AI Hallucinations in Legal Practice: A Field Guide

Everyone in legal tech has heard about Mata v. Avianca. Two attorneys submitted a brief packed with invented citations, got sanctioned, and became the cautionary tale that launched a thousand CLE presentations. It's 2026. If Mata is still the ceiling of your understanding of AI hallucination risk, you are dangerously behind.

Hallucinations are not a bug waiting to be patched. They are an architectural feature of how large language models work — probabilistic text completion dressed up as legal authority. The risk hasn't disappeared; it has matured, diversified, and in some cases gotten harder to detect. Here's what you actually need to know.

The Taxonomy: Three Distinct Failure Modes

Legal professionals tend to treat "hallucination" as a monolithic problem. It isn't. There are three meaningfully different failure modes, each requiring a different detection strategy.

Citation Fabrication is the Mata problem — the model invents a case name, docket number, court, and holding from whole cloth. It sounds authoritative because the model has learned what a citation looks like, not what one is. This is actually the easiest type to catch because the artifact either exists or it doesn't. A Westlaw search takes thirty seconds.

Statute Misquotation is subtler and considerably more dangerous. Here, the statute or regulation cited genuinely exists, but the model gets the text wrong — sometimes by a single word that completely inverts the meaning of a provision. In practice, this shows up most aggressively in regulatory work. Ask an LLM about the specific numerical thresholds in OSHA's Hazard Communication Standard or the exact definitional language in the CCPA's "sale of personal information" provisions, and you will frequently get text that resembles the real statute but departs from it in ways that matter. No fabrication — just distortion. The citation checks out. The text does not.

Case Law Distortion is the hardest to catch and, in my view, the most legally consequential. The case is real. The citation is accurate. The court said something like what the model claims. But the holding has been subtly reframed — a dissent described as majority reasoning, a narrow holding generalized into a broad rule, a case distinguished on facts the model chose not to mention. This is what happens when you ask an LLM to summarize TransUnion LLC v. Ramirez (2021) for a standing argument. The model knows the case. It also knows what you want the case to say.

Which Tasks Carry the Highest Risk

Not all legal work is equally exposed. The risk profile varies by task type.

Highest risk: Brief-writing and memo drafting that requires citation support. The model is under implicit pressure to produce authority, and it will produce authority, real or not. Secondary research synthesis — asking an AI to summarize what "the courts have held" on a particular issue — is particularly dangerous because distortion errors propagate invisibly through the summary.

High risk: Regulatory compliance advice. Regulations change frequently, and LLM training data has cutoff dates. The model may confidently describe a pre-amendment version of a rule as current law. The FTC's revised Safeguards Rule, the SEC's cybersecurity disclosure requirements that went into effect in 2023, state-by-state data privacy legislation that has been amended multiple times — these are active hallucination targets.

Moderate risk: Contract drafting and clause generation. Here the model is generating language rather than citing authority, so the fabrication vector is less relevant. The risk shifts to importing standard market terms that don't reflect current practice or that contain embedded legal conclusions the drafter hasn't evaluated.

Lower risk (but not zero): Summarizing documents the model can actually see. When you provide the full text and ask for a summary, the factual anchor reduces hallucination risk. It doesn't eliminate it — models still occasionally misread or selectively emphasize — but the failure rate drops substantially.

How to Catch Them: A Practical Protocol

Verification has to be systematic, not aspirational. "I'll spot-check it" is not a protocol; it's a hope.

For citation fabrication: every case and statute cited in any AI-assisted work product must be verified in a primary legal database before the document leaves your desk. No exceptions based on how confident the output sounds. Confidence is not correlated with accuracy in language models.

For statute misquotation: pull the primary source and compare language side-by-side. Do not trust a paraphrase. Do not trust a summary. The actual statutory text takes two minutes to retrieve from Cornell's LII or a government database. That two minutes is not optional.

For case law distortion: this requires actually reading the case, or at minimum the headnotes and the relevant section of the opinion. You need to verify not just that the citation exists, but that the proposition it's being used to support is genuinely what the court held, in that procedural posture, on those facts. Consider building a three-question check: Is the cited proposition in the majority opinion? Is it a holding or dictum? Does the factual context match your case closely enough to be relevant?

Some firms are now running AI outputs through second-pass AI checkers specifically trained for citation verification. These tools — including emerging features within Westlaw and Lexis's own AI platforms — can flag low-confidence citations automatically. They're useful. They are not a substitute for attorney judgment.

The Competence Obligation Is Not Abstract

The ABA's Formal Opinion 512, issued in 2024, makes clear that competent use of AI tools includes understanding their limitations. State bars in California, Florida, and New York have followed with their own guidance. The disciplinary exposure from an unchecked hallucination in a filed document is real, and "the AI wrote it" has never been a recognized defense to a Rule 3.3 violation.

Mata was a wake-up call. The appropriate response to a wake-up call is to stop hitting snooze. Know what type of hallucination you're looking for, know which tasks generate the most risk, and build verification into the workflow before the document goes out the door — not after opposing counsel calls you about it.