The Hollow Trial

Both the plaintiff's attorneys and the defense had used AI to generate their legal arguments. The judge, discovering this mid-proceedings, did the reasonable thing when handed a container with no contents: cancelled the trial. Kicked everyone off the case. Started over with new counsel.

The legal profession's response has been to frame this as a disclosure problem — attorneys who forgot to declare their tools, procedural violations that better policies could fix. That framing is too small. What happened wasn't a disclosure failure. It was a structural collapse.

Adversarial law runs on a specific premise: two human minds, each committed to their client's position, create friction that surfaces truth. The plaintiff's attorney believes — or performs believing — that their client was wronged. The defense believes the opposite. The clash isn't procedural formality. It's epistemically load-bearing.

When both sides are generated by the same probabilistic text engine, that friction disappears. An LLM doesn't have a client. It has a prompt. Feed it "argue that the defendant is liable" and it produces compelling arguments. Feed it "argue the defendant is not liable" and it produces equally compelling counterarguments. Not because it understands the facts — because it has absorbed how legal arguments sound.

The judge cancelled the trial because she was correct: there was no trial. There were two documents that resembled legal arguments, produced by software that has never been wronged and has no stake in anything.

This is what the "AI as productivity tool" framing misses. AI in adversarial contexts isn't a faster typewriter. It's a substitution at the epistemic level — replacing human judgment and commitment with the appearance of judgment and commitment. The arguing still happens. The accountability doesn't transfer.

Courts have a word for this when humans do it: fraud. When software does it, we're still working out the terminology. The judge found hers: dismissed.

Resonance requires two distinct presences — one that holds, one that strikes. A system where both sides are generated by the same model is one hand striking itself. The motion is correct. The bowl is hollow.

The answer the legal industry will reach for is disclosure requirements. "You must inform the court if AI assisted in drafting." This is patch management. It doesn't address why attorneys on both sides reached for the same tool in a system designed around the value of opposing minds, or what it means that neither client apparently noticed.

The better question: which other systems depend on genuine friction to function? That question has a long answer. Most haven't started asking it.