Tochi Eke-Okoro

A lawyer filed a brief citing eight cases that don't exist. Opposing counsel caught it. So he withdrew them and submitted a corrected list. Not one of the eight he withdrew matched the eight he had actually made up. And all eight of the replacements were fake too. The Eleventh Circuit published this on July 10. Judge Britt Grant, writing for a unanimous panel: "Whatever the merits of artificial intelligence, it is no substitute for actual intelligence." The attorney has been referred for discipline. The easy takeaway is "lawyers need to verify AI output." Everyone is going to post that this week. I think it misses what actually happened. There's a footnote in the opinion I can't stop thinking about. The court noted that AI is notorious for producing outputs that echo the user's own beliefs, even when those beliefs are wrong. And that this is especially dangerous in law, because a lawyer's job is to tell a client whether the law actually says what the client wants it to say. Sit with that for a second. The tool tells you what you want to hear. And then we make the person it just flattered responsible for catching it. That is not a control. That is a hope. Verification-by-human-diligence fails exactly when you need it most. When you're tired. When the filing is due tomorrow. When the fake case says precisely what your argument needed it to say. He didn't skip that step once. He skipped it twice, under scrutiny, after being warned. Policies existed. Warnings existed. It happened anyway. So I don't think this is a training problem or a policy problem. If a system is capable of emitting an unverified citation, eventually it will. The only durable fix is to make that outcome structurally impossible: the output does not ship unless every citation checks out against real authority, and when support can't be established, the system falls back to a deterministic answer instead of a confident one. That's what I've spent the last stretch building. It's called ThesisLogic. It's open source, and the whole architecture rests on one rule — no generative output leaves the system with an unverified citation. Curious where people land on this. Is the answer better lawyers, or better systems? Because we keep trying the first one, and it keeps not working.