Answers · AI evidence
A Delaware court just sanctioned a lawyer over AI-made-up quotes in a brief but let his firm off because of its AI policy and training. What did the court actually require, does fixing the quotes afterwards help, and what records should a firm keep about how AI was used in a filing?
Last updated
TL;DR: In the Delaware Court of Chancery's letter opinion of November 13, 2026 (C.A. No. 2025-0426-CDW), it did protect the firm, but not the lawyer who signed. A brief filed in January 2026 contained fabricated quotations, inaccurate citations and hallucinated legal propositions introduced when the signing partner used a GenAI tool, still in a pilot and not approved for general firm use, to draft and condense arguments. The court released the firm without sanction because, before the brief was filed, it had banned unapproved GenAI tools (March 2023), adopted a written GenAI policy requiring an experienced lawyer to check AI output (July 2025) and trained over 200 staff. The signer was fined $2,500 under Court of Chancery Rule 11, payable personally. The court said the duty to verify is personal and can't be handed to a paralegal, a junior lawyer or an AI tool; that real cases and a broadly correct position are no excuse; and that removing quotation marks while keeping the made-up legal propositions doesn't fix the brief. The order to show cause asked for a timeline of who entered the prompts and how the output went into the brief, so keep that record as you work.
What should I check, step by step?
1. Have a written GenAI policy before anything goes wrong. The court relied on safeguards that were in place before the brief was filed: a ban on unapproved tools, a written policy requiring an experienced lawyer to review AI output, and training.
- Make the person who signs read every cited authority. The court said the Rule 11 inquiry and the duty of candor are personal and nondelegable, so a paralegal's or junior lawyer's cite-check doesn't discharge the signer's duty.
- Check what each case is cited for, not only that it exists. The court said GenAI tools also make up plausible quotations and stretch real decisions to propositions they never addressed.
- When an error is flagged, go back to the sources and rewrite every affected statement. Taking the quotation marks off and keeping the made-up legal synthesis was itself criticised.
- Don't push back on the side that spotted the errors. The court said opposing counsel was right to raise fabricated citations directly with the court, without a meet-and-confer first.
- Keep a dated record of each GenAI use in a filing: the tool and version, who entered the prompts, which passages the output went into, and who checked which citations against which database. That is what the court's order to show cause asked the signer to explain.
- Keep the firm's side too: the policy text in force at the time, how it was communicated to the people involved, and training completion records. The firm was asked for exactly these.
- Treat pilot tools like any other. The tool in this case was in an internal pilot and not approved for general use; the policy still applied to the lawyer who used it.
What did the Delaware Court of Chancery decide about a brief with AI-hallucinated citations in November 2026?
The case was an advancement dispute. On January 22, 2026 the plaintiffs filed an answering brief; five days later the defendants' reply pointed out fabricated quotations, inaccurate citations and hallucinated legal propositions. The next day the signing partner filed a letter saying a GenAI tool had been used to revise the brief and that the problems were missed in a paralegal's cite-check, along with a corrected brief.
- The corrected brief removed the quotation marks around the quotes that didn't exist but kept the inaccurate descriptions of the cases. When the defendants pointed that out, the plaintiffs' response insisted the authorities were real and supported the arguments, and objected that the defendants hadn't met and conferred first.
- On July 1, 2026 the court ordered the partner and the firm to show cause why sanctions shouldn't issue, and on November 13, 2026 it resolved the matter on their sworn affidavits without a hearing: no sanction for the firm, $2,500 for the partner under Rule 11, to be paid personally to the court within 14 days. No extra AI training was ordered because he had already completed several courses on GenAI ethics and led the firm's internal training. Because the court started the inquiry itself rather than on a motion, it said it couldn't award the other side attorneys' fees under Rule 11.
Why did a law firm's GenAI policy spare it from sanctions while the lawyer who signed was fined?
Court of Chancery Rule 11(c)(1) says that, absent exceptional circumstances, a law firm must be held jointly responsible for its lawyers' violations. Drawing on Delaware's professional conduct rule on law firm management and ABA Formal Opinion 512, the court found exceptional circumstances: the firm barred unapproved GenAI tools in March 2023, set up a working group in 2024, adopted a comprehensive GenAI policy in July 2025 that requires an experienced lawyer to review GenAI output and forbids using it in place of independent judgment, and built an accredited one-hour ethics course that more than 200 staff completed by January 2026. The court said the firm should not be sanctioned for an individual lawyer bypassing those controls.
