Can Claude notice its own thoughts? The introspection study, explained
Claude can sometimes notice what is happening inside it, but only sometimes. In a study published Oct. 29, 2025, Anthropic researchers planted concepts directly into Claude's internal activity and asked whether it noticed; Claude Opus 4.1, the best performer, detected the planted concept about 20% of the time, according to Signs of introspection in large language models. The authors call the ability "highly unreliable and limited in scope."
In one sentence
Some Claude models can occasionally detect and describe changes researchers make to their internal states, which is a limited, functional form of introspection.
How they measured it
The researchers used a technique they call concept injection. They first recorded the pattern of internal activity that corresponds to a known concept. Then, while the model worked on an unrelated task, they inserted that pattern and asked the model whether it noticed anything unusual. Because the researchers knew exactly what they had put in, they could check whether the model's report was accurate. They tested several Claude models; Claude Opus 4 and Claude Opus 4.1 did best. The full paper is published as Emergent introspective awareness in large language models.
What they found
Detection works, but rarely. Even with the best method, Claude Opus 4.1 noticed the injected concept about 20% of the time. Most attempts failed.
Claude's introspection in Anthropic's 2025 study
When it works, the model names the concept. After researchers injected a pattern for text in all capital letters, the model reported something unexpected in its processing and linked it to loudness or shouting.
It can be fooled about its own intentions. When researchers made the model say the word "bread" in an unrelated context, it usually called the word a mistake. When they had also injected "bread" into its earlier internal activity, it accepted the word as intended and produced a reason for saying it.
It has some control over its internal states. Told to think about aquariums, the model showed more aquarium-related internal activity than when told not to, and both were above baseline.
What it doesn't say
The study does not say Claude is conscious. "Our results don't tell us whether Claude (or any other AI system) might be conscious," the authors wrote. It measures functional abilities, accessing and reporting internal states, not experience. It also does not show that models introspect the way people do, or that a model's description of itself can generally be trusted: most of the time, it could not tell.
Context
The work builds on Anthropic's earlier interpretability research, such as the March 2025 study that traced how Claude plans rhymes.
FAQ
Can AI models introspect?
In a limited way. In Anthropic's 2025 study, Claude Opus 4.1 detected concepts planted in its internal activity about 20% of the time; most attempts failed.
What is concept injection?
A technique in which researchers record the internal pattern for a concept and insert it into a model during an unrelated task, then ask whether the model notices.
Does the introspection study show Claude is conscious?
No. The authors say the results don't tell us whether Claude or any other AI system might be conscious.
Sources
Cite this
Can Claude notice its own thoughts? The introspection study, explained, New Mind Museum, November 6, 2026. https://newmindmuseum.com/interpretability/2026/11/claude-introspection-study-anthropic-explained/
New Mind Museum. (2026, November 6). Can Claude notice its own thoughts? The introspection study, explained. https://newmindmuseum.com/interpretability/2026/11/claude-introspection-study-anthropic-explained/
“Can Claude notice its own thoughts? The introspection study, explained.” New Mind Museum, 6 Nov. 2026, https://newmindmuseum.com/interpretability/2026/11/claude-introspection-study-anthropic-explained/.
@misc{new-mind-museum-2026-can-claude-notice-its-own-thoughts-the-i,
author = {{New Mind Museum}},
title = {Can Claude notice its own thoughts? The introspection study, explained},
year = {2026},
url = {https://newmindmuseum.com/interpretability/2026/11/claude-introspection-study-anthropic-explained/}
}Free to quote and cite with a link to this page.