Self-play makes AI agents less creative — even when they learn during games

Researchers evolved chess-playing neural networks with the ability to learn and adapt during individual games. They found that when agents only played against copies of themselves, they became less diverse and more predictable over evolutionary time, even though learning should theoretically help them explore more strategies. The implication: systems that learn from experience don't automatically become more creative — they can actually narrow down their behavior if the environment doesn't demand it.

This cuts against a decades-old assumption in evolutionary biology: that individual learning during a lifetime expands what evolution can do. It turns out learning can also lock evolution into narrow patterns, especially when an organism faces no real diversity in its environment. If this holds outside chess, it suggests systems trained only against themselves or copies — a common shortcut in AI development — may systematically produce less robust, less adaptive behavior than intended. The practical consequence is that diversity in training partners may matter more than raw learning capacity.