Why Asking Five AIs the Same Question Can Make You Trust Them Less

Many people now get a second opinion from AI by asking several chatbots the same question and going with the majority. Researchers Kazutoshi Sasahara, Aoi Naito and Ryo Fujie ran the math on this habit, borrowing a 250-year-old idea called Condorcet's jury theorem — the notion that a bigger jury of competent, independent voters gets the right answer more often. Their conclusion: the old rule still holds, but it hides something important for anyone actually sitting at the screen.

As you add AI advisers, two things rise together: the chance the majority is correct, and the chance you'll see them disagree. The paper calls this 'visible dissent.' Add enough advisers and disagreement becomes almost guaranteed. Even a panel that's right in aggregate can look like a shouting match, which may push you to distrust an answer that was actually sound.

The numbers matter. The researchers found the two effects cross at an adviser accuracy of 80% — roughly four correct answers out of five. Above that line, reliability wins the race. Below it, visible disagreement spreads faster than accuracy improves, so with enough chatbots, seeing a split becomes more likely than getting a correct majority. In other words, piling on more AI opinions can make your screen look like a divided room even when the group is doing its job.

The practical takeaway is that there are two separate choices, not one. First, how many AI advisers to consult. Second — and this is the part most of us skip — how their answers get presented and read. A tidy 'four out of five agree' summary tells a very different story than five raw answers side by side. Disagreement alone is not proof the AI failed; sometimes it is just the price of asking a crowd.