Citation Safe vs. Westlaw AI-Assisted Research

The same Stanford study evaluating Lexis+ AI also tested Westlaw AI-Assisted Research and Thomson Reuters' Ask Practical Law AI, finding hallucination rates toward the higher end of the 17%–33% range measured across the tools studied, with Westlaw AI-Assisted Research specifically cited around 33%. That is a one-in-three rate of hallucination on the tested queries — a genuinely significant number for a tool whose entire purpose is producing reliable legal research, and a strong argument for independent verification of anything it produces before it reaches a filing.

Many firms have used Westlaw for decades and reasonably trust the underlying case law database itself, which is not in question here — the 33% figure describes the AI-assisted research layer's hallucination rate, not the accuracy of Westlaw's underlying primary-source database. A citation that Westlaw's AI assistant generates and a citation you find yourself through Westlaw's traditional search are different things: one has passed through a generative AI layer with a measured, non-trivial error rate, the other is you reading the primary source directly. Independent verification matters specifically for AI-generated output, not for traditional keyword or Boolean research.

Citation Safe does not ask you to give up your Westlaw subscription or your trust in the underlying case law database. It adds a separate, deterministic check specifically on whatever citations end up in your final draft — whether they came from Westlaw AI-Assisted Research, traditional Westlaw search, another tool entirely, or an associate's memory. Given the independently measured 33% figure for Westlaw's AI-assisted layer specifically, we'd suggest treating any AI-generated Westlaw output the same way you'd treat output from any other generative AI tool: verify before filing.

The same Stanford research measured Lexis+ AI at roughly 17% and Westlaw AI-Assisted Research at roughly 33% — nearly double. Neither figure should be read as a precise, permanent measurement of either product's current performance, since both vendors continue to update their underlying models and retrieval systems, and hallucination rates in this category have generally trended downward over time as the technology matures. What the comparison does establish clearly is that meaningful, measurable gaps exist between competing commercial legal AI research tools, and that no vendor in this specific study, including the one measured lower, approached a rate low enough to skip independent verification.

As with Lexis+ AI, we'd welcome Thomson Reuters publishing its own current, continuously updated, ideally independently audited hallucination rate for Westlaw AI-Assisted Research, rather than leaving the most credible public data point to a single academic study that will inevitably age. Until that happens, we think citing the specific, sourced Stanford figure is more useful and more honest to readers than a vague reassurance in either direction — for us or against Westlaw.

It's worth noting that a 33% aggregate hallucination rate doesn't mean one in three of your own specific research results will necessarily be wrong — actual risk varies by query type and legal area, and the aggregate figure is a signal to verify systematically rather than a precise per-query prediction. Either way, a rate this far from zero is a strong argument for treating any Westlaw AI-Assisted Research output the same way you'd treat output from any other generative AI tool.

Westlaw AI-Assisted Research's independently measured ~33% hallucination rate is the highest figure in the commercial tools Stanford tested. Citation Safe's deterministic layers provide an independent check on anything it produces.