Our own content-fabrication register just crossed a line. Here is why we are not calling it a trend.

There was a time when an invented legal case, a fabricated academic reference or a quotation nobody ever said could be written off as an isolated mistake. That is getting harder to defend.

We are not running the crossover as a headline. This folio is about why, because the reason turns out to be more useful than the number.

What the register holds #

The register covers one specific failure, wherever it happens: a document goes out carrying a citation, a source or a quotation that does not exist. A court filing. A consulting report. A research paper. A police intelligence assessment. A government decision letter. Every entry is confirmed against a reputable source.11. It also excludes two things people expect to find here: plagiarism, and fabricated images or bylines. Both are real problems. Neither is a fabricated citation. If it cannot be verified, it is not listed.

As at this snapshot it holds 73 verified incidents — the earliest from April 2023 — across five sectors and 22 countries:

Read straight off that second table, 2026 has already passed 2025 with nearly five months to run. It is the line a register exists to produce. It would make a good headline, and it would be the wrong one.

Why we are not calling it a trend #

When Integrity Watch went public in June we seeded it with a one-off sweep of everything we could verify back to 2023. That sweep set 2025 at 30. Since then the register has been kept up by a weekly scan looking for what is happening now.

In the six weeks since, 2026 has gained 17 verified incidents. 2025 has gained one.

A count of what you found is not a count of what there is. Nothing goes into this register unless it can be confirmed against a reputable source — and we have been looking much harder at 2026 than at 2025.

So the register is a floor, not a census, and we think a low one. Three things have to happen before an incident reaches it. Someone has to notice the fabrication. Something has to follow — a correction, a withdrawal, a refund, a sanction. And that outcome has to become public and checkable against a reputable source. A fabricated citation nobody checks leaves no trace at all. We are counting consequences, not fabrications, and we have no way to measure the gap between the two — only reasons to believe it is wide.

The same caution applies to the sector table. Courts publish judgments; consultancies do not publish the reports a client quietly sends back. The 42 legal incidents show where fabrication gets written down, not where it happens most.

And the margin is one case.

The crossover rests entirely on it — an entry days old that our own register flags as single-sourced. A milestone you can undo by rereading your own footnotes is not a milestone.

What can be said #

A trend claim needs a denominator — a fixed population you can check repeatedly. The register has no denominator. One study does.

In May 2026 a Columbia University team led by Maxim Topaz published an audit in The Lancet covering roughly 2.5 million PubMed-indexed papers and 97.1 million references, checked against PubMed, Crossref, OpenAlex and Google Scholar. They found 4,046 fabricated references across 2,810 papers, and they measured the rate over time.

Roughly a twelve-fold rise in two years, against a stable denominator, in a single well-indexed field. That is the strongest available evidence that the underlying rate is climbing — and it is not ours.

One figure from that audit deserves more attention than it has had: 98.4% of the flagged papers are still in the literature, uncorrected. The problem is not only that fabricated references get published. It is that once published, they stay.

It also gives the gap a size. That audit found 2,810 papers carrying fabricated references, in one field. Our register holds 73 documents, across every sector and 22 countries. Those are the same unit — documents — and the distance between the two numbers is the distance between what gets written and what gets caught.

What the register is good for #

Not the rate. The texture — who, in what document, and what it cost.

Of the 73 verified incidents, 40 ended in a sanction, 20 in a withdrawal, 12 in a correction and one in a refund. The largest single category of fabrication is legal authorities, at 40 cases.

These are not typos. In a Spanish appeal in July, 24 of 32 cited authorities — 75% — were fabricated or wrong, including non-existent Supreme Court, Constitutional Court and Court of Justice of the EU rulings, several carrying invented verbatim quotations. In Nebraska, a court listed 20 examples of fictional quotes, wrong case numbers or misrepresented holdings in a single brief; by opposing counsel's count, 57 of its 63 references were defective. The attorney denied using AI, then admitted it by affidavit, and a temporary suspension from practice took effect in April.

