Do AI Models Know Their Own Thoughts: An Experimental Test

🤖 Do AI Models Know Their Own Thoughts

Researchers injected concepts into LLM activations (steering) and asked whether the model noticed the injection. Top-tier Claude versions (Jack Lindsey, Anthropic) detected it in ~20% of trials with 0% false positives. However, the model could not name the implanted concept: this ability can be disabled without affecting detection — the descriptions proved to be an artifact of steering, not introspection.

🌍 If a model detects anomalies in its own activations, this serves as an interpretability tool: monitoring manipulations and debugging, though LLM self-awareness metrics are not standardized.

👤 "Does AI think about itself" has been tested with numbers: 20% detection, 0% false positives, +53% after ablation refusals. The Quanta article and three arXiv papers can be read to see the boundary of what is proven.

Source 1: https://www.quantamagazine.org/new-research-tests-whether-ai-models-know-their-own-thoughts-20261104/ Source 2: https://arxiv.org/abs/2508.14802