Competition may reverse AI homogenization, new game theory paper finds
Using LLMs to play Scattergories reveals that competition drives diversity in generative AI.
Generative AI models have been shown to homogenize content—users produce more similar outputs when relying on the same models. But a new paper by Manish Raghavan (arXiv, Dec 2024, revised June 2026) argues that competition pushes in the opposite direction. When producers compete for customers or attention, they are incentivized to create novel or unique content, which may counteract AI-driven monoculture.
Raghavan uses a formal game-theoretic model to show that competitive markets select for diverse AI models. Crucially, a model that performs well on standard benchmarks may fail to provide value in a competitive environment because it encourages homogeneity among users. The paper validates this empirically by having language models play Scattergories, a game rewarding correct and unique answers. The results suggest homogenization is unlikely to persist in competitive markets, and competition may drive diversification in AI development.
- Generative AI in isolation reduces content diversity, but competition incentivizes uniqueness.
- Game-theoretic model shows competitive markets select for diverse AI models, not just high benchmarks.
- Empirical validation using LLMs playing Scattergories confirms that competition drives output diversity.
Why It Matters
Businesses deploying AI must evaluate output diversity, not just accuracy, to stay competitive in dynamic markets.