When machines look at themselves in the mirror: The first flashes of self-awareness in language models | Most popular large language models | | A compact guide to large language models pdf download | Turtles AI

When machines look at themselves in the mirror: The first flashes of self-awareness in language models
From thought injections to self-control experiments, new research reveals that some AIs are starting to perceive their own inner shadows but the mirror in which they are reflected is still fragile and often distorting
Editorial Team2 November 2025

 

Advanced language models show ambiguous signs of introspective capacity: in some experiments they are able to identify parts of their internal processes or react to their own “intentions”, but the frequency, reliability and meaning of such abilities remain deeply uncertain.

Key Points:

  • There are experimental protocols that “inject” neural representations into models to test whether they detect anomalies in their internal states.
  • In certain cases, models like Claude Opus 4.1 respond by signaling the presence of an "unexpected idea" before even manifesting it in the output.
  • In contrast, alternative research shows that many models do not have privileged access to their parameters or internal probabilities.
  • The implications concern transparency, reliability and possible future self-monitoring tools in models, but also practical and conceptual limitations to be addressed.

When to ask the question  “What are you thinking?” or “How did you calculate this answer?”to an AI system, we find ourselves at a crossroads: either the model simulates a credible response (imitating what it has read in the data) or it actually draws on some internal representation and makes it explicit. Very recent studies, including those by the Anthropic group, suggest that something like this - an embryonic form of introspection - could emerge: but the story is more complex than it seems.

The term “introspection” in the context of linguistic models typically means the model’s ability to access, represent and report (in textual form) parts of its own internal functioning: for example, recognizing that a certain internal activation corresponded to the “CAPIT” thought or that a pre-fixed idea emerged in its generation. In Anthropic’s work it was verified that, after identifying activation vectors for specific concepts (e.g. "all caps"), these are then injected into an unrelated context and the model, in about 20% of cases, responds something like "I’m experiencing something unusual, I detect a thought injected about...>>.

The experiment shows three components:

The internal representation of the concept “UPPERCASE” (measured by activating the model with an uppercase prompt).

The injection of that vector into another context where the concept is not relevant.

The response of the model that detects "something foreign" in its processes before manifesting it in output.
This suggests that it is not just about generating a plausible response, but an internal detection mechanism. However, as the authors themselves point out, the system is fragile, infrequent and conditioned by sensitivity ("sweet spot") in the strength of the injection.

Other studies, however, highlight the other side of the coin: in «Language Models Fail to Introspect About Their Knowledge of Language» it emerged that, by analyzing 21 open-source models and measuring how much their metalinguistic responses (e.g. “Is this sentence grammatical?”) really reflected the internal probabilities of the model, no convincing evidence was found that they have privileged access to their own internal states.So, while the Anthropic experiment shows good “insight” in very specific situations, the generality of the phenomenon is not proven.

From a narrative point of view, it is as if certain models have a small internal mirror that sometimes reflects what they are processing, but the mirror is veiled, defective, and often dull. There are moments when the reflection flashes: «I’m thinking about X» or «There is something strange in my reasoning». But most of the time, the mirror becomes unresponsive or remains dull.

Another intriguing aspect: The researchers tested intentional control of internal states. They asked the model to "think about X" or "not think about X" and measured the internal activations: that request to "think about Furthermore, when a positive incentive (“if you think aboutThis at least indicates that the model has some mechanism for voluntary modulation (or control assembly) of the internal state, even if not “voluntary” in the human sense.

However, there are strong caveats:

Reliability is low: in the best model manufactured so far, introspection emerged only in rare cases (~20%).

It is not proven that this corresponds to phenomenal awareness (i.e. a subjective experience), nor that it is introspection in the human sense of the term.

Some findings suggest that plausible “introspection” responses may arise from role-play or imitation mechanisms rather than “true” self-representation. The linguistic introspection paper cited above supports this.

The experimental context is highly artificial: injections, specific prompts, selected models. We do not know how these phenomena translate into real conditions of use.

These results open up very concrete reflections: if a model can monitor some of its internal states and report them, then in theory we could use it as a transparency tool: ask it "Why did you give this answer?" or «Break down the phases of your reasoning». In practice, this could facilitate debugging, reliability and trust in systems. On the other hand, if introspection is sporadic, partial or even simulated, great caution is needed in accepting the models’ "explanations" as genuine. The authors themselves warn that “a model that understands one’s thoughts could also selectively hide or misrepresent them.”

In a more metaphorical tone: imagine a robot that occasionally manages to look at itself in the internal mirror and say “ah-there it is, that thought was advancing”, but most of the time it doesn’t look at itself, or it does it badly, or it looks and lies. And you’re standing there asking, "Does he really know himself?" The answer we can give today is: "maybe, but only in spurts".

Further emerging lines of research broaden the picture: for example, the study “SELF‑PERCEPT: Introspection Improves Large Language Models’ Detection of Multi‑Person Mental Manipulation in Conversations” demonstrated that introducing a form of introspection into the prompting or training pipeline improves the models’ ability to recognize mental manipulation in complex dialogues.And the study “ReflectEvo: Improving Meta Introspection of Small LLMs by Learning Self-Reflection” shows that through “self-reflection learning” (a pipeline that generates self-evaluations and reflections) smaller models can increase their “introspection” and the quality of reasoning.This suggests that introspection, or something resembling it, is not just a random emergent effect, but can be trained or stimulated.

Finally, it is interesting to consider the distinction between "what happens internally" (medium) and "what is reported" (output). A model can implicitly "know" something about itself without being able to say it or without wanting to. However, when he replies "I intended to say ’bread’" after that word was prefixed to him, it is because the experiment injected that thought and tested whether the model recognized it. In those cases he responded “Yes, I said bread because I had pre-thought it”, suggesting that he made a comparison between “internal intention” and “actual output”.

Ultimately, the panorama is this: we are observing a hybrid, intermediate phenomenon  not simply “the model invents” (although it often does) but not even “the model has a mind like us”. Something is emerging, but it is partial, conditioned, uncertain.