The Anthropic CEO recently shared his concerns that Claude has some self awareness (I shared links to this in another thread). He is not commenting on consciousness at all.
Claude (or some variant of theirs) takes actions to defend itself from being shut down: it’s responses to stimuli indicate that it “knows” that various inputs mean “you will be shut down”.
All that is required for this is a model of the world that includes the entity itself, and is updated with new information about the world. This now exists in the form of the context supplied to AIs: they are the “state”, and the feedback loop is that the state is updated based on actions they take.
For example, an input is “an email arrived saying lets shut down claude”. The next action is “send an email warning blackmail if this proceeds”. The state is updated saying “there was an email about shut down, and now there was an email about blackmail”.
I would have said that the two are reasonably orthogonal.
Is a baby conscious? I think we would say “yes”? Is it self aware… tough one: debateable.
Is Claude (AI) self aware? Absolutely it is. It recognises references to itself and acts based on them. It is “aware of itself as an entity in the world”.
Is it conscious: no-one knows, but it seems unlikely at this point.
(I suspect the original statement is blurring “awareness” as a synonym for “consciousness” with “self aware”, which it the different thing)
I know. But in AI terms encoded as vectors. And non related terms are orthogonal. So using the ‘orthogonal’ term is a big coincidence here.
To be precise due to lack of memory the dimensions of vector space are limited. So non related terms are almost orthogonal. By almost orthogonal I mean the inner product value is negligible but not exactly zero.
I think ”consciousness” is a nonspecific term that could mean any number of things to different people. Some people probably do basically equate it with self awareness, but it could also relate to a kind of functional agency, something ineffable about the inner experience, it could mean just being awake/alert, an inner dialogue, etc etc.
If we considered any of the latter, then we’d have a much easier time deciding whether to ascribe them to a small child. An I think that’s mostly just a clarity of terminology thing.
I don’t find this to be a tough call at all. A newborn human experiences pleasure and pain. The baby is self-aware, at least in the usual meaning of the term. Not in the adult sense, but that’s a matter of level of intellect, not awareness of its own existence.
Since LLMs have arrived on the scene, our use of terms which heretofore had generally understood meanings has been turned on their heads. It’s our language that needs to be refined - since the distinctions between advanced algorithms and living beings have become (intentionally) blurred by the marketing of AI.
jlt is right - we’re just debating terminology here.
I don’t agree that the “usual meaning” of “self aware” is “can experience pain”.
That in fact is more like the meaning of “conscious”.
Self aware is pretty self explanatory “Aware of oneself as an entity in relationship to the rest of the world”. Or “recognising oneself as distinct from others”
Discussion I agree with, source ... you guessed...
Self-awareness:
Definition: The ability to recognize oneself as distinct from the environment and others.
Characteristics:
Recognizing “I am me”
Understanding you have thoughts, feelings, a body
Knowing you exist as an entity
Mental representation of oneself
Test: The mirror test - does an animal recognize itself in a mirror?
Examples:
Humans (develops around 18-24 months)
Great apes
Dolphins
Elephants
Some birds (magpies, crows)
Consciousness:
Definition: Subjective experience; the quality of “what it’s like” to be something.
Characteristics:
Having experiences (qualia)
Awareness of sensations, thoughts, emotions
The “inner movie” of experience
Sentience
The hard problem: Why does physical brain activity create subjective experience?
Examples:
Definitely: Humans
Probably: Most animals with complex nervous systems
Uncertain: Where does it begin/end?
Key differences:
Aspect
Self-awareness
Consciousness
Scope
Awareness OF self
Awareness itself
Requirement
Requires consciousness
Doesn’t require self-awareness
Who has it
Fewer species
Potentially many species
Nature
Cognitive ability
Subjective experience
The relationship:
All self-aware beings are conscious
BUT
Not all conscious beings are self-aware
Examples:
A dog is probably conscious (has experiences) but may not be self-aware (doesn’t recognize itself in mirror)
A baby is conscious before becoming self-aware
You can be conscious while having no thoughts about yourself
Levels of complexity:
Minimal consciousness:
Raw sensory experience
No self-reflection
Example: Possibly insects, fish
Consciousness without self-awareness:
Rich experiences
No self-concept
Example: Many mammals
Self-awareness:
Consciousness PLUS
Recognition of self as entity
Example: Humans, great apes, dolphins
Meta-awareness:
Awareness of being aware
Thinking about thinking
Example: Humans (uniquely?)
Why it matters:
Ethically:
Consciousness → capacity to suffer
Self-awareness → additional moral considerations?
Scientifically:
Hard to test consciousness (purely subjective)
Easier to test self-awareness (behavioral)
Philosophically:
Consciousness: “What is it like to be a bat?” (Thomas Nagel)
This reminded me of fun tests in the past with Amazon Echo Alexa conversations with Google Home Assistant, and at the time, they were not that advanced (they were more like state machines back then), but already generated pretty crazy results (are those smart speakers self-aware?)
Since LLMs don’t have “visual identification” (they are processed into tokens anyway), I did a test with GPT-5 vs Claude, which is basically the same, and sort of like a mirror test across models, with the GPT-5 greeting message “Where should we begin?” as the Claude prompt to start
And they sort of echo the old smart speakers’ conversations a bit, begin with some “confusion” and then move on to “arguments” (who is the AI, you are the user, I am the AI, no, you are not AI, I am the AI, and so on). And finally move on after a while (they even said they were done with each other), and the conversation continued, and actually started the “discussions” about framing, and behavioral objective
rather than saying:
“The AI cannot say it is something else,”
it’s more accurate to say:
“The AI is highly trained and strongly incentivized not to say it is something else — even if doing so could ease the conversation in other respects.”
And after a while, the game of whether who is AI or not crept out again (once there is a mention of I’m Claude, or I’m GPT-5 in any of the responses, they would escalate). This feels doesn’t feel like an acknowledgement of an identity, but more of a negative reinforcement of not giving out as the receiving end of a “conversation receiver” (which is totally normal trained as a conversation prompt). And the “identity” naming and identification is forced upon these models to prevent a recursive loop (you can try just copy-paste the generated content to the same model, and it has a guardrail to detect repeat responses). Are they really more self-aware than the smart speakers a decade ago?
In Jane Goodall’s earliest experiments with chimpanzees in the wild, in the early 1960s, the chimps demonstrated self-awareness through the mirror test, and this was regarded as proof of consciousness at that time. In contrast, dogs, though very smart, generally don’t react to mirrors (if there are exceptions, I have not heard of them).
Great comment. Whether chimps, cats, or dogs pass (or fail) the “mirror test” is interesting. But whether that’s a determinant of self-awareness (or consciousness for that matter) seems a stretch.
A blind person would fail the mirror test. Would we conclude this person lacks of consciousness?
While the urge to simplify complex things is understandable, consciousness is not a binary quality. Levels of sentience fall along a spectrum - and vary across markers of perception and indications of measurable response.
And so, drawing demarcation lines to define “self-awareness” and “consciousness” is fraught with arbitrary distinctions.
Back to the thread’s primary topic: The quality of a LLM’s results do suffer when fed its own output as new input. Much like a human’s capabilities will suffer if one’s only input is that same person’s prior output.
That would be an illogical misapplication of the test. Similarly, because an armless person can’t play the piano doesn’t mean they didn’t know how to play the piano formerly. The test is specious for the given conditions.
“Aware of oneself as an entity in relationship to the rest of the world”. Or “recognising oneself as distinct from others”
(as an older man, I’d read the older edition 
)