How to Use AI in Litigation Without Getting Sanctioned (2026 Guide)

By John Mahoney · May 2026 · 11 min read

This is the practical playbook I wish someone had written before I started using AI in my own filings: what actually causes sanctions, what the courts require for disclosure, and the verification workflow that keeps you clean.

Why hallucinated citations keep happening

Large language models are next-token predictors. When asked for "a case that supports X," the model generates the most statistically plausible shape of a citation — a federal reporter abbreviation, a year, a numeric volume, a circuit. Many of these don't correspond to real opinions.

The model doesn't know it's making things up. There's no "I don't know" path in the generation. It will produce something that looks like a real citation until you check the reporter.

Two architectural patterns produce hallucinated citations:

1. Direct generation: the prompt asks the model to output case citations directly. The model invents whatever fits the rhetorical slot.
2. Insufficient grounding: the model has a retrieval step but the retrieval database doesn't contain a real-cite answer, so the model fills the gap with a plausible-looking fake.

Where the courts are actually drawing the line

From the published sanctions opinions (Mata, Park, Schwartz, the 2026 Oregon case, and the wave of state-court parallels), three rules emerge:

Rule 1: You are responsible for every citation. Period.

Courts have rejected every "the AI did it" defense. FRCP 11(b)(2) requires the attorney to have a "reasonable inquiry" basis for every assertion. Using AI doesn't transfer that obligation. The Mata court was clear: the technology is fine, the failure to verify is the misconduct.

Rule 2: AI disclosure orders vary by jurisdiction but the floor is rising

Some judges require AI disclosure on every filing. Some require it only for substantive briefs. A handful require nothing yet. By state and federal district:

• Texas: Many district judges require pre-filing AI-use certification.
• Illinois Northern: Sample standing orders attach a checklist.
• California Central: Pre-filing certification required by name in several courtrooms.
• State courts: NY, MD, IL, and TX all have at least some judges with standing orders.

Verify your specific judge's standing orders before each filing. The Above the Law and Artificial Lawyer trackers are useful starting points but not authoritative — read the order itself.

Rule 3: Work product protection covers AI-assisted analysis if you exercise "substantive intellectual engagement"

Heppner v. Lakeshore Medical Group (Feb 2026) held that AI-generated drafts under attorney direction qualify for work product protection — but only if the attorney exercised substantive intellectual engagement with the AI output. A copy-paste workflow doesn't qualify.

The verification workflow that keeps you clean

This is what I use on every filing that involves AI-assisted drafting:

1. Generate with a hallucination-resistant tool. If the tool outputs direct case citations, treat every citation as a hypothesis to verify. If the tool outputs search queries, you're already partway clean.
2. Run every citation through Westlaw, Bloomberg Law, or CourtListener. If it doesn't pull up, don't cite it. There is no other safe path.
3. For medical literature, run every claim through PubMed. Same rule: if it doesn't pull, don't cite it.
4. Read the actual opinion before quoting from it. The model's summary of a case is often wrong even when the citation is real. Quote from the opinion itself.
5. Document your verification. Save the Westlaw / PubMed printout. If a clerk or opposing counsel raises a question, you have receipts.
6. Check your judge's standing orders before filing. If an AI disclosure is required, include it.
7. Have a human review. Not the same person who used the AI. Cross-eyes catch the hallucinations you missed.

Tools that reduce verification load (and tools that increase it)

Tools that REDUCE verification load:

• Westlaw / Lexis-native AI assistants — citations are pulled from the verified database. You still check, but the hallucination floor is lower.
• Tools that output search queries instead of citations — e.g., our analyzer outputs a PubMed query like "knee dislocation AND popliteal artery injury" rather than a specific paper. You grab the real paper from PubMed yourself.
• Retrieval-augmented tools with strict source-only citation — the AI cannot cite anything outside the retrieval set.

Tools that INCREASE verification load:

• General-purpose LLMs (ChatGPT, Claude direct) — useful for drafting but never trust a citation that comes out of them.
• ChatGPT wrappers labeled "Legal AI" — many of these are thin layers over the base model with no retrieval grounding. Treat citations as adversarial.
• Tools that mix retrieved and non-retrieved citations in the same output — you can't tell which is which without trace-back.

Our architecture (and why we built it this way)

When we built MedLegal AI's Courtroom AI deposition analyzer, the hallucination question was the load-bearing design decision. The analyzer hits a clinical claim made by a witness ("knee dislocation wouldn't have mattered for the management"). What does it do next?

The cheap version: "The analyzer cites Smith et al., J. Orthop. Trauma 2014, on missed-dislocation outcomes." Looks great. Doesn't exist. Mata sanctions await.

The architecture we shipped: the analyzer outputs a PubMed search query — "knee dislocation AND popliteal artery injury" — plus an FRE 803(18) (Learned Treatise) foundation script. The attorney clicks through to PubMed, picks the actual paper that supports their case, lays the foundation from a real published source. The hallucinated-citation failure mode is eliminated at the architectural level.

We stress-tested this against a real 2-hour public Maryland medical malpractice deposition — Tolson v. St. Agnes, published by Miller & Zois as legal-education content. 22 of 22 critical impeachment signals fired correctly. Zero fabricated case citations. Full case study with verbatim analyzer JSON output here.

Practical checklist for the next time you file an AI-assisted brief

1. ☐ Check the judge's standing orders for AI disclosure requirements
2. ☐ Run every citation through Westlaw / Bloomberg / CourtListener — kill any that don't pull
3. ☐ Read the actual opinion (not the AI's summary) before quoting
4. ☐ Save verification receipts (PDF of the Westlaw result)
5. ☐ Include AI disclosure if required by the court
6. ☐ Have a second attorney review the brief and re-verify the citations
7. ☐ Document your AI-use workflow in your case file (for Heppner-style work-product protection)

FAQ

Is using AI in litigation per se sanctionable?

No. The sanctions are about failure to verify, not about AI use. Every federal opinion since Mata has been clear on this.

What if my AI tool says it doesn't hallucinate?

Test it. Ask for cases supporting a specific narrow proposition, then verify every cite. Even tools that "don't hallucinate" can return citations that look right but are misattributed.

Can I use AI without disclosing it?

If your judge's standing order doesn't require disclosure, you don't have to disclose. Many states (NY, IL, TX, CA) have at least some judges that do require it. Check before filing.

What about AI-drafted demand letters and discovery requests?

The same rules apply, but the stakes are lower because these don't go to a judge. Still: verify every legal citation, every statutory reference, and every factual claim that you don't have personal knowledge of.

— John Mahoney
medicalai.law / [email protected]
Not legal advice. Consult your bar's professional responsibility rules and your court's standing orders.