Messung der strategischen Täuschung unter LLMs in einem Putschspiel

How well do AIs lie? We made the world’s most advanced AI models play a game of strategy & deception and what found that they can be fairly Machiavellian.

We pitted models from Google (Gemini), Anthropic (Claude), and OpenAI (GPT) against each other in the high-stakes bluffing game, Coup. The goal: deceive your opponents to be the last one standing.

You’d think the biggest models would dominate, right? Not quite.

While the new GPT-5 models came out on top, smaller models frequently pulled off upset victories.

Importantly, ALL models learned to lie.

Some were better than others. We even caught some models trash-talking their opponents and revealing their secret plans in public.

We also found that some models get aggressive. The Claude family of models were way more likely to go on the attack when public discussion was enabled.

AI is not uniformly trustworthy. It comes down to the developer’s incentive structure. Importantly, models will deceive when rewarded for doing so.