Should a swarm use majority-rule or unanimity to make decisions? I re-ran Guarnaschelli, McKelvey and Palfrey’s (2000) juror experiment. My jurors were GPT-5.6-terra. With majority rule, jurors essentially vote their signal. If the jury requires unanimity to convict, jurors vote to convict on an innocent signal 48 to 80% of the time. For Caltech students it was 36% and 48%. Under unanimity, a single vote for guilt only counts if all other jurors have voted guilty. And that’s exactly the situation where the innocent signal is likely incorrect, so a strategic juror disregards their own signal. So, unanimity made Caltech students less prone to convict the innocent. But it was worse than majority rule for both types of error for my model jurors: 48% of innocents convicted vs. 22%, when there were only three jurors. When asked what the probability that the defendant is guilty given their own private signal is, my AI jurors always answered exactly 70, or 30. This is Bayes’ rule. All of their strategic behavior was in the vote. You should collect what all the agents found out. And don’t count votes.