The Fake Source That Cites Itself Into Truth

2026-08-21

In 2023, roughly one in every 2,828 academic papers contained a fabricated citation — a reference to a study, author, or finding that did not exist. By 2025, that rate had climbed to one in 458. In the first seven weeks of 2026, researchers auditing major repositories put it at one in 277. The citations were not sloppy. They were correctly formatted, attributed to real researchers working in the right field, dated plausibly, and often published in journals that genuinely cover the topic. Reviewers waved them through because there was nothing about them that looked fake.

That is the uncomfortable part of this story. The problem was never that AI systems occasionally invent a source. Every organisation that has queried a chatbot for a quick answer already knows that happens. The problem is what happens next: a fabricated citation, once it appears in a published paper, becomes a citable fact in its own right. The next researcher — human or AI — who encounters it has no way to distinguish it from a real one. It gets cited again. Then it gets cited by whatever paper cites that paper. Each repetition adds a layer of apparent legitimacy that has nothing to do with whether the original claim was ever true.

Courts are running into the identical mechanic, at a faster pace. A public tracker maintained by a legal researcher now logs roughly 1,490 court decisions worldwide — more than 1,000 in the United States alone — where a party relied on AI-fabricated case law or filings and a judge had to respond. Penalties have moved from four-figure fines to a $15,000 sanction against a single attorney in a federal appeals case, and the first bar suspensions tied directly to AI-drafted filings. In one closely watched case filed this March, an insurer sued the maker of a chatbot directly, arguing the system had effectively practised law on a claimant's behalf — drafting 44 filings, one of them built on a case that never existed.

Fabricated case citations and fabricated academic citations are the same failure wearing different clothes. Something invents a reference. The reference gets embedded in a document that a human signs, files, or publishes. From that moment on, the fabrication has a paper trail that looks exactly like a real one, because paper trails do not carry a stamp saying who — or what — generated the first link in the chain.

This is where the citation problem stops being a curiosity for academics and lawyers and becomes a genuine reputational one. The same laundering mechanic that turns an invented case into "precedent" can turn an invented claim about a person or an organisation into apparent fact. A false statement does not need to be well-sourced to spread — it needs to be repeated in enough places that the next system summarising the topic treats the repetition itself as corroboration. A claim that appears on a low-authority forum, gets picked up by an AI-generated summary site, and then surfaces in a chatbot's answer has, by the third hop, acquired a kind of consensus it never earned. Nobody along that chain checked the first link. Most people encountering the third or fourth link have no way to.

None of this is really about AI systems getting facts wrong, though that is the part everyone notices first. It is about the disappearance of a stable record of who said what, sourced from where, and when. Academic publishing has peer review and retraction notices — imperfect tools, but tools with centuries of institutional weight behind them. Courts have sanctions and published opinions. Reputation has neither. When a person or an organisation is described falsely by an AI system, there is usually no register anywhere recording the exact wording, the date, the system that produced it, or the source it claimed to be drawing from — which means there is nothing to correct against and nothing to show a court, a regulator, or a counterparty later, once the claim has been repeated enough times to look settled.

The scientific and legal cases are a preview, not an anomaly. They show what happens once a system with no memory of its own reasoning is trusted to generate claims that get treated as durable. The fix in both domains has been the same: someone eventually built a record — a retraction database, a sanctions tracker — that could pin a fabrication to its origin before it hardened into accepted fact. Reputation is still waiting for its version of that record. Until it has one, the safest assumption for any company, law firm, or public official is that whatever an AI system says today is the version of events that gets treated as true tomorrow, by systems that were never in the room to check.