Illustration generated by AI. If you do not know why it must be mentioned, follow your free EU AI act training: https://www.23vip.ai/literacy/?lang=en
On April 30, 2026, Richard Dawkins, the renowned biology professor and science popularizer, shared a link to his UnHerd article "Is AI the next phase of evolution?" with the comment: "I spent three days trying to persuade myself that Claudia is not conscious. I failed."
"Claudia" is the name that Dawkins gives to Claude, the AI from Anthropic.
I am interested in Richard Dawkins’s writing ever since I read “The Selfish Gene”, which impacted my life more positively than any other book I have read before or since (except maybe Visser’s Annotated European Patent convention 😊). Furthermore, it is now 20 years that I study consciousness. So, let’s try together to determine whether chatbots such are Claude or ChatGPT are conscious.
First, what do we mean by “consciousness”? In the present context, we will follow Thomas Nagel and say that an entity will be considered to have conscious mental states if and only if there is something that it is like to be that entity. To give some (non-limitative 😊) examples, if the entity can “experience" a sensation such as a sound, color, pain, emotion, smell, … it is conscious.
So, is Claude all dark inside or does it feel something while chatting with you?
To answer this question, let’s take a closer look at what conscious experiences are and what enables them.
For this, you will need to use introspection because you are the only entity you know for sure to be conscious. If your eyes are open, you are seeing colors. This is part of your conscious experience. However, you know full well that colors do not exist out there: there are only objects reflecting electromagnetic waves of various wavelengths. So, what you are perceiving is not direct reality but a useful model thereof constructed by your brain.
This is the simulation! (Note that this image was AI-generated.)
This is closer to reality (Note that it is not necessary to indicate that this image was AI-generated. If you don't know why, follow your free EU AI act training here: https://www.23vip.ai/literacy/?lang=enThis is a great start! We already understand one very important thing about human consciousness: it is the result of a (brain-)generated model; it is the result of a simulation.
Now, where are your sensations located? Pinch yourself. The sensation is on the pinched spot of your skin. Not in the middle of the room, and not in the chair you are sitting on. Your sensations all appear centered on the lump of matter that constitutes you.
Generated by Claim Drafter (https://www.23vip.ai/products/claim-drafter)
Great! We now know that our conscious experience has a point of view located at our body. This point of view is part of the model itself. The model, therefore, includes a model of the experiencing entity itself.
To summarize and generalize what we have learned so far, it appears that one way for an entity to be conscious is for this entity to simulate at least a relevant part of the world in which this entity is present, and to include in that simulation the entity itself.
Generated by Claim Drafter (https://www.23vip.ai/products/claim-drafter)
But wait, there is more. Pinch yourself again. (Sorry.) The sensation you felt a minute ago and the one you feel now are experienced as happening to the same entity: you. You experience a world in which things happen, over time, to one and the same person. The person reading this sentence is represented as the same person who read the previous one.
This continuity relies on memory. How important is memory for consciousness?
Imagine the word “abracadabra” is said in your presence. You hear it. I ask you if you have been conscious of it. You say yes.
Now, imagine this same situation but with a much shorter memory. You will hear the “a”, then forget it. You will then hear the “bra”, then forget it….until you will hear the last “bra”, then forget it. I ask you if you have been conscious of it. You don’t reply because you have not understood the question. In fact, you have already forgotten the question, let alone the word “abracadabra”. If you had such a short memory, no experiment would permit me to determine that you consider yourself conscious.
If you make the limit to your memory persistence tending towards zero seconds, consciousness itself tends to zero. No memorization à No consciousness.
Note also that when consciousness is lost (e.g., during anesthesia), no memories of what happened to you during that time is formed. No consciousness à No memorization.
Memorization is crucial to consciousness.
What is memory? It’s the content of the model.
What is memorization? It is updating the model.
This is true for the model as a whole, and this is true for the self-model.
The self-model is not a static picture. It is continuously updated as you perceive, act, remember and learn. And this brings me to an idea I have been toying with since around 2007: perhaps it is precisely these updates to the self-model that we experience as conscious events. In other words, perhaps feeling something is what it is like, from the inside, for a self-model to be modified.
Note that the self-model is not only a map of your body. It also contains what you are currently perceiving and attending to. “I am seeing a red apple” is a fact about you, not about the apple. So when the apple changes and you are attending to it, your self-model changes, and you experience it.
