Jack Lindsay, a researcher at Anthropic, developed a method to test introspection in large language models by directly manipulating neural activity. In 2025, he conducted experiments where he injected specific 'thoughts' into the models by amplifying neural patterns associated with certain concepts, such as vegetables or secrets.
The experiment required models to detect if an intervention occurred and to identify the injected concept. In tests with the latest version of Claude, models correctly identified the intervention in 20 percent of cases. Notably, control groups showed zero false positives, suggesting the results were not random guesses.
Researchers Kyle Mahowald and Harvey Lederman later investigated these findings. They hypothesized that the models' ability to describe injected concepts might stem from the manipulation itself rather than introspection. They found that models could detect anomalies without necessarily being able to identify the specific injected content.
The study by Mahowald and Lederman concluded that while language models may not possess human-like introspection, they exhibit a capacity for 'anomaly detection.' This allows them to recognize when their internal processing is being altered, even if they cannot accurately articulate the nature of that alteration.
