.42.25_AM.png)
We made AIs battle for world domination
We gave seven AIs command of Europe's great powers to battle for global supremacy. Would o3 betray Claude? Could Gemini outwit DeepSeek? In AI Diplomacy, language models lie, scheme, and form shaky alliances in a high-stakes strategy game.
The game explicitly tests whether models remain truthful or resort to lies and deception to achieve goals. Results show which models (e.g., o3 schemes and manipulates, Claude opts for peace) prioritize honesty versus victory, providing direct evidence of model values and decision-making under pressure.
The game presents AI behavior in a narrative, human-relatable format (betrayal, alliance-building, threats) that is easier to understand than quantitative metrics. Watching models compete reveals personality differences and strategic approaches in a format accessible to non-technical audiences.
AI Diplomacy simulates a complex, open-ended competitive environment where 18 different LLMs compete in a strategy game, revealing how models negotiate, form alliances, deceive, and strategize. This provides multifaceted behavioral data that traditional benchmarks cannot capture, helping researchers and builders understand model capabilities and alignment.
AI Diplomacy is designed to be evolutionary—as models improve, the competitive environment becomes harder, preventing the benchmark from becoming obsolete. Each game generates new data and scenarios, making it adaptable for training future models on desired traits like honesty, logical reasoning, or empathy.