The money is real too. The largest court penalty in the register is roughly $110,000 in an Oregon federal matter — $80,000 in fees, about $15,000 in fines against one attorney and $14,000 against local counsel, with the case dismissed with prejudice. One attorney had attempted to remove the false material and refile.

The system is supposed to catch this #

A legal system has layers. A lawyer prepares a submission. An opponent scrutinises it. A judge weighs it. Authorities can be checked against databases and the court record.

The register shows fabricated material getting through those layers.

In Mississippi in June, both opposing sides filed hallucinated citations in the same matter — and it was the court, not either party, that found them. Four attorneys were removed. One out-of-state lawyer admitted she had already filed further pleadings before the same court containing undiscovered fictitious citations.

That is the deeper problem. Fabrication does not have to convince everyone. It only has to get past the next verification layer.

It is not only law #

KPMG appears in the register after 40 of 45 citations in a report were found to reference fabricated titles. Three of the Big Four are now on it.

A Danish newspaper's internal AI summarising tool, run without the source material, invented both quotations and the person who supposedly gave them — a technology professor who does not exist. Academic publishing has the same signature: one Springer Nature journal case recorded 12 of 14 references that could not be found to exist, flagged after a hospital librarian went looking for them.

The common thread is not the profession. It is the document.

The defect has moved to the document that decides #

This is the shift worth watching, and it is recent. The fabrications used to sit in what the parties filed. Over the last few weeks they have turned up in the document that resolves the matter.

In July, the UK Upper Tribunal set aside an asylum refusal built on a Morocco country policy note that the tribunal could not locate and the Home Office's own country policy team could not confirm existed. The judge held that relying on a document that did not exist was “analogous to relying on bogus evidence”.

We keep that distinction in the register and we keep it here. In several recent entries the non-existence of the cited material is established on the record, while the AI link remains a court's inference. We record what was established. A register that overstates its own entries has no business auditing anyone else's.

South Africa is not watching this from the outside #

Four of the register's entries are South African, three of them from 2026.

In April, the Department of Communications and Digital Technologies withdrew its Draft National AI Policy sixteen days after publishing it for comment, when at least six of its 67 reference-list entries turned out to reference journals and articles that had never been published. Two officials were placed on precautionary suspension. Folio 02 is what Ukweli found in that draft beyond those six citations, and Folio 03 is about the framework it was built on, which is still published.

In July, in Molawa and Others v Matjhabeng Local Municipality, the Labour Court in Durban stayed a municipal disciplinary hearing after the chairperson issued a ruling citing authorities that do not exist.

And on 31 July came the case described earlier in this folio: a full bench of the Johannesburg High Court setting out a table of citation discrepancies in an acting judge's own judgment, and referring him to the Legal Practice Council. It is the most recent South African entry in the register.

The fourth entry is older. In January 2025, in Mavundla v MEC: COGTA KZN, the KwaZulu-Natal High Court found that of nine cited cases only two existed, ordered costs against the firm, and referred the matter to the Legal Practice Council.

A policy, a municipal disciplinary ruling, a High Court judgment and a firm's heads of argument. That is not a niche professional embarrassment. It is the paperwork of the state.

The wrong question and the right one #

The instinctive response to all of this is to ask how we stop AI making things up.

That is probably the wrong question. These systems will keep generating plausible material that is not true. The more useful question is:

How do we stop an unverified machine-generated claim from becoming an authoritative document?

That moves the burden off the model and onto the systems built around it. Who checks the citation. Who verifies the quotation. Who confirms the authority exists. Who confirms the source says what the document claims it says.

And, when none of that happens — who is accountable.

What all 73 have in common #

They are not sophisticated frauds. Every one is a confident, plausible, correctly formatted citation that a competent professional read, believed and signed off — and that no reviewer caught until it was public. The failure is never in the writing. It is in the last read before the document leaves the building.

That single class of defect — fabricated or unverifiable citations, sources, quotes and claims — is what Ukweli is built to find before a document is published.