Original edition · Sept. 5, 2026
Built by Kyle Bahr with Claude Fable 5.1 and GPT-5.6 Sol.
The original study, with corrections through Sept. 6.
- Ten sections
Two editions · One dataset · September 2026
Damien Charlotin’s AI Hallucination Cases database passed 2,000 decisions. Read the study in two editions.
Find additional discussion in the original LinkedIn thread.
Expanded edition · Sept. 9, 2026
Built by Kyle Bahr with GPT-6 Astra and Claude Fable 5.1.
Rebuilt with the newest frontier models.
- Thirteen sections
Switching between editions. Each page carries an Original | Expanded control beside the theme button. It opens the same section in the other edition.
What the expanded edition adds
Three of the additions came from comments on the original LinkedIn thread.
01
What about gender?
The pronouns that court documents and professional biographies use for the 420 entries in the study’s identified-lawyer roster, with the limits stated plainly.
Inspired by a LinkedIn comment from Carolyn Elefant.02
Before ChatGPT
What the public record can and cannot say about false authority and factual-record errors before generative AI: fourteen selected historical examples (including a quoted anecdote), six sources compared, and a proposal to estimate a before-and-after rate.
Inspired by a LinkedIn comment from Doug Hall.03
Contract review
Six exercises and a checklist for transactional lawyers, and why a model’s draft goes wrong in deal work.
Inspired by a LinkedIn comment from Rafie Faruq.04
Litigation labs
Six exercises on fictional AI output, each a decision a litigator has to make, with the exact repair. The original real-briefs training is kept.
05
Verification workbench
Compare a quotation with its source word by word, work the filing checklist, keep a handoff record, and download the templates and the team exercise pack.
06
Sources and method
How the edition was made: the instruction Kyle gave GPT-6 Astra, how Codex and Claude reviewed each other’s work, and what was verified against which source.