- For the signer, the court stressed three points. The duty to make a reasonable inquiry and the duty of candor are personal and can't be delegated to a paralegal, a junior lawyer or an AI program. It is no excuse that the cited cases are real or that the overall position is arguably supported by general law. And once errors are found there is an affirmative duty to review the underlying authorities and correct every inaccurate quotation and characterization.
- The opinion states the general rule plainly: a fake quotation or a fabricated legal synthesis generated by an algorithm is not existing law, and presenting unverified GenAI output to a court violates Rule 11(b)(2). Using GenAI is neither forbidden nor inherently a problem; doing it without the required care is.
What records should a lawyer keep about how AI was used in a court filing?
The order to show cause asked the signer for a timeline of how and when the GenAI tool was used, including who entered the prompts and how the generated output was incorporated, plus a description of the cite-check: instructions to the paralegals, the databases used, whether lawyers verified the cited text, and whether the cite-check edits made it into the filed version. It asked the firm for its written GenAI policies in force at the time, how they were communicated to the people involved, and its safeguards and training.
- Those questions are much easier to answer from records made at the time than from memory months later. A practical set: for each AI-assisted draft, the tool and model version, a hash of the prompt and the output, the document and section it went into, and a cite-check log listing each authority, who checked it, against what source, and when. Keep the policy version and the training records alongside.
- Records don't replace reading the cases. They show what was done; the Delaware court made clear the signer still has to check every authority personally before filing.
How do I ask Rightshold about this (free, no key)?
Add the MCP server https://www.rightshold.com/mcp to Claude, Cursor, ChatGPT, Grok or VS Code, then ask: “Record a signed receipt for one GenAI drafting step in a brief: the model version, hashes of the prompt and the output, and in params the matter number, the section of the brief the output went into, the policy version in force and the id of the cite-check log entry, so we can later show what the tool produced and that it was checked.”
{
"tool": "record_ai_run",
"arguments": {
"model": "<model id and dated version>",
"provider": "<tool or model provider>",
"prompt_sha256": "<sha256 of the prompt as entered>",
"output_sha256": "<sha256 of the full output as returned>",
"params": {
"matter": "<internal matter number>",
"document": "answering-brief-draft-3",
"section": "Argument II.B",
"genai_policy": "genai-policy v2025-07",
"cite_check_log": "cc-2026-11-14-04"
},
"ran_at": "2026-11-14T20:30:00Z"
}
}- Needs an API key. Send hashes and the fields shown; keep the prompt, output and brief in your own document system, and keep client names and confidential facts out of params. params are stored as a hash, so keep the exact params object.
- Record each AI-assisted step as it happens, and keep the cite-check log (who checked each authority, against which database, when) in your own system under the id you put in params.
- verify_ai_run is free and needs no key: anyone with the receipt, its signature and the output text can confirm it matches what was recorded.
- A receipt shows what was recorded and when. It doesn't check citations, read cases or confirm that anything in a brief is accurate.
Rightshold isn't a law firm and isn't affiliated with any court, bar, firm or party in this case. A receipt is evidence of what was recorded, not proof that a filing was checked or complies with Rule 11. Details are from the Delaware Court of Chancery's letter opinion of November 13, 2026 and its order to show cause of July 1, 2026 in C.A. No. 2025-0426-CDW (LWW), read November 14, 2026. The decision applies Delaware's Court of Chancery Rule 11; other courts and rules may differ. Not legal advice.
What are the sources?
- Delaware Court of Chancery: letter opinion regarding Rule 11 sanction, C.A. No. 2025-0426-CDW (LWW), November 13, 2026
- Delaware Court of Chancery: letter order to show cause, C.A. No. 2025-0426-CDW (LWW), July 1, 2026
- Delaware Courts: Court of Chancery opinions list
Related
- How do I check AI-generated citations for hallucinations?
- How to prove an AI-generated report cited sources that actually said what it claims
- How to show a customer or regulator that a human reviewed an AI decision before it went out
- How to prove which model version, prompt and settings produced an AI output
- Are your ChatGPT or Claude chats privileged, or can they be used as evidence in a lawsuit?
Not legal advice. JSON · Markdown · All answers
Get a free API key in one call
No card and no email needed. 100 free credits; the license and policy checks stay free without a key.
curl -s -X POST https://www.rightshold.com/v1/accounts/signup -H "Content-Type: application/json" -d '{"accept_terms":true}'
export RIGHTSHOLD_API_KEY=rs_... # the api_key from the reply (shown once)
curl -s https://www.rightshold.com/v1/licenses/check -H "Authorization: Bearer $RIGHTSHOLD_API_KEY" \
-H "Content-Type: application/json" -d '{"license":"MIT","intended_use":{"commercial_use":true}}'