cover seven weeks. We hold ourselves to the same standard below.
The rate was flat for years. It turned upward after large language models became broadly available. A fabricated reference is a citation that is plausibly formatted, attributed to real researchers, passes review, but does not exist.
We believe AI risk is reduced when the tool is built for the domain. Evidence inside it; claims checkable against a source. A general-purpose AI model has neither.
Medicines are not approved because someone feels they will help. They are approved on comparators, endpoints, time periods and method. Software for medical affairs deserves the same proof. When any vendor, including us, offers an AI tool for work that touches the evidence behind a medicine, ask for evidence of capability on four axes.
How much of the therapeutic narrative gets surfaced. The value of a document is the substantive claims it puts on the page.
References trace to the original trial data, not to reviews restating findings at a distance.
The narrative draws on pivotal trials, real-world data, meta-analyses, guidelines and health economics, not one of them.
Every claim traces to a specific source, and every source can be verified against the claim it supports.
We designed and ran these studies ourselves. Whole systems, the whole task, against a general-purpose AI model run the same way. Scored blind where a reviewer could tell the systems apart. Misses stated. Results are summarized on the product pages; speak to us for the details.
Posteros against the best of three general-purpose AI models, claim by claim, with the per-drug spread and the limits stated.
No figure yet for the screen itself. Instead, every decision is on the record: each paper placed with its reason, six questions with denominators, four self-audit views, and your overrule on any of them.
A simulated panel's votes against the recorded outcome, beside a general-purpose AI model run the same way, misses counted.
Searching the literature, screening and classifying sources by evidence class, extracting the data points in each publication, proposing grades. No team can staff this work; it is what we built the AI to do. Once the search is done, the AI works only from the evidence it retrieved. It has no access to the open internet.
So is the review. Our claims study went through independent medical review and came back with six precision corrections, kept as errata. The AI’s work is treated as a claim to be checked, never as an answer to be trusted.
Statement to evidence to publication, link by link. Challenge any statement and we recheck it, then send you the full trace. Your writers start from a sourced draft and hold the pen.
Our tests against general-purpose AI: how they were run, and what they missed.