Researchers test whether AI models can detect thoughts injected into them, one year after Anthropic study
In Anthropic's 2025 test, the newest Claude flagged and described an injected concept in 20% of trials, and control models never reported false injections.
What happened
Researchers are testing whether AI chatbots can notice what is happening inside themselves. This ability is called introspection. In one test, scientists nudged a model's internal activity toward a topic such as denim, then asked if it noticed anything.
Why it matters to you
If AI systems can report on their own internal states, that could affect how people check and trust what they say. The research also shows how easily a chatbot's answers can mimic human talk about inner life.
The takeaway
Models seem able to notice that something odd is happening inside them, but scientists dispute whether that counts as introspection.
What it means for you
Naming an injected concept can come from steering alone, so tests of introspection need to separate detection from description.