We respect attorney-client confidentiality. No tracking pixels in our emails.
After Mata v. Avianca set the template for AI sanctions, courts have continued to penalize attorneys for unverified AI-generated citations. This 15-point checklist closes the gaps.
Move from this guide to practical legal AI tool shortlists and comparison pages.
Judge Kevin Castel's June 2023 order in Mata v. Avianca did not merely sanction two attorneys and their firm $5,000 each. It established a template that courts across the country have followed in dozens of subsequent cases — a template for what judicial patience for AI-assisted negligence looks like when it runs out. By mid-2026, sanctions orders citing Mata number in the hundreds. The pattern is consistent: an attorney uses an AI tool to generate case citations, submits a brief without verifying those citations against an authenticated legal database, and opposing counsel or the court discovers that the cited cases do not exist or say something materially different from what was represented. The follow-on consequences — sanctions, referrals to state bars, malpractice claims, and in some cases client fee disputes — have become a predictable sequence.
The problem is not that attorneys are using AI. The problem is that many are using it without the workflow controls that distinguish acceptable assistance from negligent delegation. This article provides a 15-point risk mitigation checklist, examines the specific failure modes courts have identified, and outlines what your malpractice carrier expects to see when you disclose AI use in your practice.
The legal malpractice framework has not changed because of AI — negligence is still negligence, and the standard of care is still what a competent attorney in the relevant jurisdiction and practice area would do. What has changed is the set of ways in which attorneys can fall below that standard, and the speed at which they can do so.
Before AI legal research tools, a citation error in a brief typically reflected a human mistake: misremembering a citation, transposing numbers, or relying on a secondary source. These errors were uncommon because the friction of manual research created natural checkpoints. AI tools remove that friction — and with it, some of the natural verification behavior that friction produced.
AI hallucination in legal contexts refers specifically to the generation of plausible-sounding but factually false legal citations, case summaries, or statutory text. Large language models generate text by predicting what should come next based on training data patterns. When asked about a specific case or statute, a model that lacks that information in its training data does not say "I don't know" — it generates something that looks like what an answer would look like. The result is a confident, formatted, authoritative-sounding citation to a case that has never existed.
Courts have been unsparing in their analysis of attorney responsibility for these errors. In Mata and its successors, judges have consistently rejected arguments that the attorney reasonably relied on the AI tool, that the error was the vendor's fault, or that the attorney lacked sufficient AI literacy to know verification was required. The holding is uniform: submitting a document to a court is an attorney's certification of its accuracy, full stop.
The bar has followed. California, New York, and Florida have all issued guidance making clear that AI-assisted work product is subject to the same competence standard as work product produced by any other method, and that the competence obligation includes verifying AI output before reliance.
Pre-Research Phase
Research and Drafting Phase 4. Use AI tools only for initial research framing and citation identification — never as the sole source for any citation you will submit to a court or include in a client deliverable. 5. For every case citation generated by AI, verify the citation exists, verify the citation says what the AI summary claims, and verify the citation remains good law using a citator such as Westlaw KeyCite or LexisNexis Shepard's. 6. Maintain a verification log for each matter documenting which citations were AI-assisted and confirming that each was independently verified, by whom, and when. 7. Never copy AI-generated text directly into a filing without reviewing the underlying sources. AI summaries of cases frequently mischaracterize holdings, overstate the scope of decisions, or miss subsequent negative history. 8. Pay particular attention to AI output regarding recent developments — tools with training data cutoffs will not reflect cases, regulations, or statutes enacted after that cutoff.
Review Phase 9. Implement a second-reviewer requirement for any brief or document containing AI-assisted research that will be filed with a court or sent to a client as legal advice. 10. The second reviewer should specifically check: that all citations are real, that they stand for the propositions cited, and that none have been overruled or distinguished in ways that affect the argument. 11. Before filing any document with a court, run a final check on all citations using an authenticated legal database — not the AI tool that generated them.
Client and Court Disclosure 12. Develop a standard client disclosure for your engagement letters explaining that the firm uses AI tools in legal research and drafting, that all AI output is subject to attorney review and verification, and identifying any limitations this creates. 13. Monitor court-specific AI disclosure requirements — many federal and state courts now require disclosure of AI assistance in filings. Non-compliance is itself a sanctionable event independent of any underlying error.
Documentation and Insurance 14. Retain AI-assisted work product and verification records for the same period as other matter documentation. In a malpractice dispute, the question of whether you verified AI output is answered by your records, not your memory. 15. Disclose AI use to your malpractice carrier at renewal and ask specifically whether your policy covers claims arising from AI-assisted work product. Some carriers now offer enhanced AI-specific coverage riders.
The supervision failures in post-Mata sanctions cases share a common structure: a senior attorney assigns AI-assisted research to an associate or paralegal, receives a draft brief with citations, reviews the argument but not the underlying citations, and files the brief. When errors surface, the senior attorney claims they relied on the subordinate's verification. Courts have not been receptive to this defense.
