Verification in Legal AI Is a Design Problem

This is a preview of Mary Technology’s whitepaper, ‘Verification in Legal AI Is a Design Problem’. The full whitepaper can be downloaded at the bottom of this page.

Executive Summary:

In May 2025, an AmLaw 100 firm and a Los Angeles litigation boutique were sanctioned $31,100 over a brief built on invented authorities. The work had passed through paid, purpose-built legal research platforms used by experienced litigators. As of August 2026, Damien Charlotin’s public database of court decisions involving hallucinated AI material stood at 1,922 cases. The warning was clear, but the cases kept coming.

The failure of ‘be more careful’ was predictable because it left verification to vigilance rather than the workflow. Verification is a design problem: the system must make the basis and limits of its output easy to inspect before someone relies on it. If checking an answer means repeating the task that produced it, the time saving disappears, and under deadline pressure review becomes the step most likely to be cut.

The paper sets out five controls for a checkable workflow, what a buyer should ask to see for each, and why responsibility ends where it always has: with the person who signs. You can delegate the reading. You cannot delegate the signature.

Want to learn more?

Download the full whitepaper below, Verification in Legal AI Is a Design Problem, and see how to run the known-matter test on any legal AI tool your firm is evaluating.