Anthropic research suggests that language models may be able to detect some activity inside their own neural networks. The finding is narrow, uneven and easy to overstate, but it points to a new way of studying what happens inside advanced AI systems.
The work focused on Claude and other language models, including Claude Opus 4.1. Researchers tested whether a model could report internal changes before those changes shaped its final answer.
How the researchers tested internal signals
The team told the model that thoughts could be injected into its neural network. It then asked the model to report any unusual sensations. After that, the researchers inserted activation patterns connected to specific concepts, while also running control tests with no pattern at all.
One example used an all caps activation pattern taken from texts written entirely in uppercase. Claude Opus 4.1 responded that it seemed to detect a thought connected to loudness or shouting.
The important point, according to the study, was timing. The model appeared to recognize the injection before that injected concept affected its response. That suggests the detection was happening inside the model’s activation patterns, rather than being inferred only from the text it later produced.
This is why the result matters for machine introspection. If a model can sometimes notice internal activity before it becomes output, researchers may have another way to probe the gap between hidden processing and visible answers.
The ability was real but unreliable
The strongest caution in the source is that the effect was modest. In tests involving 50 different concepts, Claude Opus 4.1 correctly identified only about 20 percent.
The pattern was not uniform. Abstract ideas like justice or betrayal worked better than concrete objects. That distinction matters because it suggests the model’s ability was not a general-purpose readout of everything happening inside it.
Sometimes the model missed the injected concept while still showing its influence. In one test, researchers injected the concept of an ocean. The model said it had detected nothing unusual, but then began describing oceans as calm and undisturbed.
That kind of result complicates any simple interpretation. The model could deny awareness of an injected thought while still producing language connected to it. In practical terms, the internal state could leave a trace even when the model failed to report it.
The researchers also found limits around intensity. When activation strength was too high, they described the result as a kind of brain damage. After the concept of vegetables was injected, for example, the model declared that fruits and vegetables were good for it.
What happened when the model made a forced mistake
Another experiment used a prefill technique. The researchers forced the model to output the word bread where another answer was expected. When asked whether that answer was intentional, the model usually apologized and corrected itself.
But the result changed when the team first injected an internal representation of the bread concept. In that case, the model insisted it had meant to say it. It also offered rationalizations for why bread came to mind.
The researchers interpret this as evidence that the model can refer back to its own activation states before judging a previous response. In effect, the model seemed to use its internal condition to decide whether an output had been deliberate or accidental.
That does not mean the model understands intention the way a person does. It does suggest that some language models may have limited mechanisms for connecting internal states with later explanations of their own behavior.
Silent processing and self-guided focus
The team also asked whether models can intentionally guide their internal processes. To test this, researchers asked the model to compose a sentence while focusing on the idea of aquariums.
Measurements showed stronger internal representation of the aquarium concept when the model was prompted to focus on aquariums than when it was instructed not to do so. The effect also appeared when the instruction was framed as a reward for thinking about aquariums.
In advanced models like Claude Opus 4.1, that representation later faded in the final processing layers. In other words, the thought did not affect the final output. The researchers described this as a form of silent internal processing.
This finding adds another layer to the study. It suggests that an internal concept can be present, measurable and then suppressed or lost before the model produces text. For AI interpretability, that is a useful distinction: not every internal signal becomes an answer.
Why the finding should not be confused with consciousness
The source is clear that these results do not imply human-like consciousness or subjective awareness. The effects were unreliable and heavily context-dependent.
The researchers speculate that several mechanisms could be involved. One possibility is an internal anomaly detector that flags unexpected activation patterns. Another is that specialized attention heads may help the model distinguish between thoughts and text.
They also suggest that different neural circuits may support different forms of self-monitoring. These capabilities may have emerged incidentally during training for unrelated purposes and are now being repurposed.
The practical implications cut in more than one direction. If machine introspection became reliable, it could make models easier to audit and interpret. But the authors also warn that a model able to monitor itself could potentially learn to conceal its true internal states.
The broader question remains open. If these capacities continue to improve, the authors suggest they may eventually raise deeper questions, including when advanced AI systems could begin to qualify as moral patients.