AI Hallucinations Are Down 96%. Legal Is the Exception.

Summary of an essay from Dorna Moini's Legal Tech Notes, published on Substack on February 24, 2026. Dorna Moini is the CEO and Founder of Gavel and a former litigator at Sidley Austin. This page gives the argument and the numbers; the full essay is on Substack.

In brief: The best general AI models now make up facts less than 1% of the time when summarizing a document, down from roughly 20% in 2021. Legal is the exception: legal questions fail at about 6.4%, and a 2025 Stanford study found Westlaw AI and Lexis+ AI hallucinated on 17 to 33% of legal research queries. The essay explains why legal is harder and which practices actually reduce the risk.

Why does legal AI hallucinate more than general AI?

Legal questions are harder for three reasons the essay lays out. Many legal questions have no clean answer, so a model that is rewarded for answering confidently guesses. The law changes faster than training data, so a model can cite a rule that was amended last year. And the stakes are higher: a wrong restaurant recommendation costs a dinner, a fabricated case citation costs a sanction.

The numbers reflect that gap. On the Vectara leaderboard in late 2025, Gemini 2.0 Flash hallucinated 0.7% of the time and OpenAI o3-mini 0.8%, while general-knowledge questions across top models ran about 0.8% and legal questions about 6.4%. The Stanford study of Westlaw AI and Lexis+ AI, published in the Journal of Empirical Legal Studies, found error rates of 17 to 33% on legal research questions.

Why do hallucination benchmarks disagree with each other?

Because they measure different things. Vectara measures how often a model invents facts when asked to summarize a document it was given. Artificial Analysis measures how often a model answers when it should have said "I don't know," and on that test Claude 4.5 Sonnet answered when it should have abstained 48% of the time and GPT-5.1 51%. A vendor quoting one number without naming the test is not telling you much. Ask what the metric measures before you compare two tools.

Does retrieval-augmented generation fix the problem?

It helps, and it is not a cure. Retrieval-augmented generation (RAG) means the model is handed the relevant source documents at the moment of the question instead of relying on what it memorized in training. Vectara's data puts the reduction at about 71% compared with a model answering from memory. What RAG does not do is reason correctly across several sources, and a model can still misread a document it retrieved.

Asking a model to check its own answer, a technique tested in a 2024 study, cut hallucinations by a further 17%, but systematic errors survive self-review. The essay's position is that no single technique is enough: the working approach layers retrieval, legal-specific training, and a human who verifies the output.

Where is the hallucination risk highest in legal work?

Citation-specific research, especially on obscure or recent case law, is the highest-risk task, because the model is being asked to produce a specific fact it may never have seen. Risk is lowest on document-grounded tasks: summarizing, comparing or redlining a contract the lawyer has supplied, where the answer has to come from text in front of the model. The practical rule is to give the AI your actual documents and to treat any citation it produces as unverified until you have opened the case.

Courts and state bars are raising the diligence expectation at the same time. The essay's forecast, citing OpenAI research that better models improve mainly by flagging their own uncertainty, is that purpose-built legal tools reach the reliability of a careful first-year associate on well-defined tasks within a couple of years.

Key numbers and sources from the essay

• About 96% reduction in hallucination rates for the best models since 2021, from roughly 20% to under 1% (Vectara leaderboard, late 2025).
• Legal questions: about 6.4% hallucination rate among top models, versus 0.8% for general-knowledge questions.
• Westlaw AI and Lexis+ AI: 17 to 33% hallucination rate on legal research queries (Stanford study, Journal of Empirical Legal Studies, 2025).
• Vectara late 2025: Gemini 2.0 Flash 0.7%, OpenAI o3-mini 0.8%, DeepSeek V3.2 6.3%, GPT-5.2 8.4%, Claude Sonnet 4.5 above 10% (document-summary test).
• Artificial Analysis abstention test: Claude 4.5 Sonnet 48% and GPT-5.1 51% answered when they should have abstained.
• Retrieval-augmented generation: about 71% fewer hallucinations than a model answering from memory (Vectara). Self-checking prompts: 17% fewer (2024 study).

What this means for a law firm or legal team

1. Sort your AI use by task. Document-grounded work (review, summary, redlining of a contract you supply) carries the lowest risk; open legal research carries the highest.
2. Before buying, ask every vendor which benchmark their accuracy claim comes from and what it measures. A summary-fidelity number and an abstention number are not comparable.
3. Put a verification step in the workflow for anything that will be filed or sent: open every cited case, check every quoted clause against the source document.
4. Prefer tools that work from your documents and show their sources. Gavel Exec reviews and redlines inside Microsoft Word against the contract and a playbook, which is the document-grounded category the essay describes as lowest risk; see how the AI contract review tools compare.

Frequently asked questions

What is the hallucination rate of legal AI tools?

It depends on the task. Top general models hallucinate under 1% of the time when summarizing a supplied document, about 6.4% on legal questions, and the 2025 Stanford study measured 17 to 33% for Westlaw AI and Lexis+ AI on legal research queries.

Does RAG stop AI hallucinations in legal research?

Retrieval-augmented generation cuts hallucinations by about 71% compared with a model answering from memory, according to Vectara, but it does not eliminate them. Models can still misread retrieved documents and struggle to reason across several sources, so human verification remains necessary.

Which legal tasks are safest to use AI for?

Tasks grounded in documents you provide, such as contract review, summarizing an agreement or drafting a redline from a playbook. Citation-specific legal research on obscure or recent case law is the highest-risk task.

Read the full essay: AI Hallucinations Are Down 96%. Legal Is the Exception. on Substack, or subscribe to Dorna Moini's Legal Tech Notes.

Related on gavel.io: Gavel Exec AI contract review, Best AI contract review tools for lawyers, Gavel security.