We ran it on our own patent application

Ten models. Ten citations. Eight of them weren't real.

Before filing, we asked ten leading AI models to find prior art against our own patent application — a real task, with real money and a real deadline riding on the answer. Between them they produced ten patent citations. We checked every one against the USPTO register.

Eight of the ten were fabricated or mis-attributed

Not one of these models was being careless. They were doing the task we gave them, on the kind of question people are already trusting them with.

One model's headline finding — the prior art it called the single greatest threat to our application — does not exist as described. Acting on it would have meant narrowing our claims against a patent that isn't there.

!

Seven were real patent numbers on the wrong invention

This is the part that matters. The numbers resolve. They're genuine, granted patents. Only the attribution is wrong.

So a “does this citation exist?” check passes nine of the ten. Existence is not the test. Attribution is — and it's the one almost nobody runs.

What one of them looked like
✕ What the model said

US 10,853,527 — IBM, “Verification of factual assertions”

A real patent number. A plausible title. Exactly the prior art you'd fear.

✓ What the USPTO register says

US 10,853,527“System for Loss Prevention and Recovery of Electronic Devices”, Top Dawg Media.

The patent is real. It's about stopping people stealing your phone.

And it caught us, too. Our own filed application states that a particular patent belongs to Microsoft. Defensible — Microsoft owns LinkedIn. But the assignee of record is Linkedin Corporation, and our own gate flagged it. A claim can be reasonable, widely held, and still not what the register says. That's the entire product, and it found us out.

Read this honestly: ten citations on one real task, checked against the USPTO register on 15 July 2026. It is a count, not a rate — we are not telling you that AI is wrong 80% of the time, and anyone who quotes it that way (including us) is over-reading it. What it does show is the failure mode: a real identifier attached to the wrong record, which every existence check in the world will wave straight through.

AttestAlly — because “the AI cited it” isn’t a defence.

Mata v. Avianca, S.D.N.Y., June 2023. When counsel asked the model whether the citations were real, it confirmed they were.
Read the CNBC report on Mata v. Avianca →

Every one of these was found by somebody else: a judge, opposing counsel, a researcher. None were caught in-house.

THE ATTESTALLY LEDGER S.D.N.Y. · 22 June 2023
COURT RECORD

Six cases.
None of them existed.

A $5,000 Rule 11 sanction after a brief cited six decisions that were never written. Asked to confirm the citations were real, the model said they were.

Read the original: CNBC, June 2023 →
THE ATTESTALLY LEDGER S.D.N.Y. Bankr. · 18 April 2026
THE FIRM DID NOT CATCH IT

Wall Street firm apologises to the court.

Sullivan & Cromwell filed wrong and fabricated citations in the Prince Group wind-down. Opposing counsel found them — not the firm’s own review.

Read the original: Bloomberg Law, April 2026 →
THE ATTESTALLY LEDGER Canberra · October 2025
NOT ONLY LAWYERS

A quote from a judgment that was never said.

A Deloitte report on a AUD 440,000 government contract carried a fabricated Federal Court quote and references to papers that do not exist. A university researcher found it.

Read the original: AP / Fast Company, Oct 2025 →

Every documented case we track →

Summaries by AttestAlly. Each clipping links to the original reporting.