Artificial intelligence may be developing a limited ability to notice its own internal activity—but the evidence stops well short of showing human-like self-awareness. New research into AI introspection is testing whether language models can identify thoughts planted inside their neural networks, rather than merely generating convincing explanations after the fact.
The most striking work comes from Anthropic’s interpretability team, which used a technique called concept injection. Researchers first identify a pattern of neural activity associated with an idea, then insert that pattern into a model while it is handling an unrelated prompt. The model is subsequently asked whether anything unusual happened and, if so, what the injected concept was. In some trials, Claude detected the interference before producing any text that would have revealed the concept indirectly. Anthropic’s findings suggest that Claude models possess some ability to monitor and influence their own internal states, although the company emphasizes that the ability is unreliable and narrow.
The numbers provide plenty of room for skepticism. Even Claude Opus 4.1, the strongest model in Anthropic’s tests, correctly detected and identified injected concepts only about 20 percent of the time. The result also depended on the strength of the intervention: weak signals went unnoticed, while stronger ones could trigger confused or hallucinated responses. Anthropic tested multiple Claude generations and found that post-training appeared to matter, while model size alone did not consistently predict better introspective performance.
A second experiment asked whether models could recognize when an earlier response had been forced into place. Researchers prefixed an answer with an unrelated word, then changed the model’s hidden activity to make that word appear intentional. The model’s explanation shifted accordingly, a result consistent with it consulting a representation of its prior “intention.” But that behavior could also reflect specialized circuits responding to unusual patterns, rather than a broad inner observer. Anthropic itself says the experiments do not establish consciousness and may involve several narrow mechanisms instead of one general introspective faculty.
Follow-up research has made the debate sharper. One study found that large models could often detect that an internal anomaly had occurred while failing to identify the anomaly’s specific content, suggesting generic error or novelty detection rather than access to a private thought. Another 2026 paper found that apparent metacognitive performance could sometimes be reproduced from the model’s input alone, challenging the claim that the system had privileged access to its hidden representations. The “Reality Check” analysis concluded that current behavioral tests are not enough to prove genuine metacognitive monitoring.
That distinction matters beyond philosophical arguments about machine consciousness. If models can reliably report what is happening inside them, introspection could become a practical tool for debugging deceptive behavior, tracking hidden goals, or auditing why a system reached a dangerous conclusion. For now, however, the safest interpretation is more modest: some advanced models can occasionally detect and influence aspects of their internal processing, but they do not consistently know what they are thinking—and researchers still lack a decisive test separating introspection from sophisticated anomaly detection.
