AI Models Show Ability to Detect Internal Changes

A team of researchers at Anthropic conducted experiments to determine if large language models could detect when their internal processes were altered. They injected specific neural patterns into models and observed whether the models could recognize these changes.

In one trial, models were asked to identify if they had been injected with a thought. In 20% of cases, the most recent and powerful models correctly flagged the injection and described the concept associated with it.

Control groups, which did not receive injections, never reported detecting any changes. This lack of false positives suggested the successful trials were not random guesses.

Despite these results, researchers remain divided on whether the models' ability to detect internal changes qualifies as introspection. Some argue it may be a form of anomaly detection rather than true self-awareness.

Other researchers tested alternative explanations, finding that models could be influenced by the steering technique used in the experiment. This suggested that the models might not be introspecting but instead responding to external cues.

The debate continues as researchers seek to define what constitutes introspection in both humans and AI. While the models' ability to detect internal changes is notable, whether this represents a form of self-awareness remains unclear.