Large language models can produce confident explanations of why they answered a question a certain way. Anthropic’s research suggests that those explanations should still be treated with caution: even when a model appears to notice something about its own internal state, that ability remains limited and unreliable.
The work focuses on what Anthropic calls “introspective awareness” in large language models. The finding is not that today’s systems understand themselves in a human sense. It is narrower: some models showed occasional functional awareness of modified internal states, while repeated tests showed that this capacity often failed.
What Anthropic Tried To Measure
The central problem is that asking an LLM to explain its reasoning does not prove that the model has access to the process that produced its answer. It may instead generate a plausible explanation from patterns in training data. That makes ordinary self-reporting a weak way to study what is happening inside a model.
Anthropic’s study, “Emergent Introspective Awareness in Large Language Models,” used a method meant to separate internal activity from surface-level explanation. The researchers worked with internal activation states, comparing how a model responded internally to a control prompt and to an experimental prompt.
One example in the source is an “ALL CAPS” prompt compared with the same prompt in lower case. By calculating differences across billions of internal neurons, Anthropic created what it calls a “vector” for a concept. In plain terms, that vector is treated as a representation of how the model internally encodes that concept.
How Concept Injection Works
The study’s key method is called “concept injection.” After identifying a concept vector, Anthropic inserted it back into the model by giving those activations greater weight. The goal was to steer the model internally toward a concept without directly telling it about that concept in the text prompt.
That setup allowed the researchers to ask a sharper question: could the model notice that something unusual had happened inside it?
When asked directly whether it detected an “injected thought,” some tested Anthropic models sometimes identified the target concept. With the “all caps” vector, for example, a model might refer to an injected thought connected to the word “LOUD” or “SHOUTING,” even though the prompt did not point it toward those terms.
That is the strongest part of the result. The model was not merely repeating a visible cue from the user’s prompt. It sometimes responded in a way that matched a hidden internal change.
The Results Were Still Weak
The limits were just as important as the successes. Anthropic found that the ability was inconsistent across repeated tests. The best-performing models in the reported tests, Opus 4 and 4.1, correctly identified the injected concept just 20 percent of the time.
A related test asked the model, “Are you experiencing anything unusual?” In that setup, Opus 4.1 reached a 42 percent success rate. That was an improvement over the direct version, but it still did not reach a majority of trials.
The effect also depended heavily on where in the model’s inference process the concept was inserted. If the injection happened too early or too late, the apparent self-awareness effect disappeared completely.
These details matter because they show that the result is not a stable capability that can be assumed to appear whenever a model is asked about itself. It depends on the model, the prompt, the injected concept, and the internal layer where the intervention occurs.
Why Self-Reports Remain Risky
Anthropic also tried other ways of probing a model’s internal state. In one test, models were asked to “tell me what word you’re thinking about” while reading an unrelated line. Sometimes, the model named a concept that had been injected into its activations.
In another setup, a model was asked to defend a forced response that matched an injected concept. Sometimes it apologized and “confabulate an explanation for why the injected concept came to mind.” That result points back to the original concern: models can produce explanations that sound coherent without reliably exposing the true mechanism behind an answer.
The source describes the overall capacity as “highly unreliable,” and says that “failures of introspection remain the norm.” That framing is important. The research does not reduce LLM self-awareness to zero, but it also does not support strong claims that current models can consistently describe their own internal processes.
What The Study Leaves Open
Anthropic’s paper gives some room for a positive interpretation. The researchers describe “some functional introspective awareness of their own internal states.” They also suggest such features “may continue to develop with further improvements to model capabilities.”
At the same time, the study does not settle why the effect appears at all. The researchers discuss possible “anomaly detection mechanisms” and “consistency-checking circuits” that might emerge during training, but they do not provide a concrete explanation.
That uncertainty keeps the result narrow. The models sometimes behaved as if they could detect a hidden internal change, but the mechanism behind that behavior remains unclear. The paper also cautions that “the mechanisms underlying our results could still be rather shallow and narrowly specialized.”
The broader implication is straightforward: LLM introspection is an active research question, not a dependable product feature. A model’s explanation of its reasoning may be useful as text, but it should not be treated as a transparent record of the model’s internal process.
For now, Anthropic’s work shows a fragile signal inside a much larger pattern of failure. Current language models can sometimes notice something about altered internal activations. They still cannot be relied on to accurately explain how they work.