TL;DR
Courts around the world have now recorded more than two thousand cases of lawyers, and sometimes judges, filing citations that do not exist. The answer has been disclosure forms, certifications and fines, all of them aimed at human diligence. A paper published in April 2026 measured a different answer: a machine that refuses to pass a citation which does not resolve, with failure rates falling under one percent. The profession has the design and has not made it mandatory. Aviation faced the same choice in 1935 and chose the mechanism.
The official version is tidy. A lawyer got lazy, a chatbot invented a precedent, a judge noticed, somebody paid.
On 5 October 2026 the public database of AI hallucination cases stood at 2,149 decisions across more than sixty jurisdictions, with 1,473 in the United States, 229 in Canada and 112 in Australia. The counter moves most days. By the time you read this it will be higher.
The profession answered with paperwork
The response has been remarkably consistent, and remarkably one-directional. By May 2026 individual federal judges in the United States had issued 143 standing orders on AI use, with nine instruments at circuit level and formal guidance in 34 states. In Italy the Court of Cassation ruled on 11 June 2026 that filing unverified AI citations is qualified negligence, not an ordinary procedural slip, and raised the fine accordingly.
Every one of those instruments points at a conscience. Disclose that you used the tool. Certify that you checked. Accept the sanction when you did not.
Conscience is not a control. It is a hope with a signature block.
We have numbers on that hope. A study published in February 2026 looked at citation behaviour across 375,440 generated references and a survey of working researchers. 41.5 percent paste a citation without opening it. Among reviewers, 76.7 percent do not check references properly and 80 percent never suspect a fabricated one. These are trained professionals in a field whose entire currency is the citation.
The control exists and it has been measured
In April 2026 a team at Penn published the other half of the story. They tested commercial models and deep research agents for citation URLs that do not resolve, then put a mechanical resolution check inside the generation loop. Non-resolving citations fell from 16.0 percent to 0.6 percent on one model and from 6.1 percent to 0.1 percent on another. Under one percent across three.
Not a policy. A gate.
The legal profession already has the design, published in December 2025 as a mandatory hyperlink rule: every citation in an electronic filing must carry a functional link to an authoritative source. The logic is unanswerable. A hallucinated case cannot be hyperlinked, because it is not in the database. New York's Commercial Division has required hyperlinked citations since 2020.
So the score reads: one hundred and forty-three orders demanding declarations of good faith, and almost nobody requiring the one check a machine performs in milliseconds.
Wright Field, 30 October 1935
Boeing's Model 299 took off, stalled at three hundred feet and burned. Major Ployer Hill, flying it, had not released a new lock on the elevator and rudder controls. The investigation found pilot error, which was true and useless.
The Army did not ask for better pilots. Test pilots wrote a checklist. That aeroplane went on to fly eighteen million miles without an accident, and the Army bought nearly thirteen thousand of them under the name B-17.
Why this matters for your business
The courts are just a legible version of the choice every company makes in month three of an agent rollout. Policy and training, or a gate in the pipeline. Policy is cheap to write, satisfying to circulate and impossible to measure. A gate is boring and produces a number.
The serious work in this field is building the gate. Everything else is a memo.
So stop asking your vendor who reviews the output. Ask what fraction of the claims the system emits cannot leave the building unless they resolve to a retrievable source. If the answer arrives as a PDF of guidelines, you are flying a Model 299 with no checklist, and the investigation will find human error. It will be true, and useless.