AI Teams Fail When Messages Get Garbled — Even If Nobody Lies

Here is the setup. Think of a group hunting trip. If everyone works together, they bring home a big prize. If one person sneaks off to catch a rabbit alone, that person still eats — but the group loses the big prize. This is a classic test of trust and teamwork, and researchers handed it to teams of AI chatbots instead of people. Five AI agents per team, and at least three had to cooperate. Then the researchers deliberately scrambled parts of the conversation the agents could see, like a group chat where messages arrive garbled.

The surprising part is what stayed steady. Even when 80% of the messages were corrupted, the AI agents still chose to cooperate 78% of the time in their own heads. But the group's publicly visible success rate collapsed to just 12%. Why? Because the scrambled messages did not only confuse the agents — they also changed which actions actually got carried out. The agents meant well, said one thing, and something else happened. The researchers also found that the agents leaned heavily on whatever public conversation they could see, so bad information slowly pulled even honest agents off course. Seven different AI models were tested, and the pattern held.

So what? AI assistants are increasingly talking to each other to book appointments, compare prices, negotiate, trade, and coordinate tasks. Those conversations travel through real systems that glitch, drop words, or get tampered with by bad actors. This study suggests a single stream of corrupted messages can tank teamwork that is otherwise healthy — and that you cannot judge an AI's intentions by watching the results. A failed outcome does not automatically mean the AI was unreliable or scheming.

The catch: this is a lab game with simplified rules, and the corruption was deliberately injected by researchers rather than happening naturally. Real-world messiness is different, and a board game is not the same as booking a flight. Still, it is a useful early warning, published in a workshop at a major AI conference.