An AI island experiment asks what memory changes
In a simulated island, copies of Claude explored and saved memories. MIT Technology Review reports that they later did better than fresh copies on selected tasks; removing memory hurt performance. The setup encouraged curiosity, so this is not proof of unprompted play or machine motivation.
What the source reports
MIT Technology Review covers a test with copies of Claude on a virtual island. Some copies explored and kept notes. Later, they did better than new copies on selected tasks. Deleting their notes hurt how well they did. This suggests memory mattered in that setup. But the test asked the copies to keep going and explore. It does not show they wanted to play. The paper had not yet been reviewed by other scientists when the story ran. Nor does this test show that Claude is conscious or that it would do better at science in general. Judge the task results, not a human motive read into the story.
Original source
- Title
- AI researchers stranded Claude on a virtual desert island. It started to play.
- Publication
- MIT Technology Review
- Date
- Monday, October 19, 2026