AI in Litigation: A Hallucination and Verification Protocol

The principle

A growing body of guidance and discipline treats AI-assisted drafting the way it treats any other drafting: the lawyer who signs and files is accountable for what is in the document. ABA Formal Opinion 512 addresses generative AI and the duties of competence, confidentiality, communication, candor, and supervision — read it directly and confirm how your jurisdiction’s rules and any state or local ethics opinions track or depart from it, because adoption and interpretation vary. Courts across the country have sanctioned lawyers for briefs containing citations that did not exist; the specific decisions and penalties differ and continue to accumulate, so look to current authority in your own jurisdiction rather than rely on any summary.

The operative point does not change with the caselaw: a draft from a model is owned by the person who files it, exactly as a draft from a junior associate or a contract attorney is. “The tool generated it” is not a defense. This is general information, not legal advice, and AI rules are moving quickly — treat every item below as a prompt to verify current requirements, not a statement of settled law.

The verification protocol

Run before anything is filed, served, or sent to a client. Each item is nondelegable in substance even where the mechanical work is delegated.

• Verify every authority against the primary source. Pull each case, statute, rule, and regulation and confirm it exists, says what the draft claims, remains good law, and supports the proposition cited. A citation that is perfectly formatted and plausibly named can be entirely fabricated. The same applies to every quotation and pincite — read the underlying text, not a model’s paraphrase of it.
• Verify factual assertions, not just citations. Confirm record cites, dates, dollar figures, party names, and procedural history against the actual record and documents. Hallucination is not limited to caselaw; a fluent draft can invent facts as easily as authorities.
• Check the court’s standing orders and certification requirements. A number of judges have issued standing orders or local rules on AI use — some require disclosure, some require a certification that authorities were human-verified, some restrict use outright. These vary by judge and are proliferating; confirm the current requirements for the specific court, division, and judge before filing, and calendar a re-check, because they change.
• Confirm the tool was approved and confidentiality preserved. Entering case facts, client data, or privileged material into a consumer or non-approved AI tool can breach confidentiality and may risk waiver. Confirm the tool is on the organization’s approved list, that inputs were handled under its terms, and that nothing sensitive went into a system that trains on or retains the data.
• Address AI-generated and AI-handled evidence. Consider authentication, reliability, and potential challenges to anything AI produced or processed — summaries, translations, transcriptions, synthetic or enhanced media. Consider also that prompts, inputs, outputs, and the AI system itself may be discoverable; preserve them where a litigation hold may reach, and coordinate with your litigation hold.
• Document that a human verified the primary source. Record who checked each authority and assertion against the source, and when. If a court asks how the document was prepared, the answer should be a verification step a person actually performed — not an assumption that the tool, or someone else, had already checked.

Supervising outside counsel

The duty travels with the matter even when the drafting leaves the building. Many in-house teams find it useful to set the expectation in writing rather than assume it.

1. Ask outside counsel to confirm their AI use and verification practice.

   Not whether they use AI — most do, somewhere in the stack — but how they verify AI-assisted work product, who is accountable for it, and how they handle client confidential information in any tool.

2. Put it in the engagement letter and outside-counsel guidelines.

   Address permitted tools, human verification of all authorities and factual assertions, confidentiality of client data, compliance with court AI orders, and whether and how AI use will be disclosed to you. A pattern that has served clients well is to make verification an affirmative obligation of the engagement, not an implied courtesy.

3. Treat fabricated authority as a reportable event.

   If a hallucinated citation surfaces — in your own work or counsel’s — correct it promptly with the tribunal where candor requires, and treat the miss as a process failure to run down, not a one-off to bury.

Where the failures actually come from

Confidentiality and privilege questions around AI inputs deserve the same care as the citations; when in doubt about a specific tool, jurisdiction, or court order, confirm against current rules and the primary sources before relying on anything here.