For law firms
$145,000 in one quarter. The AI sanctions cases are not about AI.
At least $145,000 in sanctions for fabricated AI citations in the first quarter of 2026 alone (ComplexDiscovery, April 6, 2026). That is the aggregate that gets forwarded around a firm. The individual cases are more instructive, and they all fail in the same place.
In every one of these cases a document containing legal authority left the firm without a named person pulling the cited cases and reading them — and that, not the model, is the defect. A better tool would not have fixed it. The same failure produced bad citations in briefs written from bad memory long before any of this existed. Generative AI raised the volume and the plausibility. It did not invent the failure mode.
Which means the fix is available to you this month, costs nothing to write down, and is already the thing your carrier is asking about.
The cases, and the schedule Oregon quietly adopted
Whiting v. City of Athens, Sixth Circuit, No. 25-5425, decided March 13, 2026. Two attorneys sanctioned $15,000 each, plus appellee fees and double costs, over 24 or more fabricated or incorrect citations. Reported as the largest federal appellate AI-hallucination sanction on record.
Couvrette v. Wisnovsky, District of Oregon: $15,500 for 15 false citations and 8 fabricated quotes.
And the Oregon Court of Appeals has done something quietly significant. It has effectively adopted a schedule: $500 per false citation, $1,000 per fabricated quote. Applied in Watson in December 2025 for $2,000, and in Ghiorso in March 2026 for $10,000.
Read the last one again. A per-unit price. Not a discretionary sanction weighed against the totality of the circumstances — arithmetic. Courts price things they expect to see more of.
The obvious conclusion — AI is dangerous, be careful — is so incomplete it is useless. Firms acting on it either ban the tools, which converts documented use into undocumented use, or circulate a memo, which does less.
Two failures repeat, and neither one is technological
A verification step did not happen. The model produced plausible-looking citations because producing plausible-looking text is what it does. Nobody checked. A citation checker in the research platform the firm already pays for would have caught most of these in minutes. The failure was not capability. It was that the step was not in the process.
It was unclear who was supervising. The same ambiguity appears again and again: an associate or contract attorney drafted, a partner signed, and there was no moment at which a named person owned the verification. Everyone assumed someone else had done it. Rules 5.1 and 5.3 exist precisely because that ambiguity is predictable, and ABA Formal Opinion 512 of July 29, 2024 applied them to this directly — firm leadership must set written policy and supervise lawyers and staff in AI use.
That distinction decides what you do next. If the problem is AI, you buy or ban a tool. If the problem is a missing step and an unclear owner, you write down the step and name the owner.
The control, in eight lines
It applies to any document containing legal authority that will be filed, served or sent to a client, where any part of the drafting involved an AI tool. That trigger is deliberately broad, because the alternative is asking a drafter to self-report at 6pm.
- Every cited authority is pulled — opened in Westlaw or Lexis from the citation as written in the draft.
- Every citation is read for the proposition it is offered for. The common real-world failure is no longer the wholly invented case; it is the real case cited for something it does not say.
- Every quotation is matched to the source text, character for character. Oregon prices a fabricated quote at twice a false citation for a reason.
- Every pin cite is confirmed. Page numbers are exactly what a language model generates fluently and wrongly.
- Every authority is checked for subsequent history.
- The verifier is named on the record — not "verified," but verified by [name], [date].
- Where the drafter is not the supervising attorney, the verifier is someone other than the drafter.
- The supervising attorney signs having seen the completed verification record.
Eight lines. It fits on a card and it goes on the matter file.
Then the thing that is not on the card and matters as much: a near-miss log. When a verifier catches a bad citation before it goes out, record it. Not to discipline anyone — to know. A carrier now asks whether the firm has had any incident involving inaccurate AI-generated content, and "no" is only a real answer if you would have found out.
Your carrier has already connected these dots
Per a 2026 EPIC carrier survey reported in specialty insurance trade press, seven of thirteen major lawyers professional liability carriers reported AI-related claims increases, eight of thirteen reported overall claim-frequency increases — the first rise in five years — and more than 60% now ask AI questions at renewal. CNA has sent supplemental AI questionnaires since 2025 covering tools, governance policy, training, work-product verification and incident response. Trade-press sourced; ask your broker.
Note the fourth item. Work-product verification is a named section. The underwriter is asking you to describe the checklist above. "Attorneys are responsible for verifying their own work" is a true statement and a weak answer: it describes a duty, and the question asks about a process.
What to do Monday
- Write the eight lines above onto one page and date it.
- Decide the one case where the verifier may be the drafter, and write down why.
- Start the near-miss log today, even empty, with a date on it.
- Ask your research platform how its citation-checking tool is licensed to you. You are probably already paying for it.
- Pull your last professional liability application and read what the firm said about verification.
What we will not do
We will not automate the verification step and tell you it is handled. We could build something that pulls citations and flags the ones that do not resolve; plenty of vendors sell that well. What we will not do is let a tool stand as the answer to the question the carrier is asking, because the failure in every case above was that no person owned the check. The agent can do the pulling. A named person reads and signs. That order does not reverse in anything we build.
Signet, a division of Circle Square Consulting. AI that holds up. Radnor, Pennsylvania. The Renewal Dry Run is free and takes 30 minutes: /renewal-dry-run