What sanctions are courts imposing for AI hallucinations, and what AI standards should I set for outside counsel?

What sanctions are courts actually imposing for AI hallucinations?

As of mid 2026, the range runs from a few thousand dollars to a case-ending order. Most rulings still fine the filing lawyer a modest amount. But the ceiling has moved: an Oregon magistrate imposed the largest AI-hallucination sanction in American courts and dismissed the underlying claim with prejudice. A running tracker counts well over a thousand filings worldwide with AI-fabricated authority.

The Oregon order is the one to know. The case was an elder-abuse claim with a seven-figure ask. Judge Clarke did not stop at the fine. He entered terminating sanctions and dismissed the claim with prejudice, meaning it can never be refiled. He fined the lawyers and killed the case, and he wrote that if there was ever an appropriate case for terminating sanctions over AI misuse, this was it.

Are courts now sanctioning the firm, not just the lawyer?

Yes, and that is the shift executives should watch. On July 1, 2026, Vice Chancellor Lori W. Will of the Delaware Court of Chancery ordered a firm, not just the partner who signed, to show cause and to file its written generative-AI policies onto a public docket. She invoked Rule 11's joint-responsibility clause. The firm's internal AI governance became discoverable by any client or adversary.

In that ruling, Leiske v. Kidd, the court found that counsel's corrected brief merely deleted the quotation marks while preserving the fabricated legal synthesis. The firm and the signing partner were given until July 15 to answer. The mechanism matters: once a firm has to describe its AI policies, training, and safeguards on the record, those policies stop being a document on a shelf and become something clients can demand to see.

The lesson for a general counsel or claims executive is direct. If a court can pull your panel firm's AI governance onto a public docket, you should be able to pull it into your outside-counsel guidelines first. What the bench is now willing to compel, you can require by contract before the sanction ever lands.

Where do courts stand on AI and privilege?

Split, and moving fast. On February 10, 2026, Judge Jed Rakoff in the Southern District of New York held that documents a defendant created with consumer Claude were not protected by attorney-client privilege or work product. The same day, a Michigan magistrate reached the opposite result on work product, holding that AI tools are tools, not persons. Two courts, two answers.

• Southern District of New York (Rakoff): consumer-tier AI chats are not privileged; no reasonable expectation of confidentiality when the vendor is not contractually bound.
• Eastern District of Michigan (Warner v. Gilbarco): sharing materials with an AI tool is not disclosure to an adversary, so work product protection survives.
• The practical takeaway: the tier and the contract decide the outcome. Consumer tools that train on inputs sit on the wrong side of both rulings.

Until this settles, treat the tier as the privilege question. A lawyer running client facts through a consumer chatbot may be generating a discovery target. The same work on a properly contracted enterprise tool, where the vendor is bound to confidentiality, sits on much safer ground.

What AI standards should I set for outside counsel?

Set them in writing, in your guidelines, before the next matter opens. The bar is not exotic: disclose AI use, verify every citation, keep client data out of consumer tools, and shift the cost of a sanction back to the firm that earns it. As of mid 2026 the case law is unsettled, so write standards that hold regardless of which way a given court comes out.

1. Disclosure of AI use: require every panel firm to tell you which drafting workflows use AI, which tool, and what verification process is in place.
2. Verification duty: make the filing attorney responsible for confirming every citation and quotation against the actual source. The Ninth Circuit has framed the violation as occurring at the point of signing and filing, not at research or drafting.
3. No client data in consumer tools: prohibit entering client or matter facts into consumer-tier AI that trains on inputs; require enterprise tools contractually bound to confidentiality.
4. Sanction indemnity: put language in your outside-counsel guidelines making the firm responsible for the cost and consequence of any AI-driven sanction on your matters.
5. Audit your own house: apply the same standard to any internal AI use, and require vendors to attest to security, zero data retention, and privilege protection in the master agreement.

This is also the standard CaseGlide holds itself to. Every fact its AI extracts stays traceable to the report it came from, and answers are drawn only from your own case data. That is the model to demand from any AI touching your litigation: grounded in a real source, and never a black box.