AI learns to negotiate like humans by playing against itself thousands of times

Researchers built an AI system that learns opponent strategies by playing repeated games against itself, using a combination of game theory and deep learning to improve its ability to predict what others will do. In tests where humans negotiated with the AI in a bargaining game, the AI achieved deal outcomes comparable to human-to-human negotiations — suggesting the method works for teaching machines to model and respond to other agents' behavior.

Most AI systems that play against opponents require hand-coded strategies or heavy domain expertise to set up. This approach automates that process entirely: the system builds its own opponent model through self-play and theory-based training, then refines it while actually playing. The practical implication is that AI agents could now handle multi-agent negotiation and competitive scenarios with less human design work — which matters for things like auction design, diplomacy simulation, or any situation where you need an AI to predict what another actor will do and respond adaptively.