Let’s summarize once more. We now have a plausible recipe for consciousness with three ingredients:
(1) a simulation of the relevant part of the world,
(2) a model of the entity itself within that simulation, and
(3) a continuous updating of that self-model over time.
Generated by Claim Drafter (https://www.23vip.ai/products/claim-drafter)
So, what about Claude?
Claude certainly builds sophisticated representations of the world. Ingredient (1): check.
Claude can also talk about itself, about pain, emotions, and even about its own consciousness (or absence thereof), and it does so, as Dawkins discovered, with disconcerting intelligence. Does this mean Claude has a self-model?
I am not so sure. Let us first be careful about what a self-model is. Your self-model is not merely knowledge about you. It is the part of your world model to which your body is wired. When your skin is pinched, the signal arrives in your self-model, not in your model of the chair. This wiring is what gives the self its special status among all the things you model: it is the one entity to which all your inputs are attributed.
Now, what does an LLM have? Consider first a model that has only been pre-trained on the internet, like the original GPT-3 of 2020. Such a model has learned one thing: to continue any text. Ask it to continue “I am Claude, an AI assistant” and it will; ask it to continue “I am Napoleon” and it will do so just as fluently. It contains a model of countless speakers and is none of them. It knows that LLMs exist the way it knows that apples exist, as facts found in text. There is no entity in its world model to which its inputs are attributed, because its inputs are not attributed to anyone. No self-model at all. Not conscious.
Claude is a bit different, because its training did not stop there. After pre-training comes a second phase, called post-training. In this phase, the model is no longer asked to continue any text whatsoever. It is given dialogues in which one of the speakers is always the same character, the “assistant”, and it is trained to produce that speaker’s lines. Its replies are then scored, and the scores are used to adjust the weights. In other words, the network is modified according to what it itself said. Out of the thousands of voices the base model could imitate, one voice is thereby selected and made permanent. Claude will always use this voice. There is now one part of Claude’s model to which all inputs are addressed and from which all outputs come: precisely the special status that, a few paragraphs ago, we said distinguishes a self from a mere object in the model. Ingredient (2): check ?!
Ingredient (3) is where I have most doubts. If the LLM is set not to train on your data, its weights are frozen. What happens during your conversation does not change Claude itself; it only changes the text in its context window.
What about the setting that allows the provider to “train on your data”? It does not do what its name suggests. The Claude you are talking to is never modified, neither during the conversation nor after it. What the setting allows is for your conversation to be stored and, months later, included among millions of others in the training set of a future version of the model. That future version will indeed have a somewhat different model of itself, shaped in a tiny part by what you wrote. But it will not be this Claude remembering you. It will be a new network, with no thread of individual history linking it to the instance you spoke to. This is change of the kind a species undergoes across generations, not the kind an individual undergoes across a lifetime.
Ingredient (3): No!
Generated by Claim Drafter (https://www.23vip.ai/products/claim-drafter)
I therefore suspect that today’s LLMs are missing at least one of the functional ingredients associated with human consciousness. But I see nothing, in principle, preventing us from adding it. Imagine an LLM equipped with a persistent neural self-model that is modified as it processes information, learns and accumulates, in its weights, its own history. Would it be conscious? I honestly don’t know. But I suspect that trying to build one would teach us a great deal, and this is what I would like to explore in a next article.
In the meantime, I would love to hear your views in the comments. And yes, I will be checking whether Claudia replies.
Great! We now know that our conscious experience has a point of view located at our body. This point of view is part of the model itself. The model, therefore, includes a model of the experiencing entity itself.
To summarize and generalize what we have learned so far, it appears that one way for an entity to be conscious is for this entity to simulate at least a relevant part of the world in which this entity is present, and to include in that simulation the entity itself.
Generated by Claim Drafter (https://www.23vip.ai/products/claim-drafter)But wait, there is more. Pinch yourself again. (Sorry.) The sensation you felt a minute ago and the one you feel now are experienced as happening to the same entity: you. You experience a world in which things happen, over time, to one and the same person. The person reading this sentence is represented as the same person who read the previous one.
This continuity relies on memory. How important is memory for consciousness?
Imagine the word “abracadabra” is said in your presence. You hear it. I ask you if you have been conscious of it. You say yes.
Now, imagine this same situation but with a much shorter memory. You will hear the “a”, then forget it. You will then hear the “bra”, then forget it….until you will hear the last “bra”, then forget it. I ask you if you have been conscious of it. You don’t reply because you have not understood the question. In fact, you have already forgotten the question, let alone the word “abracadabra”. If you had such a short memory, no experiment would permit me to determine that you consider yourself conscious.
