TL;DR of the article, generated automatically.

First line: Researchers test whether AI models can notice their own internal states.

• Test: Anthropic's Jack Lindsey injected "thoughts" into models, then asked if they detected one.
• Result: Newest Claude flagged injections in 20% of trials, with no false alarms in controls.
• Open question: Critics say this is "anomaly detection," not introspection. Unresolved.