CYPHR Register

Why we publish the source next to the answer

This page is about a measured problem in the general use of AI for legal questions. It is not a claim about any particular vendor, and it is not a claim that our own answers are beyond checking. It is the reason the Register is built the way it is.

The measurement

Stanford RegLab found that the AI legal research tools made by LexisNexis and Thomson Reuters each hallucinate between 17% and 33% of the time. The study is the first preregistered empirical evaluation of these tools, and it tested the products that law firms actually buy rather than a general chatbot.

Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning and Daniel E. Ho, published in the Journal of Empirical Legal Studies, volume 22, issue 2, pages 216–242 (2025), DOI 10.1111/jels.12413. We fetched the abstract carrying the figure above from the Stanford RegLab publication page on August 25, 2026; the publisher's own copy blocks automated retrieval, so we did not fetch it. An earlier May 2024 preprint of the same work reports a narrower finding across two tools, so the figure above is quoted from the published version, not the preprint. reglab.stanford.edu

What it looks like in court

Damien Charlotin's AI Hallucination Cases database listed 1,961 court decisions involving hallucinated material when we checked it on August 25, 2026. The database tracks decisions where a court or tribunal addressed the use of AI in more than passing reference, typically fake citations. That count was 1,959 on August 20, 2026. It moved by two in five days, which is why every figure here carries the date we checked it.

Damien Charlotin, AI Hallucination Cases database, checked August 25, 2026. The database describes itself as “1961 cases identified so far” and states that it is a work in progress. damiencharlotin.com

What follows from it

A confident answer and a correct answer look identical on the screen. The only thing that separates them is whether you can get to the source without taking anyone's word for it. So every Register entry carries the words of the instrument, a link to the official text, the date we fetched it, and a plain statement of what we did not check.

The last of those matters most. A gap we name is a gap you can work around. A gap we quietly fill is one you find out about later, from someone else.

We are not claiming our answers cannot be wrong. We are claiming you can check them, and that we tell you where we stopped looking. Read the Register, and hold us to it.