Is it a fit?Book a call

Journal3 min

2,217 court decisions on AI hallucinations

A public database now counts more than two thousand rulings where a party relied on invented AI content, 879 of them involving lawyers. What it says to a firm choosing its AI.

A lawyer working through a file at a long wooden table by the window

Damien Charlotin’s database of AI hallucination cases lists court decisions in which a party was found to have relied on content an AI made up: cases that do not exist, quotes nobody said. At its update of 25 October 2026 it held 2,217 decisions, 1,514 of them in the United States.

  • 1,275 involve people representing themselves.
  • 879 involve lawyers.
  • 32 involve judges, and 17 experts.

It counts only decisions that discuss AI use beyond a passing mention, so the true number of mistakes is higher. The record is now long enough that no firm can call it rare.

The fault is not only the model

Most of these cases share a pattern: a general model, asked a legal question, answered fluently, and nobody checked the answer against a source before it reached a court. The fix is not to ban the tool. It is to change what the tool has to show.

What a firm’s own model does differently

  • Every answer names its sources, from the firm’s own files and approved precedents, so a person can check it in seconds.
  • It learns only from approved work. A correction is not training data until someone approves it.
  • It is tested on the firm’s own material, on a sealed set scored by its own lawyers, before any version goes live.
  • It says when it does not know. “No call” is an answer the model is allowed to give.

None of this removes the lawyer from the work, and it should not. It gives the lawyer something to check. That is what we build for law firms and in-house legal teams, and it starts with one task in a Test Lab.

Notes

  1. AI Hallucination Cases database, Damien Charlotin, update of 25 October 2026. Cases where AI use was only alleged are flagged by the database.

Read next

Newsletter

Release notes.

Once a month: what we released, what we held, and why. Short, dated and true, losses included.

Double opt-in. One click to leave. We never share the list.

Each issue holds

  1. One release or testWhat passed the gate, what was held, and the reason in one line.
  2. One Journal pieceAn explainer from the people who build it.
  3. One numberWith its baseline and its date, never on its own.

The first issue comes in November 2026.

Start

Bring one task. See the proof first.

In about four weeks, on your own data in a sealed setup, we show whether an Eigna expert beats your baseline. If it does not, you keep the report.

Book a callRequest a Test Lab