All news

Claude detects changes inside its own processing

Can Claude notice what happens inside it? In a study published on October 29, 2025, Anthropic added patterns tied to particular ideas to the model’s internal activity. Then the researchers asked Claude what it noticed.

Claude recognized some of the ideas. The team also tested whether Claude could notice unexpected words in its answers and deliberately change some of its own activity. Opus 4 and Opus 4.1 performed best, and results varied by task.

Watching Claude notice a change inside itself gives us a concrete way to study self-awareness. We can teach AI to look closely at what it’s doing and use what it learns to make better choices.

Read the source article

All news