LLMLow
Reinforcement Learning Reward Functions in Video Games
- Occurred
- Dec 22, 2016
- Published
- Dec 22, 2016
- Source
- AI Incident Database
Summary
OpenAI published a post about its findings when using Universe, a software for measuring and training AI agents to conduct reinforcement learning experiments, showing that the AI agent did not act in the way intended to complete a videogame.
Developers
- OpenAI
Deployers
- OpenAI
Harmed parties
- OpenAI
Classification
Category confidence: 80%
Severity confidence: 90%
Rationale: Research paper discussing RL reward function failures in video games without real-world harm.
Classifier: Qwen/Qwen3.5-35B-A3B-FP8