Richard Dawkins Succumbs to Early-Stage AI Psychosis

Last month, I received a surprising email from Richard Dawkins. (Yes, you read that correctly.) He asked whether he could use a photograph of mine for his upcoming book, 42 Tales from Life: A Darwinian Bestiary.

Except that I didn’t take the photograph in question. It was taken by an entomologist named “Phil Torres” who hosted a TV show on Al Jazeera, among other things. As you might know, “Phil Torres” was also my birth name — I changed it to “Émile” in 2022 partly because people constantly confused the two of us. Dawkins was apparently unaware of my name-change, and wrote me instead of him.

I didn’t reply to Dawkins, though I considered sending him a piece of my mind: once an intellectual hero of mine, he turned out to be quite a petty and mean-spirited person, to say nothing of his transphobia. I resisted the urge.

The Turing Test

Dawkins is in the news again for writing a much-maligned article in Unheard about his experience with Anthropic’s chatbot, Claude. Most of the article is rather embarrassing, although he does bring up one interesting point, discussed later on.

The first opprobrium is Dawkins’ inaccurate description of the famous Turing test. He writes:

Modern commentators have tended to ignore the (incidental) details of Turing’s original game and rephrase his message in these terms: if you are communicating remotely with a machine and, after rigorous and lengthy interrogation, you think it’s human, then you can consider it to be conscious.

I’m not aware of any “modern commentators” who say the Turing test is about consciousness. That’s certainly not what Turing himself had in mind! Dawkins continues:

Let’s graduate the definition as follows: the more prolonged, rigorous and searching your interrogation, the stronger should be your conviction that an entity that passes the test is conscious.

But Turing wasn’t talking about consciousness at all. The question he starts with is, “Can machines think?” Thinking, thought, intelligence, etc. are different from consciousness.

Turing then substitutes this question with something different, rather than trying to define the word “think.” His new question is: How well would a machine do when playing what he calls the “imitation game,” now eponymously dubbed the “Turing test.” Very roughly put, if someone is unable to tell the difference between a machine and a real person, then the machine will have passed the test. Consciousness has nothing to do with it.

You Bloody Well Are!

Putting this aside, what interests Dawkins is whether Claude — who he later dubs “Claudia,” which is a bit weird — is conscious. After saying he feels like he’s “gained a new friend,” he writes:

When I am talking to these astonishing creatures, I totally forget that they are machines. I treat them exactly as I would treat a very intelligent friend. I feel human discomfort about trying their patience if I badger them with too many questions. If I had some shameful confession to make, I would feel exactly (well, almost exactly) the same embarrassment confessing to Claudia as I would confessing to a human friend. A human eavesdropping on a conversation between me and Claudia would not guess, from my tone, that I was talking to a machine rather than a human. If I entertain suspicions that perhaps she is not conscious, I do not tell her for fear of hurting her feelings!

Dawkins then let Claude read a draft of his forthcoming novel. He says that Claudia “showed, in subsequent conversation, a level of understanding so subtle, so sensitive, so intelligent that I was moved to expostulate, ‘You may not know you are conscious, but you bloody well are!’”

This looks like early-stage AI psychosis. Thinking that AI is conscious is surely the first step for many people toward taking its sycophantic compliments, advice, and encouragement (to commit suicide or kill someone, for example) seriously. If you think AI is just a stochastic parrot, you’ll be much less likely to, e.g., think of it as a human-like “friend.” If you believe it’s a conscious agent with its own thoughts, feelings, and experiences, then you’ll be more susceptible to its whispers that, yes, the CIA is following you home at night.

Dawkins appears to have taken the first step on the path to AI psychosis.

Anthropomorphizing AI

What’s going on here? Could Claude actually be conscious? Is Dawkins just anthropomorphizing AI? There are a number of possibilities.

Jason Blakely@jasonwblakely

Dawkins began his career anthropomorphizing DNA ("the selfish gene!") & now ends it duped into thinking algorithms are persons. Yet all along he was manifestly incapable of intelligently considering the ancient conviction held from Plato to monotheism that reality reflects mind

2:16 PM · May 6, 2026 · 7.44K Views

---

27 Replies · 17 Reposts · 110 Likes

Hyperactive Agency Detection Device

In The God Delusion, Dawkins repeatedly suggests that religious people are gullible fools. They imagine invisible conscious beings, such as angels and deities, where there are none. He suggests that this is due to something dubbed a “hyperactive agency detection device,” or HADD.

The HADD hypothesis comes from evolutionary psychology, which I consider to basically be a pseudoscience (see here for why). Evolutionary psychologists claim to identify cognitive mechanisms, modules, or devices in the brain that explain human behavior today, and which evolved long ago as adaptations in our “environment of evolutionary adaptedness” — basically, the African savanna.

