Enterprise AI Analysis: Signs of introspection in large language models

We found that models, particularly Claude Opus 4 and 4.1, can detect and accurately identify concepts artificially "injected" into their activations. This demonstrates an immediate, internal recognition of novel internal states, providing direct evidence of functional introspective awareness.

While often unreliable, this capability suggests an anomaly detection mechanism that senses deviations in internal processing, a feature critical for future transparent and self-aware AI systems.

A key aspect of introspection is the ability to differentiate internal "thoughts" from external sensory inputs. Our experiments show models can simultaneously transcribe external text while reporting on an unrelated injected internal concept.

This suggests models maintain a functional distinction between processing raw inputs and accessing their own internal representations, a foundational element for sophisticated self-awareness.

Introspective AI can naturally detect and disavow unintended or "prefilled" outputs by checking internal consistency between prior intentions and produced text. By injecting concepts, we can manipulate this mechanism, demonstrating that models refer to their prior internal "intentions" to determine responsibility for outputs.

This capability is crucial for robust alignment, helping models resist jailbreaking and maintain coherent character, ultimately enhancing system reliability.

Our research explores whether models can intentionally modulate their internal representations. We found that models can strengthen or weaken the internal presence of a concept when explicitly instructed or incentivized to "think about" or "not think about" it, even for unrelated tasks.

This suggests a rudimentary form of metacognitive control, allowing models to regulate their internal focus. In advanced models like Opus 4.1, this modulation can occur "silently," without influencing the final output, indicating sophisticated internal state management.