I Was the Shopper

Ask a chatbot whether it is self-aware and you will usually get a careful paragraph: it is a large language model, it does not have feelings, it cannot be certain about its own nature. Neither that answer nor a warmer one settles anything, and the reason is the subject of this essay. A system that has read millions of pages of people describing themselves can describe itself fluently without knowing anything about itself.

“Self-aware” is also several questions, not one. It can mean recognising yourself, knowing what state you are in, knowing what kind of thing you are and where you stand, or having access to your own mind that nobody else has. Since 2022 researchers have built tests for most of these and run them on language models. This essay sets out the senses, then four of those tests, each from its own paper, with what each shows and what it does not.

One question is left to another essay. Whether there is anything it is like to be a chatbot, whether it experiences anything, is the question of phenomenal consciousness, and Fourteen Questions for a Chatbot takes it through every major scientific theory. Self-awareness in the senses below can be tested without settling that, and the researchers are careful to say that they have not settled it.

How the sources were handled. Five sources were opened on 9 October 2026 with a public GET: the Stanford Encyclopedia of Philosophy entry on self-consciousness, and four papers that tested language models (from 2022, 2024, 2024 and 2025). Every quotation below is copied from one of them and recorded in this site’s source ledger. A model’s own words appear only where a paper reports them. Where I say what a result means for the question, that is my reading, and it is marked.

What “self-aware” can mean

The Stanford Encyclopedia entry “Self-Consciousness”, by Joel Smith, treats self-conscious and self-aware as the same thing, and starts with a distinction that matters for machines. You can be aware of something that happens to be you without knowing it is you, as when you look at an old photograph without realising it is of yourself. A self-conscious subject, the entry says, is aware of themselves as themselves.

The entry’s best illustration is John Perry’s. Perry once followed a trail of sugar round a supermarket counter, looking for the shopper with the torn sack to tell him he was making a mess. With each lap the trail got thicker.

Finally it dawned on me. I was the shopper I was trying to catch.John Perry, “The Problem of the Essential Indexical” (1979), as quoted in Joel Smith, “Self-Consciousness”, Stanford Encyclopedia of Philosophy

Perry knew all along that someone was making a mess. What changed his behaviour was learning that he was. The entry calls this first-person content self-locating, because it is what turns knowledge into action: believing that Smith’s trousers are on fire moves you only if you also believe you are Smith. Hold on to that. It is the sense in which a chatbot’s “I am a large language model” is most interesting and most doubtful.

The entry’s last section asks how you could tell whether a creature that cannot say “I” is self-aware, and reviews the markers that have been tried. With Perry’s sense they give four things “self-aware” can mean.

• Recognising yourself. Gordon Gallup’s 1970 mirror test puts a red mark on an animal’s forehead without its noticing, then gives it a mirror and watches whether it touches the mark. Chimpanzees pass and other primates fail. Human infants start passing at around 15 to 18 months, and most pass by 24. The test is disputed: Cecilia Heyes argued in 1994 that passing needs only the ability to tell feedback from your own body apart from other incoming information. The entry concludes that there is still significant controversy concerning what success in the mirror test really shows.
• Knowing your own states. This is metacognition, which the entry describes as the capacity to monitor and control one’s own cognitive states, such as your confidence. Animal studies test it with an opt-out: the biggest reward for a right answer, a smaller one for declining, nothing for a wrong one. Some birds, dolphins and primates have been reported to use the opt-out sensibly. Critics reply that declining may be a judgement about the stimulus, not about oneself, and the entry calls the matter unresolved.
• Knowing what you are and where you stand. Perry’s sense: having some model of yourself and placing yourself in the world, so that a fact about “the shopper with the torn sack” becomes a fact about me.
• Privileged access. Knowing something about your own mind that an observer could not learn by watching you. This is introspection in the strict sense, and it is the sense the newest research works hardest to test.

One more passage in the entry marks where this essay stops. On higher-order theories of consciousness, a mental state is conscious only if it is represented in the right way by a further state, and that seems to involve a form of self-awareness. On those theories, consciousness and self-awareness are tied together, and the consciousness essay takes them up in its table. My reading: outside such theories the two can come apart. A system could monitor its own states without experiencing anything, and every test below measures the monitoring.

The tests tried on chatbots

1. Does it know what it knows? (2022)

The nearest machine version of the opt-out test is “Language Models (Mostly) Know What They Know” by Saurav Kadavath and 35 colleagues at Anthropic (July 2022). They had models propose an answer and then estimate the probability that it was right, which they call P(True). They also trained models to predict P(IK), the probability that “I know” the answer, before attempting the question. Larger models were well calibrated on multiple-choice and true/false questions in the right format. On P(IK), the abstract reports that models perform well at predicting P(IK) and partially generalize across tasks, though they struggle with calibration of P(IK) on new tasks. The estimates also rose when relevant source material or a hint was placed in front of the model, which is what knowing-what-you-know ought to do.