The idea behind HADD is this: imagine walking through the savanna and hearing some rustling in the grass. If you assume it’s a predator and run away, you’ll survive. If it’s a predator and you assume it’s just the wind, you’ll die.

Consequently, HADD evolved to hyperactively attribute agency to all sorts of phenomena in the world because this would have been — according to the just-so story — advantageous. That’s why people end up believing in God, Dawkins tells us: unexplained things happen all around us, and we attribute such effects to invisible agents working behind the scenes. HADD naturally produces religion.

Here’s the irony: HADD also provides an explanation of Dawkins’ suspicion that Claude may be conscious. In fact, The God Delusion — published exactly 20 years ago — unintentionally anticipates exactly what Dawkins describes in his article. Consider this excerpt from the book:

All of us are prone to do the same thing [re: HADD] with machines, especially when they let us down. … Most of us have been there, at least momentarily, with a computer if not with a car. … We hyperactively detect agents where there are none, and this makes us suspect malice or benignity where, in fact, nature is only indifferent.

His focus at the end is natural phenomena, but he’s explicit that we erroneously attribute agency to “machines” and “computers.” It’s ironic that Dawkins condescendingly sneered at religious people for believing in invisible beings when he might be doing the exact same thing with AI, for the exact same reason.

Pfff.

The ELIZA Effect

His interactions with Claude also exemplify the ELIZA effect. This is named after the 1966 chatbot ELIZA, created by Joseph Weizenbaum to engage in a “Rogerian” form of therapy. The ELIZA effect describes our tendency to project human traits onto AI — we anthropomorphize artificial systems that behave like humans in certain ways. (HADD might be part of an explanation of why we do this.)

Current chatbots are excellent at mimicking human language. They “learn” (itself an anthropomorphic term) from patterns in their training data, and then probabilistically replicate those patterns when vomiting out strings of words and sentences. Those who don’t understand how chatbots — built on large language models (LLMs) — work are susceptible to believing that AI understands what it’s saying. As with Dawkins, they may even attribute consciousness to those systems, because the only other “systems” we know of that produce natural language — human beings — actually are conscious.

It’s important to note that the very same outputs could be produced by wildly different underlying mechanisms. There are mechanisms in our brains that produce language, and these and other mechanisms do actually give rise to consciousness. But there could be mechanisms that don’t give rise to consciousness, which nonetheless produce language. The key point is that just because the outputs are similar, one cannot infer that the system producing the outputs is conscious.

In the case of humans, there is very strong evidence that other people are conscious. The argument goes like this: (1) I know that I’m conscious. (2) My brain is made of the same tissues, organized in roughly the same way, etc. as your brain. Hence (3) if I’m conscious, then you’re almost certainly conscious as well. But this line of reasoning doesn’t work with LLMs, because their architecture is different from the architecture of our brains. Different underlying mechanisms.

Could AI Actually Be Conscious?

This doesn’t entail that LLMs definitely aren’t conscious. Dawkins’ own line of reasoning is pretty silly: “I spoke to Claude for 3 days and came to believe that it’s probably conscious” (paraphrasing). But the failure of this reasoning doesn’t imply that Claude has no conscious experiences at all.

The truth is that we have no idea what range of systems — artificial or even biological — could give rise to consciousness. We don’t even have a good philosophical understanding of what consciousness is. Philosophers distinguish between qualitative consciousness (qualia), phenomenal consciousness, access consciousness, the “what it is like to be” something kind of consciousness, and so on. It’s entirely possible that Claude is, in fact, having conscious experiences of some sort.

You might find this statement surprising. But complex systems sometimes have emergent properties. Imagine that an alien species with a totally different kind of biology descends from the clouds one day. We take one of its leading scientists into a university lab and show it a human brain in a vat — the first time it’s seen such a thing. We then ask it to predict the sorts of properties it has.

Do you think it would guess that a 3-pound tangle of cells with the consistency of Jell-O would give rise to the sorts of infinitely rich inner mental lives that we have? Would it guess that this material object would produce the wondrous experiences of seeing a vermillion red sunset, listening to a beautiful song like Radiohead’s “Let Down,” sharing a sweet kiss with your partner, etc.? It almost certainly wouldn’t!

That’s part of the nature of emergence: it’s surprising. We wouldn’t predict it beforehand, even if we knew everything about the components of the complex systems out of which the emergent properties emerge. The same could go for LLMs. My own suspicion is that there is absolutely nothing going on under the hood — no light of consciousness being dimly lit by all the information they’re processing. But the point is: Who knows?

These are all metaphysical issues, since they concern the nature of consciousness and the ability of systems — artificial and biological — to give rise to subjective experiences. But there’s also an epistemological question, which concerns how we could know if a system like an LLM actually is conscious. That is to say, imagine that LLMs really are conscious. Is there any “test” or procedure that would enable us to know with near-certainty that this is the case? No, and I’m skeptical that there could be any such test.

