Update February 20, 2026

Since we first published this post, the stakes surrounding AI in the legal profession have increased. Alongside cases emerging nationwide involving issues with artificial intelligence-induced citation hallucinations, there are increasing instances of attorneys using heretofore trustworthy sources for research that have been found to be laden with errors. We have become aware of several instances in which attorneys conduct good-faith research on seemingly trustworthy legal research sources only to find out the cases cited for an issue do not match the citation. Some of these attorneys now find themselves ensnared by State Bar investigations, facing assertions of incompetence or making misrepresentations to the court.

Attorneys receiving citations from individual courts for incorrect information or citations are sometimes required to report themselves to the State Bar, and we know of instances in which this led to the initiation of a State Bar investigation. Furthermore, these cases are approached by the bar as issues of moral turpitude, which can have serious consequences up to and including disbarment. The California State Senate has introduced a bill that would require lawyers to safeguard confidential information and prevent it from being entered into AI systems, as well as to personally review any AI-generated work and prohibit decision making to machines. These updates further underscore the need for extreme caution when considering the use of AI in anything to do with the practice of law.

Attorneys are submitting briefs to courts containing fabricated case citations, fabricated quotations, and entirely invented legal authorities. These are not just careless mistakes or isolated incidents. They are the predictable result of delegating critical legal work to artificial intelligence systems that their vendors claim are reliable but demonstrably are not.

In Noland v. Land of the Free, L.P., 2025 WL 2374381 (Cal. Ct. App. Sept. 12, 2025), the California Court of Appeal confronted a recent example of this in which counsel submitted an opening brief in which 21 of 23 legal quotations were entirely fabricated. The attorney had used generative AI tools to draft the briefs, then used additional AI tools to verify the citations—a practice equivalent to fact-checking rumors by asking about other rumors. He never consulted the cases themselves. The court’s response was surgical: a $10,000 sanction, a referral to the State Bar of California, and notification to the client.

But the damage extends far beyond one attorney’s misconduct.

What Are AI Hallucinations?

When generative AI systems encounter a request they cannot answer from their training data, they are less likely to say they don’t know the answer. Instead, they will make one up. These made-up sources tend to appear entirely authentic (complete with case names, citations, court designations, etc.). The systems do this out a desire to be helpful to the user (uncertainty is very unhelpful!).

Large Language Models (AI systems) are pattern-matching systems. They are not capable of reasoning. Instead, they take input data, compare it to patterns in the data they already have in their system, and produce a guess at the correct answer. Sometimes their guesses are really good, but they do not understand law, logic, or arguments. In fact, they cannot think like a lawyer. They cannot think at all.

Still, vendors (often recent undergraduates or college dropouts) market these tools as reliable legal research platforms. They explicitly claim the tools are “hallucination-free” and these claims are easily disproven. Research has shown disturbingly high levels of hallucinations across all of the readily available AI legal research tools, including both the more general tools (e.g., Chat-GPT or Claude) and the specialized ones marketed to lawyers specifically (e.g., Lexis Nexis and Thomson Reuters). The marketing is incorrect.

The Scope of the Legal Hallucination Trend

A comprehensive study of 114 federal court cases involving AI hallucinations in attorney filings reveals a troubling pattern. Small law firms account for the vast majority of incidents, but this is likely only because larger firms have more resources to catch errors (i.e., multiple attorneys reviewing each document before submission). The legal profession is experiencing economic strain across the board, and it’s likely that AI tools will grow in popularity as more and more firms struggle to manage large workloads and growing price sensitivity. AI tools offer the illusion of sophistication and efficiency while delivering a new way of making the same mistakes.

A Profession-Wide Problem Requiring a Profession-Wide Response

The American Bar Association issued formal guidance on AI use in July 2024. State bars have begun developing continuing legal education requirements. Yet incidents continue at an accelerating pace. Dozens of new cases have been added to hallucination databases since the ABA’s guidance was published.

Individual attorney vigilance is necessary but often insufficient. The legal profession must treat AI hallucinations as a systemic threat requiring coordinated action: mandatory, rigorous CLE on AI risks; clear ethical guidelines prohibiting delegation of citation verification to machines; bar association monitoring of AI tool performance; and heightened scrutiny of filings incorporating AI research.

But the most essential rule remains unchanged and cannot be automated. Read your briefs. Check your citations against primary sources. Verify everything yourself. Understand that the tools you are relying on are fundamentally less reliable than their marketing claims suggest.

AI can assist with legal work. Verification, however, cannot be delegated to machines. The profession must treat that distinction as non-negotiable, or it will continue paying the price in sanctions, bar discipline, and lost public confidence.

< See all News / Events