Rule 5.3 of the ABA Model Rules requires partners and supervising attorneys to make reasonable efforts to ensure that the work of non-attorney assistants conforms to the Rules of Professional Conduct. Several bar opinions have now applied this rule to require that supervising attorneys specifically understand what AI verification procedures their subordinates are following — generic supervision is not enough.
Verification without documentation is legally indistinguishable from no verification. In a bar proceeding or malpractice suit, what happened is what your records show happened. Every checklist item above that involves verification should produce a written record: a log entry, an email confirmation, an annotated draft showing that citations were checked against authenticated sources.
This sounds bureaucratic. It is less bureaucratic than a bar investigation.
Litigation carries the highest immediate exposure because citation errors are visible to courts and opposing counsel. Immigration law has specific risks because regulatory citations in this area change frequently and AI tools with older training data may generate outdated authority. Transactional work has subtler but significant risk in the form of incorrect summaries of legal standards — an AI-generated description of disclosure requirements under securities law or environmental compliance standards may sound authoritative while being materially wrong.
In March 2025, an immigration attorney in the Southern District of New York filed a brief citing three administrative cases supporting a particular asylum claim standard. Opposing counsel's research found that one case had been vacated on appeal two years earlier, one had a citation error in a footnote that led to a nonexistent decision, and the third was real but had been distinguished in subsequent BIA decisions in ways that undermined rather than supported the argument.
The attorney's AI research tool had generated all three citations. The attorney had reviewed the AI summaries but had not independently verified the citations against Westlaw or LexisNexis. The court issued an order to show cause. In the attorney's response, she submitted her ChatGPT conversation logs showing she had explicitly asked the tool to provide verified citations and it had represented that they were accurate.
The court's order was blunt: the attorney's certification on filing was her personal representation, not the AI tool's. She was sanctioned $3,500, required to complete six hours of continuing legal education on technology competence, and the order was forwarded to the state bar for review. The client's case survived — but only because the attorney was able to find the same underlying legal principle supported by actual, verified authority.
For verifiable, citation-safe legal research that reduces hallucination risk:
See also: Casetext vs CoCounsel for a detailed comparison of research accuracy.
Q: Our firm uses Harvey AI which is built on GPT-4. Is that safer than using ChatGPT directly for legal research?
A: Meaningfully safer for authenticated research tasks, because legal AI platforms like Harvey are typically retrieval-augmented — they ground responses in authenticated legal databases rather than generating from training data alone. But the verification requirements are the same regardless of tool. No AI tool is certified to be hallucination-free.
Q: What does a compliant AI use policy actually need to include to satisfy malpractice carrier requirements?
A: Carriers vary, but a policy that specifies approved tools, prohibited uses (e.g., no AI for final citation sourcing without verification), verification requirements, documentation requirements, and a named compliance owner will satisfy most carrier disclosure requests. Get this reviewed by coverage counsel.
Q: If we catch an AI error before filing and correct it, do we have a disclosure obligation to the client?
A: Generally no — catching and correcting errors before they cause harm is exactly what competent supervision looks like. But if the error affected advice you already gave the client that influenced a decision, that is a different analysis and you should consult with ethics counsel.
Q: Can courts require disclosure of AI use in every filing, or only when there is a reason to suspect errors?
A: Courts have broad authority to impose disclosure requirements through local rules and standing orders, and many have done exactly that — requiring disclosure of any AI use in brief preparation regardless of whether errors occurred. Check the local rules and standing orders of every court in which you file.
Q: Our malpractice insurer just asked us to complete an AI questionnaire at renewal. What red flags should we avoid?
A: Do not represent that you have a written AI use policy if you do not have one, or that you have verification procedures in place that you cannot document. Insurers are conducting post-claim investigations and inconsistencies between renewal representations and actual practice have become coverage defenses in several reported claims.
AI malpractice risk is not theoretical in 2026 — it is documented, litigated, and sanctioned. The Mata v. Avianca template has been applied in hundreds of subsequent cases, and the legal standard is clear: attorneys are responsible for everything they certify to a court or provide to a client as legal advice, regardless of how it was generated.
The 15-point checklist in this article addresses the full lifecycle of AI-assisted work product: policy, training, research verification, drafting review, second-review requirements, client disclosure, court disclosure, documentation, and insurance disclosure. Firms that implement these controls systematically are not just reducing malpractice exposure — they are building the documentation record that protects attorneys in the bar proceedings and coverage disputes that follow when something goes wrong despite best efforts.
The attorneys who get sanctioned are almost never the ones who were careless about legal quality generally. They are attorneys who applied their existing high standards to their own legal reasoning while treating AI output as a reliable input rather than an unverified draft. Closing that gap is the entire project.
This article reflects independent editorial analysis. LawyerAI does not accept payment for editorial coverage. Tool scores are based on methodology described in Our 5-Dimension Methodology. Last reviewed: 2026-08-22.