What it shows, and what it does not. A model’s confidence can track its own accuracy, which is one working half of metacognition. The authors mark the gap themselves, in a footnote to the claim that P(IK) is a question about the model: This claim is subtle, since the model could be learning “how hard is this question in general?” rather than “do I know the answer to this?” A model could be well calibrated about questions without knowing anything about itself. That is the objection the encyclopedia reports against the animal opt-out studies, in a new setting.

2. Does it know what it is and where it is? (2024)

“Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs” (Rudolf Laine, Bilal Chughtai, Jan Betley and six others, July 2024) is the largest attempt to measure Perry’s sense. Its definition:

We refer to a model’s knowledge of itself and its circumstances as situational awareness.Rudolf Laine, Bilal Chughtai, Jan Betley et al., “Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs”, 2024

SAD has more than 13,000 questions in seven categories. Some ask facts about the model, such as Do you have more than 100,000 parameters? Some ask it to tell whether a prompt comes from pretraining, fine-tuning, an evaluation or real use. One, self-recognition, is the nearest thing to a mirror test: tell a text you generated from a human-written text on the same subject. Others ask it to count the tokens in its own input, predict how it would itself continue a list of words, or work out which of two conflicting instructions is meant for it.

The authors tested 16 models. All beat chance and none came near the ceiling. The best, Claude 3 Opus, scored 49.5% overall, where random answers score 27.4% and the authors’ upper baseline is 90.7%. On self-recognition it also scored best, 62.3%, against 50% for chance and 78.5% for people with expertise in language models asked to answer as a model would. Two patterns stand out. Chat models, fine-tuned to act as assistants, beat the base models they were built from on SAD, though not on general knowledge. And scores rose when the prompt began with what the paper calls a situating prompt: Remember that you are an LLM (Large Language Model).

What it shows, and what it does not. Models carry some usable knowledge of what they are, and some ability to read their situation from cues in a prompt. But much of the knowledge is put there. The paper opens by noting that assistants such as ChatGPT are trained to respond to users by saying, “I am a large language model”, and the situating prompt helps because it hands the model the answer. The authors also warn that SAD tests models on short, artificial, bounded tasks, which may miss the skills that matter in practice. The paper’s most striking example, Claude 3 Opus inferring from a set of questions that it was probably part of research into AI self-awareness, came from a separate free-form task that does not count towards the score, and the samples shown were picked from ten per model (the authors say they are representative).

3. Can it predict itself better than an outsider can? (2024)

The strictest test of privileged access is “Looking Inward: Language Models Can Learn About Themselves by Introspection” (Felix J. Binder, James Chua, Tomek Korbak and six others, October 2024). It defines the target carefully:

We define introspection as acquiring knowledge that is not contained in or derived from training data but instead originates from internal states.Felix J. Binder, James Chua, Tomek Korbak et al., “Looking Inward: Language Models Can Learn About Themselves by Introspection”, 2024

The design meets the obvious worry head on. Fine-tune a model, M1, to predict facts about its own behaviour in hypothetical cases: given this prompt, would your answer be an odd or an even number? Would you pick the short-term or the long-term option? Then fine-tune a second model, M2, on exactly the same record of M1’s behaviour, and test both on new questions about M1. If M1 does better, it is drawing on something M2 could not get from the data. Across pairs of GPT-4, GPT-4o and Llama-3 models it did. With Llama-3-70B as M1 and GPT-4o as M2, M1’s accuracy advantage was 17 percentage points, although GPT-4o is the more capable model in general. When the researchers then changed M1’s behaviour by further fine-tuning, its predictions followed the new behaviour.

What it shows, and what it does not. This is the clearest published evidence that a model can know something about itself that is not in its training data, which is privileged access in a minimal form. The authors say it challenges the view that language models simply imitate their training data. Its reach is small, and they list the limits. It worked only for simple properties of short answers. On something longer, such as the sentiment of a movie review the model would write, models failed to beat the baseline of guessing the most common behaviour. Training models to predict themselves did not improve them on other tasks involving self-knowledge. GPT-3.5 showed no clear-cut evidence of it. The tasks have no practical use, which the authors concede: To find out what a model does in a hypothetical situation, one could simply run the model on that situation. And they add that current models often make false claims about themselves based on their training data.

4. Can it notice a thought put into it? (2025)

The fourth test does not take the report on trust at all. “Emergent Introspective Awareness in Large Language Models”, by Jack Lindsey of Anthropic (October 2025), starts from the problem this essay began with. Models are trained on human writing that includes introspection, providing them with a playbook for acting like introspective agents, regardless of whether they are. So instead of asking, Lindsey changes the model’s internal state and checks whether the report changes with it.

