Hallucination decisions · Legal RAG limits

Does RAG reduce hallucinations enough for legal work?

Direct answer. No. Retrieval-augmented generation can ground a draft in documents you supply, but it does not make any published Truth Score a licence for final legal work. Claude Opus 4.8 at Truth Score 92 and Claude Sonnet 4.6 at 96 remain drafting marks. Keep a qualified lawyer on the signature.

RAG is often sold as the fix for legal hallucinations: retrieve the statute, then generate. That reduces some unsupported invention when the corpus is complete and the citation path is checked. It does not clear the decision-rights line on this site. Legal questions still need a specific authority a model can miss, mis-rank or invent around. The published Truth Scores — Claude Sonnet 4.6 at 96, Claude Opus 4.8 at 92, GPT-4o at 80, Mistral Large 3 at 80, Grok 4.1 at 79, DeepSeek V3 at 68 — measure drafting reliability, not permission to file. HIPAA and GDPR flags in the dataset are data-handling marks, not a bar licence. This page decides whether RAG alone is enough for legal work. It is not. Mitigation patterns stay on the hallucination-reduction guide; sector risk stays on the legal and medical page. The match engine may still choose which model drafts inside a RAG stack.

What RAG changes — and what it does not

RAG can attach retrieved passages to a draft and lower the chance of a wholly invented citation when the index is sound. It does not guarantee the right passage was retrieved, that the model quoted it faithfully, or that the legal conclusion is correct. Unsourced citation risk is still on unsourced citations.

Truth Scores stay drafting marks

ModelTruthPrivacyRole with RAG on legal text
Claude Sonnet 4.69694Strongest draft mark. Not the signature.
Claude Opus 4.89295Frontier draft for hard analysis. Still review.
GPT-4o8080Eligible on HIPAA true for data handling. Not counsel.
Mistral Large 38092EU path for document drafts. Lawyer still owns the filing.
DeepSeek V36852Hosted path fails GDPR/HIPAA on this record for client data.

Figures from /data/ai-comparison-2026.json. Last verified 2026-06-26. Scoring version v1.0. Affiliate links are off.

The decision for legal work

Use RAG to summarise and draft against a corpus a lawyer already trusts. Do not treat retrieval as a substitute for that lawyer. Block final filings and advice per when to block final decisions. The sibling page on legal and medical questions owns the HIPAA-as-data-gate argument.

Choosing a draft model inside a RAG stack

Prefer higher Truth and privacy marks when the corpus includes client matter. Opus and Sonnet lead those columns; DeepSeek V3's hosted privacy 52 is the wrong default for client PII. Methodology weights put safety and privacy at 10 percent each — another reason the overall weighted score cannot clear a legal answer. See methodology and best AI for document processing.

Pick the draft model, not the signer. Use the match engine for the stack. Keep the qualified reviewer on the filing.

FAQ

Is RAG enough to let a model file a legal brief?

No. RAG can ground a draft in retrieved documents. Final legal work stays with a qualified lawyer.

Which profiled model has the highest Truth Score for hard legal drafts?

Among Wave 2 profiles, Claude Opus 4.8 has Truth Score 92. Claude Sonnet 4.6 remains highest in the full dataset at 96.

Does HIPAA true mean RAG legal answers are cleared?

No. HIPAA in this dataset is a data-handling flag. It is not a licence to practise law.