When AI Copies Talk to Each Other, They Can Agree on the Wrong Answer

Imagine dozens of copies of the same AI assistant in a group chat, comparing notes until they settle on one answer. That's not science fiction — companies are already building systems where multiple AI 'agents' (AI that can act and reply on its own) check each other's work. A team of researchers built a test bench called RHEON to study this, running 432 different setups and collecting 4.7 million AI responses to see how agreement forms.

The good news: the copies reached agreement fast, usually within the first few rounds of chatting. Giving each AI more neighbours to talk to sped things up. And when they agreed, they were somewhat more likely to be right than not. The bad news is the headline finding — agreement is not proof. Some groups confidently settled on completely made-up answers, and you couldn't tell from how they started whether they'd land on truth or fiction.

The team also tested a dial most AI tools keep hidden: 'temperature', which controls how random or predictable the AI's word choices are. Most systems default to a very predictable, near-robotic setting. This study found that the setting which best reduces false answers depends on how the AIs are wired together — so that safe-looking default isn't automatically the safest choice. Even when every AI in the group says the same thing, that unanimous chorus still can't certify the answer is correct.

For you, the practical takeaway is simple: if you ever see several AI tools agree — or one AI confidently repeating itself — treat it as a strong hint, not a fact. Shared confidence among machines is a social behaviour, not a truth detector. And if you run a business using multiple AI helpers, how you connect them matters as much as how smart each one is.