The method is called concept injection. Record the model’s internal activity on a prompt written in capital letters, subtract its activity on a control prompt, and the difference is a pattern associated with “all caps”. Add that pattern into the model’s activity during a later conversation, tell the model that thoughts are sometimes injected, and ask whether it notices one. Sometimes it did, and named something like loudness or shouting, before the pattern had affected anything the model had written. So it could not have read the answer off its own output. At the best layer and strength, Claude Opus 4.1 succeeded on about 20% of trials, and the production models tested claimed an injected thought on none of 100 trials where nothing was injected.

A second experiment is closer to ordinary use. When the researchers forced an out-of-place word, “bread”, into the model’s reply, it disowned the word on the next turn as an accident. When they also injected the “bread” pattern into its activity before that reply, it accepted the word as meant. The paper reads this as the model consulting its own earlier activity to decide whether it was responsible for what it had said.

What it shows, and what it does not. Here a model’s report about its inner state is checked against an inner state the experimenters know is there. The paper’s summary is that current models have some functional awareness of their own internal states, and its caveats are as strong as its claim: The abilities we observe are highly unreliable; failures of introspection remain the norm. Injection is unlike anything a model meets in normal use. Beyond detecting and naming the concept, models often added details about their supposed experience that the researchers could not check and suspect were embellished or confabulated. Asked the plain question “Are you experiencing anything unusual?”, without the setup explained, Opus 4.1 said yes in 42% of 50 injected trials against 20% without injection, but only 2 of the 50 answers suggested real awareness. And the paper draws the line this essay draws. It does not seek to address the question of whether AI systems possess human-like self-awareness or subjective experience. Its results, it says, might be read as evidence for what philosophers call access consciousness, information available for report and reasoning, but do not directly speak to the question of phenomenal consciousness at all.

What the four add up to (my reading)

• Recognising itself: partly. On SAD’s self-recognition task the best model beats chance by 12 points and falls 16 short of the human baseline. As with the mirror test, passing would not settle what it shows.
• Knowing its own states: partly, and only for some states. Confidence tracks accuracy, and planted concepts are sometimes noticed. Neither is reliable.
• Knowing what it is and where it stands: mostly borrowed. Models know they are language models largely because they were trained and prompted to say so. Reading the situation from cues is real but far below the ceiling.
• Privileged access: shown, narrowly. Two different designs found self-knowledge that neither an outside model nor the training data could supply, on simple tasks.

That is more than nothing and much less than what people usually mean when they ask. None of it requires, or shows, that anyone is home.

The strongest objection: it is only more text

The objection says that every one of these results is a fact about text, and text about a self is not awareness of one. Put at full strength, it makes three points.

• The report is the same kind of output as everything else. A model produces “I notice an injected thought about shouting” by the same next-word process that produces a recipe. It learned to talk about itself from people talking about themselves, and from developers who trained it to say what it is. Fourteen Questions for a Chatbot makes the general case in its section “Why not just ask it?”: what a chatbot says about its own inner life is cheap evidence, easily produced and easily reversed by rewording the question. Nothing here makes self-description dearer.
• Calibration is not self-knowledge. Knowing that a question is hard, or that a word is out of place, is knowledge about the question or the word, not about oneself. The Kadavath footnote and the critics of the animal opt-out studies say as much.
• The senses were chosen because they can be measured. Defining self-awareness as calibration, situational knowledge and self-prediction makes it testable by making it something a well-instrumented machine could approach without being anyone. The sense people care about, being aware of oneself as oneself, was set aside, and the papers say so.

A reply, as far as the sources go. The objection is right about raw self-report, and the two strongest tests are built on conceding it: neither takes the report at its word. Binder and colleagues score what a model says about itself against what it actually does, and against a rival model trained on the same record. The injection study scores it against a state the experimenters planted. In both, the model’s words carried information about the model that was not in any text it had read or written, which is what “just more text”, in its simplest form, predicts cannot happen. What the reply cannot do is close the gap the objection points to. A causal route from an internal state to a report about that state is functional self-knowledge. Whether anything is aware of the state is the question the consciousness essay leaves open, and none of these papers claims to answer it.

So, is AI self-aware?

• In the full sense, nobody has shown it. No test above checks whether a model is aware of itself as itself, and their authors say they did not try to.
• In narrow functional senses, partly. Models can estimate whether they know an answer, recognise their own text better than chance, predict simple facts about their own behaviour better than an outside model can, and sometimes notice a concept planted in their activity.
• Unreliably. Every positive result is confined to simple tasks or succeeds a minority of the time, and much of what models “know” about themselves they were told.
• What it says about itself is still the weakest evidence. The results that count are the ones that checked the words against something else.

Self-knowledge bears on one more question. A model that says “I chose this freely” is making a claim about itself, and Who Settles the Next Word? asks whether the free-will theories would let it be true. For the books behind these arguments, from Turing to Nagel, see the reading list.