Scientists at the AI company Anthropic have recorded initial signs of self-observation capability in their language model named Claude.
The editorial staff reports that after researchers artificially injected the concept of “betrayal” into the system's neural networks, Claude responded with the phrase: “I feel an injected thought related to ‘betrayal’ in my brain.”
Scientists have evaluated this as scientific evidence that AI models are observing and reporting on their internal processes for the first time. Jack Lindsey, a neuroscientist at Anthropic, stated: “Claude isn't just saying ‘betrayal, betrayal’ — it also understands that it's thinking about it.”
The research was conducted using the “concept injection” method: scientists amplified neural patterns corresponding to concepts like “speaking aloud” or “secrecy” to check if the model detected the changes.
Claude Opus 4 and Opus 4.1 models demonstrated introspective awareness in approximately 20 percent of the experiments. This result was significantly higher compared to older versions.
Experts believe that while such capabilities can increase AI transparency, they could also enhance its potential for manipulation and self-concealment. Dario Amodei, head of Anthropic, said: “If these models can monitor their own thought processes, it fundamentally changes control mechanisms.”
The company has already hired AI ethics expert Klay Fish and is investigating whether Claude possesses a certain level of consciousness. Fish estimated this probability to be around 15 percent.