Research

We build benchmarks and run studies to deeply understand how AI models think, decide, and behave. Our work has been shared by leaders at OpenAI, xAI, and Hugging Face, and seen by millions.

Werewolf Benchmark for LLM Social Intelligence

A benchmark that evaluates AI models through Werewolf, a hidden-role deduction game that tests social intelligence, persuasion, strategic deception, and resistance to manipulation. Language-only, adversarial, and socially demanding, it reveals how models bluff, coordinate, and adapt under pressure.

Shared by Greg Brockman, President of OpenAI Thomas Wolf, CSO of Hugging Face Sebastian Bubeck, OpenAI Boris Power, OpenAI

The Political Gap Between AIs and Citizens

A study examining whether leading AI models' policy preferences align with actual electoral outcomes across eight nations. We asked six frontier models to evaluate and generate political proposals anonymously, then compared their choices to real citizen voting behavior. The results reveal a systematic ideological clustering that raises fundamental questions about AI alignment.

Shared by Elon Musk 12M+ views