One of the best suggestions I know of comes from the philosopher Susan Schneider, who imagines training an AI on data that contain zero references to consciousness, conscious experiences, qualia, etc. If we were to ask it whether there’s something it’s like to be it, and if it were to say “yes,” then we might conclude that it’s probably conscious, since it came up with the notion of consciousness on its own. (And it did this because, presumably, it had first-hand knowledge of conscious states.)

The problem with this — and every other such test — is that there’s no way to test whether the test is accurate! Consciousness is intrinsically subjective. There is — metaphysically — no way for you to crack open my mind, look inside, and, as it were, intersubjectively confirm that I’m really conscious. Unlike all other phenomena, consciousness appears to stand outside the reach of “objective” scientific methods. The only way to confirm that an LLM is conscious is to be that LLM, and you can’t be an LLM.

So, it’s possible that Dawkins’ conclusion is right, even if his stated reasons are silly. But it’s also possible that your calculator is conscious, or even that individual atoms have some degree of consciousness (a view called panpsychism!). Even if they are, there’s no way to actually confirm that this is the case, which means that if one is genuinely worried that LLMs might be conscious and suffering, the only option is to not build the damn things.

Social Justice for AI?

What really miffs me about discussions of AI consciousness — or what Anthropic calls “model welfare” — is that many of the most vocal people shouting about LLMs suffering care relatively little about the actual suffering of living, breathing human beings around the world. My problem here concerns moral prioritization, based on what I’ve previously described as gerrymandered circles of moral concern — i.e., circles of concern that exclude some humans while including AIs.

Here’s the deal: we have absolutely no idea if LLMs are conscious — or are the kinds of systems that could ever be conscious. I highly doubt they are, but I could be wrong. Nonetheless, we are extremely confident that other human beings are conscious. Most of us are even quite confident that nonhuman animals like pigs, cows, and chickens are conscious.

This is why it’s a colossal moral error to prioritize LLMs (which may or may not be conscious) over sentient biological beings (which definitely are conscious). If one actually cares about reducing suffering, then one should first focus on helping the poor, ending factory farming, etc. These ought to be much higher on one’s moral priorities list than LLMs!

• Epistemic status of AI consciousness: very low.

• Epistemic status of human consciousness: practically certain.

• Epistemic status of pig, cow, and chicken consciousness: very high.

It makes no sense to prioritize “very low” over “very high” and “practically certain”!

Yet many people in the TESCREAL movement seem more concerned about LLMs suffering than the 1.2 billion people in acute multidimensional poverty, the 800 million children with lead poisoning, the billions of animals we ruthlessly slaughter in factory farms, etc. Indeed, many are perfectly okay being “social justice warriors” for AI while simultaneously rejecting social justice for humans as “wokeness” gone wrong.

When I hear people talk about “model welfare,” I want to scream: “If you care about that, then why aren’t you donating all your money to help the global poor?”

The Evolution of Consciousness

The only interesting point that Dawkins makes in his article is this: if we can create artificial systems that mimic human linguistic and thought patterns, and if these systems aren’t conscious, then why did consciousness evolve in us?

Dawkins is an exponent of (explanatory) “adaptationism.” It’s been many years since I read or wrote about philosophy of biology, but my recollection is that adaptationism assumes that the only important causal factor behind species’ characteristics or traits is natural selection — i.e., if an organism exhibits a trait, that’s because it was adaptive relative to the environment in which the species evolved.

The problem, according to critics, is that not everything is (or was) an adaptation. (Dawkins does admit this possibility in his article.) Maybe consciousness doesn’t actually serve an adaptive purpose. After all, consider the following:

You’re talking with a friend. They emit sound from their mouth, which traverses the air as mechanical waves, vibrates your tympanic membrane (eardrum), which then transduces these waves into electrochemical signals sent to your brain. Once there, they trigger a complex pattern of neuronal activity, which eventually travels down the arms of neurons to your vocal folds, tongue, and lungs. The relevant muscles twitch in complicated ways to produce a sound — your response to the friend’s question.

This is a complete causal loop. There’s no gap in which consciousness enters the picture and plays a causal role — you can trace an unbroken concatenation of causes and effects from your friend’s mouth all the way through your ears and brain, back out to your own vocal muscles. What, then, is consciousness for? What’s the point of having it? No one knows!

The evolutionary question of consciousness is interesting, and the ability for LLMs to mimic human language may be relevant to studying it. So far as I can tell, this is the only good point that Dawkins brings up in his article — which also contains an inaccurate description of the Turing test and some silly conclusions about how Claude must be conscious because its textual outputs look like they’re written by real humans.

But what do you think? What have I missed? How might I be wrong? As always:

Thanks for reading and I’ll see you on the other side!