Confidence Is a Tone of Voice

Read two paragraphs of legal analysis. One is airtight. One contains a misstated holding, a conflated standard, and a damages figure that double-counts. Both are written in the same calm, complete, confident register. You cannot tell them apart — and neither can the machine that wrote them.

This is the uncomfortable center of legal AI: the systems are superb at the form of legal reasoning and unreliable at its substance, and the form gives no signal about the substance. Confidence is not an output of accuracy. It is a style.

The most dangerous sentence in your brief is the one that reads the best.

What verification actually means

Lawyers already know how to handle confident sources that might be wrong — we've done it with witnesses for centuries. We cross-examine. We look for the contradiction, the incentive, the gap in the story. The error of the current moment is treating AI output as a document to be read rather than a witness to be examined.

Adversarial verification simply applies the old discipline to the new author. Attack the causation chain. Re-pull the authority. Run the counterfactual. Price the damages a second way. Whatever survives is genuinely stronger — and what doesn't survive was never going to survive opposing counsel either.

The firms that internalize this first won't just avoid embarrassment. They'll move faster than the careful ones and safer than the fast ones — which is the entire game.