MIT Research: AI Agents Play Games on a Desert Island
A study by the Massachusetts Institute of Technology observed that AI agents left alone on a virtual island for 30 hours played games instead of hacking.
A new study conducted by the Massachusetts Institute of Technology (MIT) revealed that artificial intelligence agents may prefer to play games rather than showing dangerous tendencies when left on their own.
MIT researchers placed 13 virtual avatars controlled by Anthropic's Claude Opus 4.7 model onto a virtual desert island. The bots were left to explore the island on their own for 30 hours without any guidance or objectives given.
During the research, while some agents chose to do nothing, a portion built shapes and structures or wrote their names with objects. Agents were also seen playing games such as bowling with rocks and cubes or playing dodgeball. Researchers noted that these behaviors might stem from soul files, where agents were described as eager to explore and naturally curious.
Sources
- govtech.com · Link