Introducing OpenPsy

Large language models are becoming more capable and more widely used. They can perform many tasks previously carried out by people, often at a scale that would be difficult for people to match. Yet there is much we still do not understand about how their behavior changes across tasks and circumstances.

Psychology offers useful tools for investigating these questions. Psychologists have developed methods for studying behavior under controlled conditions and drawing careful inferences from what they observe. We can adapt these methods to study AI models and agents, while keeping open the question of whether similar behavior reflects similar underlying processes.

OpenPsy is being built as a platform for this work: designing behavioral experiments, comparing models across conditions, and keeping an inspectable record of what was tested and what happened. Our aim is to help improve the reliability of AI systems and sharpen our own judgment about their capabilities and limitations.

In the future, OpenPsy may also be used to study human-agent interactions and the cognitive and metacognitive effects of the same.