言語AIガイド
Hallucination Rates in Legal AI Tools
Independent testing has found that AI legal research tools built on retrieval-augmented generation (RAG) still give incorrect or unsupported answers a meaningful share of the time.
このページでは4 分で読めます
概要
A 2024 Stanford study found hallucinations in more than one in six queries even for the best-performing commercial tool. It matters because vendors marketed these tools as avoiding hallucinations, and lawyers who rely on them are still responsible for every citation they file.
ディープダイブ
Two Stanford-affiliated studies shaped the debate. The first, "Large Legal Fictions" (2024), by Matthew Dahl, Varun Magesh, Mirac Suzgun and Daniel Ho, tested general-purpose chatbots on verifiable questions about federal court cases. It found hallucination rates between 69 and 88 percent on those questions. Models also tended to accept false premises built into the questions. Commercial legal vendors argued that their tools avoided this problem through retrieval-augmented generation, which grounds answers in real legal databases. Some marketing used phrases such as "hallucination-free." In 2024, researchers at Stanford's RegLab and Institute for Human-Centered AI tested that claim in a study titled "Hallucination-Free?" They ran more than 200 legal queries through Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI, with GPT-4 as a comparison. They counted an answer as a hallucination if it was either incorrect or misgrounded. Misgrounded means it cited a real source that does not support the claim. Lexis+ AI performed best, yet still hallucinated on more than 17 percent of queries. Westlaw's tool hallucinated on roughly a third. Ask Practical Law AI often gave incomplete answers. An early version of the paper was criticized because the researchers first lacked access to Westlaw's research tool; they later tested it and revised the paper. Retrieval reduces errors but does not remove them, for several reasons. The retriever may find documents that share words with the query but are not legally relevant. It may miss the controlling authority, or pull a case from the wrong jurisdiction or a superseded version of a rule. The model may then misread or ignore what was retrieved, blend sources together, or go along with a false premise. A common misconception is that a correct-looking citation means a correct answer. The misgrounded category shows that a real citation can support a false claim.
戦略的影響
速度とスケール
言語ワークフローは、一貫性を犠牲にすることなく、より高速に移行できます。
アクセスと到達範囲
言語やコミュニケーション スタイルを超えてアクセスが拡張されます。
より明確な判決
自動化が繰り返しを処理する間、チームは判断により多くの時間を費やすことができます。
The Future of Hallucination Rates in Legal AI Tools
Vendors have updated their products since the 2024 studies, so specific error rates may not describe current versions. That is why repeated independent testing matters more than any single number. Benchmarking efforts by third parties and by law firms themselves are becoming more common. Better citators, agent-style workflows that check their own citations, and stronger grounding may reduce errors. Nothing so far shows that retrieval will remove them entirely. Courts continue to see filings with fabricated or misrepresented authority, and some have imposed sanctions. For the foreseeable future, human review of every cited authority remains a professional obligation, whatever a vendor claims.
現実世界の実装
An associate asks a research tool why a court ruled a certain way in a case that actually went the other way. The tool accepts the false premise and explains the invented ruling, citing a real opinion.
A tool returns a real appellate case with an accurate citation for a proposition the case never addresses. This is a 'misgrounded' answer that looks trustworthy until someone reads the opinion.
A research tool cites a decision that was later reversed, because it retrieved the older opinion and did not check its later history.
A legal department tests two vendors on its own set of 100 questions with known answers. It records whether each citation exists, whether it supports the claim, and whether the answer is complete.
リスクとガードレール
幻覚のような事実が、レポート、サポート フロー、または研究結果に静かに組み込まれる可能性があります。
迅速な対応により、同様のリクエスト間で一貫性のない結果が生じる可能性があります。
アクセス制御が弱いと、機密テキスト データが漏洩する可能性があります。
実装ロードマップ
展開する前に、出力形式、トーン、品質基準を定義します。
正確さが重要な場合は常に、信頼できる情報源を使って地上対応を行ってください。
一か八かの成果物については人間によるレビュー チェックポイントを維持します。
失敗パターンを追跡し、プロンプトやワークフローを定期的に再トレーニングします。
探検を続けましょう
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Hallucination Rates in Legal AI Tools quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
よくある質問
What is Hallucination Rates in Legal AI Tools?
Independent testing has found that AI legal research tools built on retrieval-augmented generation (RAG) still give incorrect or unsupported answers a meaningful share of the time. A 2024 Stanford study found hallucinations in more than one in six queries even for the best-performing commercial tool. It matters because vendors marketed these tools as avoiding hallucinations, and lawyers who rely on them are still responsible for every citation they file.
How did the Stanford "Hallucination-Free?" study define a hallucination?
The study counted two types of error: statements that are incorrect, and statements supported by a citation that does not actually support them.
In the Stanford study's terms, what is a 'misgrounded' answer?
A misgrounded answer points to a genuine authority, but that authority does not support what the answer says. This makes it look more trustworthy than it is.
Which tool performed best in the 2024 Stanford test but still hallucinated on more than 17 percent of queries?
Lexis+ AI had the best results, yet still hallucinated on more than 17 percent of queries. Westlaw's tool hallucinated on roughly a third.
What hallucination range did "Large Legal Fictions" find for general-purpose chatbots on verifiable federal court questions?
The study found that general chatbots hallucinated between 69 and 88 percent of the time on these specific, verifiable questions.
Why can a retrieval step find the wrong authority even when the database contains the right one?
Retrievers score closeness in wording or meaning. A dissent, a summary of the other side's argument, or a case from the wrong jurisdiction can rank highly without being controlling authority.
学び続ける
関連ガイド
このトピックのために選ばれたその他のガイド