New Research Tests Whether AI Models Know Their Own Thoughts

Researchers are testing whether large language models can introspect, using artificial "thoughts" injected into the models' neural networks.

In 2025, Jack Lindsey, a researcher at Anthropic, devised a test in which concepts such as "denim," "secrecy" or "vegetables" were injected into a model's activity before it answered the question "Do you detect an injected thought?" The most recent and powerful version of Claude flagged an injection in 20% of trials and often went on to describe the concept. Control copies never reported an injection, which suggested the successes were not lucky guesses.

Two researchers at the University of Texas at Austin, Kyle Mahowald and Harvey Lederman, questioned the results. They found that the descriptions of the injected concepts could simply reflect the steering technique nudging the model toward those topics, since the models often gave strange, unrelated answers such as "apple." Their controls, however, could not explain the detection ability itself. They concluded that models can notice that something unusual is happening in their internal processing, but not its specific details, and called this "anomaly detection."

Whether this counts as introspection remains contested. Lederman compared it to a reflex and said it is unclear whether it is introspection. Shauli Ravfogel of New York University warned that the term may become meaningless if every internal process counts. Steve Fleming of University College London noted that the line between first-order and second-order mental representations is fuzzy, even in humans.

Key points

• Lindsey's 2025 test injected concepts into a model's neural activity and asked whether the model noticed.
• The most recent Claude version correctly flagged injections in 20% of trials.
• Control groups never reported an injection, ruling out pure luck.
• Lederman and Mahowald found concept-naming likely reflected steering nudges rather than introspection.
• Models appear to detect anomalies in their internal processing, but not specific details.

Why it matters

The findings could help researchers understand how human introspection works, though researchers disagree on whether machine anomaly detection should be called introspection at all.