What did the study actually find?

A published paper's odds of containing a fabricated citation are now roughly 1 in 277 — and that number has been rising fast, not sitting still.

The Lancet published the research on 7 May 2026: Columbia University researchers, led by Maxim Topaz, analysed more than 2 million papers and 97 million citations. In 2023, fabricated references showed up in 1 of every 2,828 papers. By 2025, that had jumped to 1 in 458 — six times worse. In the first seven weeks of 2026 alone, it hit 1 in 277. More than a third of the fabricated citations traced back to just two large open-access publishers, which the researchers declined to name. Their own description of what's accumulating in the published record: "slop."

Why does an academic-publishing story belong in a business newsletter?

Because the mechanism isn't academic. It's the model.

A citation, a client testimonial, a competitor statistic, a regulatory figure — these are all the same shape of claim to a language model: something it's being asked to produce that has the form of a fact. When the model doesn't actually know the answer, some produce something that looks like one anyway, complete with a name, a date, and a source that was never checked. Peer review is specifically designed to catch this, with editors and reviewers whose job is exactly that. And it's still getting worse, sixfold in two years.

If a formal review process with trained specialists is missing 1 in 277, the honest question is: what's the miss rate for a solo founder pasting an AI-drafted paragraph straight into a client email?

Doesn't AI usually get things right, though?

Usually, yes — which is exactly what makes this dangerous rather than obviously broken.

Stanford's AI Index (2026), using its AA-Omniscience benchmark, tested 26 leading models on hard factual questions specifically designed to have no clean answer — the point being to see whether a model would guess or admit it didn't know. The result: models guessed instead of abstaining 22% to 94% of the time, depending on the model. Even the best-performing model in that test was wrong roughly a quarter of the time it should have said "I don't know." A model that's right 90% of the time still hands you a confident, wrong answer on the tenth try — and it reads exactly as confident as the other nine.

What does this actually change about how I use AI tools?

Nothing about whether you use them. Something about which parts of the output you trust by default.

The distinction that matters: framing, opinion and structure from an AI tool are low-risk, because there's nothing to fabricate — they're judgement calls, and you're the editor either way. A specific number, name, date, price or cited source is a different category entirely, because it's a checkable claim being presented as already checked. This week's hub sets out the five-question pass for exactly that category, and Tuesday's post gives you the 30-second version to run right now, on the next thing an AI tool hands you with a source attached.

Is this rate going to keep climbing?

Nobody knows, and anyone who tells you with confidence is doing the exact thing this post is about.

What's measurable is the direction: three data points across three years, each one worse than the last, from a research team with no reason to inflate the number — if anything, academic publishers have every incentive to undercount their own error rate. Thursday's post goes deeper into why models behave this way rather than simply refusing to answer, because the "why" changes what a useful fix looks like. For now, the practical takeaway doesn't need the trend to continue: the rate is already high enough today to justify checking anything you're about to publish with someone else's name on it.

For the guide that finds every place AI already touches your business — free, 5-min read: The 10-Hour Audit.

Related Decodes

• The Citation That Doesn't Exist — this week's hub.
• Three Sizes of AI Just Launched — You Probably Need the Cheap One — an earlier News Decode on model capability claims worth the same scepticism.
• The 30-Second Check — tomorrow's post: the concrete habit.
• Your AI Would Rather Guess Than Say I Don't Know — Thursday's post: why models behave this way.

AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.

Not financial or legal advice. Figures describe academic-publishing research (The Lancet, May 2026) and a model benchmark (Stanford AI Index 2026); they illustrate a measured trend, not a probability for any specific AI output.

TL;DR: A Lancet study (7 May 2026, Columbia University, 2M+ papers and 97M citations analysed) found fabricated citations in peer-reviewed research rose from 1-in-2,828 papers in 2023 to 1-in-458 in 2025 to 1-in-277 in early 2026 — a sixfold increase in two years, despite formal peer review. Separately, Stanford's 2026 AI Index found leading models guess rather than admit uncertainty on hard factual questions 22-94% of the time. The lesson isn't to stop using AI tools; it's that framing and structure from AI are low-risk, while any specific number, name, date, price or source is a checkable claim that needs checking, because a model that sounds confident and a model that's correct are not reliably the same thing.