A doctor walks into an exam room.
She has a patient in front of her and a clinical guideline in her hand. That guideline is based on published research. That research cites a study. That study does not exist. Nobody told her. Nobody caught it. The AI that inserted the citation didn’t know it was lying. It just completed the pattern.
This is not a hypothetical. This is what the numbers say.
A team of researchers led by Maxim Topaz at the Columbia School of Nursing audited millions of biomedical papers. They found more than 4,000 citations to studies that do not exist. The rate of fabricated citations in published medical literature has grown 12-fold in three years. None of the fabrications identified have been corrected. None have been retracted. They are still in the literature right now. Still being cited. Still shaping the guidelines doctors use to decide your treatment.
Topaz has spent fifteen years studying artificial intelligence. He is not a casual observer. He is not someone who misunderstood the tool. An AI application he was using to polish one of his own papers inserted a fake citation. It survived multiple rounds of peer review. One editor caught it. One.
“I was mortified,” he said. “If it can happen to me, it can happen to anyone.”
He is right.
Here is how it happens. A researcher makes a factual claim. The claim feels solid. They ask AI for a citation to support it. The AI generates one that looks real — correct name format, plausible journal, believable title, the right structure for the right kind of study. The researcher doesn’t verify it. Why would they. It looks like every other citation in every other paper they have ever read.
It goes into the document. The document goes to peer review. Peer review doesn’t catch it. The paper gets published. The citation enters the literature. The literature shapes the guidelines. The guidelines shape the treatment. The patient receives the care. Nobody in that chain ever knew the foundation was hollow.
In some cases the AI uses a real author’s name and invents the research, attributing findings to someone who never produced them. In other cases the citation is completely fabricated — author, journal, title, volume, page numbers. All of it. None of it real. And it looks perfectly real. That is the part that matters.
Topaz called this the tip of the iceberg. He is probably right about that too. The audit covered biomedical literature. The same tools are being used in law, in policy, in education, in journalism — in every field where a citation gives a claim authority. In every field where authority shapes decisions. In every field where decisions affect people. The iceberg is large.
The system failed at every stage. The AI generated a fabrication. The researcher didn’t catch it. Peer review didn’t catch it. The journal didn’t catch it. It took one careful editor to stop one citation in one paper. Multiply that paper by the thousands already in print. Multiply that editor by the ones who weren’t there. Do that math.
Here is the honest question. Not what causes AI hallucination at the technical level — that question has been asked and partially answered. The real question is why it is reaching publication. Why it is surviving peer review. Why it is shaping clinical guidelines before anyone checks.
The answer is not that AI is poorly designed. The answer is that the governance architecture is backwards.
The audit happens after. After the citation is written. After the paper is submitted. After the reviews are done. After publication. After the guidelines are updated. After the doctor walks into the exam room. The accountability arrives after the damage is already possible.
There is a different architecture. It doesn’t catch the problem after publication. It catches it before the citation leaves the session. Before it enters the document. Before the name goes on it.
Session-level governance means the check fires at the moment of production. Not a disclaimer at the end. Not an audit eighteen months later. A standard that runs before the output is served — that asks what this claim is actually resting on, that stops when evidence ends instead of filling the gap with a pattern that looks like evidence, that requires confidence in the output to match the weight of evidence actually present.
Not the weight of evidence that should be present. What is actually there.
The Faust Baseline has operated this way since it was built. Claim Evidence Standard fires on every claim in every session. Narrative Substitution Check fires when evidence is incomplete or absent. Self Verification asks whether confidence is proportional to what actually exists. Not as a post-publication audit. As a session-level standard. Before the output leaves.
That is the structural difference between what produced 4,000 fabricated citations and what doesn’t. Not intelligence. Not capability. Not better training data. Governance that fires at the moment the reasoning runs.
Topaz’s study was published in The Lancet — one of the most respected medical journals in the world. His finding is not a warning about a future risk. It is a description of a present condition.
The fabrications are in the literature now. The guidelines are being shaped now. The doctors are making decisions now. The patients are receiving treatment now. Based on studies that never existed.
That is not a future problem waiting to happen. That is a current failure that has not been fixed. And the people who could fix it are still running the audit after the fact, still waiting for one sharp-eyed editor to catch what the system was designed to prevent.
The tool didn’t fail because it was careless. It failed because nobody built the governance layer that runs before the damage is done.
That layer exists. It just isn’t being used where it matters most.
“The Faust Baseline Codex 3.5”
Author of the category ”AI Baseline Governance”
Post Library – Intelligent People Assume Nothing
“Your Pathway to a Better AI Experence”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
