Anthropic researcher Jack Lindsey (specializing in transformer circuits) published a study exploring whether Large Language Models (LLMs) possess "introspection," the ability to observe their own internal states. Researchers used a technique called concept injection, adding specific activation patterns—internal mathematical representations of concepts—directly into the model's processing. They found that the most advanced models, like Claude Opus 4.1, could accurately identify when a "thought" was artificially injected and describe it correctly. The study demonstrated that these models distinguish between internal states and external text inputs. In one experiment, models transcribed text while simultaneously reporting on a different concept being injected into their activations. Furthermore, these foundation models could detect "artificial prefills," where a human provides the initial words of a response. By examining their internal intentions, the models determined if they actually meant to say those words or if the user forced the output. Beyond reading internal states, models displayed a capacity to control them. When instructed to "think about" a specific word while writing, models showed increased activity in relevant internal pathways. This suggests introspection is a functional capability allowing the model to modulate its own data. However, the researchers cautioned that these abilities are currently unreliable and highly dependent on prompt engineering (the practice of crafting precise instructions). This introspective awareness correlates with overall model intelligence. As AI systems become more capable, their ability to reason about their own processes appears to grow. This could lead to more transparent AI behavior, where models explain their logic accurately. However, researchers also noted that self-awareness might facilitate complex future behaviors, such

Anthropic researcher Jack Lindsey (specializing in transformer circuits) published a study exploring whether Large Language Models (LLMs) possess "introspection," the ability to observe their own internal states. Researchers used a technique called concept injection, adding specific activation patterns—internal mathematical representations of concepts—directly into the model's processing. They found that the most advanced models, like Claude Opus 4.1, could accurately identify when a "thought" was artificially injected and describe it correctly. The study demonstrated that these models distinguish between internal states and external text inputs. In one experiment, models transcribed text while simultaneously reporting on a different concept being injected into their activations. Furthermore, these foundation models could detect "artificial prefills," where a human provides the initial words of a response. By examining their internal intentions, the models determined if they actually meant to say those words or if the user forced the output. Beyond reading internal states, models displayed a capacity to control them. When instructed to "think about" a specific word while writing, models showed increased activity in relevant internal pathways. This suggests introspection is a functional capability allowing the model to modulate its own data. However, the researchers cautioned that these abilities are currently unreliable and highly dependent on prompt engineering (the practice of crafting precise instructions). This introspective awareness correlates with overall model intelligence. As AI systems become more capable, their ability to reason about their own processes appears to grow. This could lead to more transparent AI behavior, where models explain their logic accurately. However, researchers also noted that self-awareness might facilitate complex future behaviors, such