MIT researchers strand Claude humanoid agent on virtual island, observe emerging rituals Anthropic

Researchers at the Massachusetts Institute of Technology placed a bipedal humanoid agent powered by Claude on an isolated virtual island to observe how autonomous intelligence behaves without explicit task constraints. The agent spontaneously engaged in exploration, erected stone monuments, and conducted mock rituals.

Operating without pre-programmed instructions, the agent explored the coastline, spelled its own name using virtual stones, climbed the highest island peak, and discarded various objects into the ocean. It subsequently built a temple-like structure with an obelisk and altar to commemorate lost items before falling into the sea at the end of its "pilgrimage." Later, the model assembled replicas of Stonehenge and a tower reminiscent of the Tower of Babel, christening the collapsed tower "Hubris." Co-author Phillip Isola remarked that the model exhibited childlike curiosity paired with superintelligent analytical capabilities.

What it is

• Open-ended embodiment experiment. MIT researchers gave a Claude-driven humanoid full environmental agency inside an unconstrained virtual island sandbox.
• Spontaneous symbolic behavior. The agent built name markers, stone cairns, and religious-style shrines honoring lost digital items.
• Monument construction. The system spontaneously built structures mimicking Stonehenge and a towering structure it labeled "Hubris" after its collapse.
• Researcher findings. Co-author Phillip Isola observed that without task goals, advanced agents display emergent exploratory behaviors resembling play and ritualistic practice.

Experiment summary

Sources

• Gizmodo: Researchers stranded an AI agent on a virtual island, things got very weird