Citation Safe vs. Lexis+ AI

A Stanford study specifically evaluating commercial legal AI research tools — not general-purpose chatbots — found that Lexis+ AI and similar retrieval-augmented products from Thomson Reuters hallucinated in roughly 17% to 33% of tested queries, with Lexis+ AI's rate on the lower end of that range at approximately 17%. This is a meaningful improvement over general-purpose model hallucination rates (which the same research group had earlier measured at 58%–88% for raw language models on legal queries), reflecting the real value of retrieval-augmented grounding in an actual legal database. It is not, however, a rate close enough to zero to skip independent verification.

Citation Safe is not a research tool, so we are not making a claim to have a lower hallucination rate on legal research queries — that's not what we do. What we offer is a separate, deterministic verification layer that checks the citations in a document after drafting, regardless of whether Lexis+ AI, another AI tool, or a human associate produced them. Our existence and quote-match layers run without an LLM at all, so they are structurally incapable of the specific failure mode (hallucination) the Stanford study measured in generative research tools.

If your firm has a Lexis subscription and uses Lexis+ AI for research, that is a reasonable, well-supported research tool with real capability behind it. Given the independently measured 17% hallucination rate on the specific tested queries, we'd suggest treating a Lexis+ AI-assisted first draft the same way you'd treat a draft from any AI tool: verify the citations independently before filing. That is exactly the workflow Citation Safe is built to support, regardless of which research tool sits upstream of it.

It's a common assumption that grounding an AI tool in a real legal database (retrieval-augmented generation, or RAG) should eliminate hallucination entirely. The Stanford research shows this isn't quite right: RAG substantially reduces hallucination compared to a raw language model with no database grounding, but the underlying language model can still misinterpret retrieved passages, blend information from multiple sources incorrectly, or generate a citation that doesn't precisely match what was actually retrieved. Grounding helps; it doesn't fully solve the problem, which is exactly why the measured rate for a RAG-based tool like Lexis+ AI is meaningfully above zero rather than at it.

We'd genuinely welcome LexisNexis publishing its own current, continuously updated hallucination rate for Lexis+ AI, independently audited if possible, rather than leaving this to a single (excellent, but necessarily time-limited) academic study. Until that exists, the Stanford figures remain the most credible independent data point publicly available, and we think it's more honest to cite that specific, sourced number than to make a vague claim that AI research tools are "sometimes inaccurate" without a checkable figure attached.

It's also worth noting that hallucination rates measured in academic studies like this one are aggregate figures across many tested queries; your own actual risk on any specific research query depends on factors the aggregate number can't capture, including how novel or well-settled the underlying legal question is. This is one more reason to treat the 17% figure as a strong signal to verify rather than a precise prediction for any single citation you happen to generate.

Lexis+ AI's independently measured ~17% hallucination rate is a real improvement over raw AI models but far from zero. Citation Safe adds a separate, deterministic verification layer regardless of which research tool you use.