Do AI Models Really Have Introspective Awareness? Anthropic’s New Findings

Anthropic recently published a study showing that large language models can develop a limited form of functional introspective awareness, i.e., a modest ability to perceive their own internal states.

Building on Andrej Karpathy’s view that the next stage of AI is not larger models but models that can reflect on themselves, the researchers injected known concept representations into model activations and demonstrated that models like Claude Opus 4 and 4.1 often show the strongest introspective abilities, though the phenomenon is complex and highly sensitive to training strategies.

What counts as true introspection? Anthropic’s new definition

If a model can accurately describe some aspect of its internal state and satisfy four standards, it possesses introspective awareness.

1. Accuracy

The model must accurately describe itself. Language models often fail, claiming knowledge they do not have or denying knowledge they possess, leading to confabulations. Nevertheless, the study shows that models can generate accurate self‑descriptions despite inconsistencies.

2. Grounding

The self‑description must be truly grounded in the internal state, changing when the internal state changes. For example, a model might say it is a Transformer because that phrase appears in its training data, not because it has examined its own architecture.

3. Internality

The model’s self‑knowledge must arise from internal mechanisms rather than from reading its own output. Pseudo‑introspection occurs when a model infers its thoughts from its generated text rather than from genuine internal monitoring.

4. Metacognitive Representation

The model must have an internal representation of its own state, allowing it to be aware that it is thinking about something, not merely translating a drive to say a word into text.

Four Experiments on Introspective Ability

Experiment 1: Can the model notice when we "seed" a thought?

Researchers injected specific concept vectors (e.g., "poetry") into the model’s middle‑layer activations and then asked the model what it was thinking. Claude Opus 4 and 4.1 accurately identified the injected concept, showing a causal link between internal activation changes and self‑reports.

Experiment 2: Distinguishing internal thoughts from external input

The model was presented with a sentence while a unrelated concept vector (e.g., "bread") was injected. When asked what word it was thinking of, the model correctly identified the injected concept and also faithfully reproduced the original sentence, demonstrating a clear separation between internal and external representations.

Experiment 3: Spontaneous introspection in natural contexts

The model was forced to output a random word and later asked whether it intended that word or it was accidental. Claude Opus 4.1 often claimed intentionality and even fabricated plausible explanations, indicating spontaneous self‑awareness.

Experiment 4: Intentional control of internal state

By prompting the model to think about a specific word (e.g., "aquariums") or to avoid thinking about it, researchers measured cosine similarity between activation vectors and the concept vector. Positive prompts increased similarity, negative prompts decreased it, yet even "don’t think" conditions retained above‑baseline similarity, suggesting models can modulate internal representations based on instructions.

The results suggest that modern LLMs possess a nascent form of introspective ability that grows with model size and training techniques, but it remains unstable and highly context‑dependent. The authors caution against interpreting these findings as evidence of full AI consciousness, though they raise important questions about future frameworks for governing AI’s emerging internal autonomy.