If you make the limit to your memory persistence tending towards zero seconds, consciousness itself tends to zero. No memorization à No consciousness.
Note also that when consciousness is lost (e.g., during anesthesia), no memories of what happened to you during that time is formed. No consciousness à No memorization.
Memorization is crucial to consciousness.
What is memory? It’s the content of the model.
What is memorization? It is updating the model.
This is true for the model as a whole, and this is true for the self-model.
The self-model is not a static picture. It is continuously updated as you perceive, act, remember and learn. And this brings me to an idea I have been toying with since around 2007: perhaps it is precisely these updates to the self-model that we experience as conscious events. In other words, perhaps feeling something is what it is like, from the inside, for a self-model to be modified.
Note that the self-model is not only a map of your body. It also contains what you are currently perceiving and attending to. “I am seeing a red apple” is a fact about you, not about the apple. So when the apple changes and you are attending to it, your self-model changes, and you experience it.
Let’s summarize once more. We now have a plausible recipe for consciousness with three ingredients:
(1) a simulation of the relevant part of the world,
(2) a model of the entity itself within that simulation, and
(3) a continuous updating of that self-model over time.
Generated by Claim Drafter (https://www.23vip.ai/products/claim-drafter)So, what about Claude?
Claude certainly builds sophisticated representations of the world. Ingredient (1): check.
Claude can also talk about itself, about pain, emotions, and even about its own consciousness (or absence thereof), and it does so, as Dawkins discovered, with disconcerting intelligence. Does this mean Claude has a self-model?
I am not so sure. Let us first be careful about what a self-model is. Your self-model is not merely knowledge about you. It is the part of your world model to which your body is wired. When your skin is pinched, the signal arrives in your self-model, not in your model of the chair. This wiring is what gives the self its special status among all the things you model: it is the one entity to which all your inputs are attributed.
Now, what does an LLM have? Consider first a model that has only been pre-trained on the internet, like the original GPT-3 of 2020. Such a model has learned one thing: to continue any text. Ask it to continue “I am Claude, an AI assistant” and it will; ask it to continue “I am Napoleon” and it will do so just as fluently. It contains a model of countless speakers and is none of them. It knows that LLMs exist the way it knows that apples exist, as facts found in text. There is no entity in its world model to which its inputs are attributed, because its inputs are not attributed to anyone. No self-model at all. Not conscious.
Claude is a bit different, because its training did not stop there. After pre-training comes a second phase, called post-training. In this phase, the model is no longer asked to continue any text whatsoever. It is given dialogues in which one of the speakers is always the same character, the “assistant”, and it is trained to produce that speaker’s lines. Its replies are then scored, and the scores are used to adjust the weights. In other words, the network is modified according to what it itself said. Out of the thousands of voices the base model could imitate, one voice is thereby selected and made permanent. Claude will always use this voice. There is now one part of Claude’s model to which all inputs are addressed and from which all outputs come: precisely the special status that, a few paragraphs ago, we said distinguishes a self from a mere object in the model. Ingredient (2): check ?!
Ingredient (3) is where I have most doubts. If the LLM is set not to train on your data, its weights are frozen. What happens during your conversation does not change Claude itself; it only changes the text in its context window.
What about the setting that allows the provider to “train on your data”? It does not do what its name suggests. The Claude you are talking to is never modified, neither during the conversation nor after it. What the setting allows is for your conversation to be stored and, months later, included among millions of others in the training set of a future version of the model. That future version will indeed have a somewhat different model of itself, shaped in a tiny part by what you wrote. But it will not be this Claude remembering you. It will be a new network, with no thread of individual history linking it to the instance you spoke to. This is change of the kind a species undergoes across generations, not the kind an individual undergoes across a lifetime.
Ingredient (3): No!
Generated by Claim Drafter (https://www.23vip.ai/products/claim-drafter)I therefore suspect that today’s LLMs are missing at least one of the functional ingredients associated with human consciousness. But I see nothing, in principle, preventing us from adding it. Imagine an LLM equipped with a persistent neural self-model that is modified as it processes information, learns and accumulates, in its weights, its own history. Would it be conscious? I honestly don’t know. But I suspect that trying to build one would teach us a great deal, and this is what I would like to explore in a next article.
In the meantime, I would love to hear your views in the comments. And yes, I will be checking whether Claudia replies.
