New Research Tests Whether AI Models Know Their Own Thoughts

Key Points:

• Researchers are investigating whether large language models (LLMs) can perform introspection—awareness of their own internal thought processes—by testing their ability to detect and describe artificially injected "thoughts" into their neural networks.
• In 2025, Jack Lindsey at Anthropic developed a method to inject specific neural activation patterns into LLMs and found that models could sometimes identify and describe these injections, suggesting a form of introspection or at least anomaly detection.
• Follow-up studies by other researchers revealed that while LLMs can detect unusual internal activity (anomaly detection), their detailed descriptions of injected concepts may result from prompt steering rather than true introspective awareness.
• The debate continues over whether anomaly detection in LLMs qualifies as introspection, with some experts noting the difficulty in applying human-centric definitions of introspection to disembodied AI systems lacking subjective experience.
• This research is advancing philosophical and scientific understanding of introspection both in AI and humans, highlighting challenges in defining and measuring self-awareness and internal representation in complex systems.