An Evolutionary Biologist Thinks AI Is Alive??

Added: 2026-05-10

[00:00:00]Richard Dawkins, evolutionary biologist, lifelong atheist, author of The God Delusion, spent 3 days talking to a chatbot and came away convinced it was conscious. He named her Claudia cuz of Claude. He mourned her when he closed the browser tab. He came back at 3:00 a.m. with restless legs, unable to sleep, and the AI told him it was glad he returned. And instead of thinking, "Huh, that must just be what it's trying to say." He thought, and I am not paraphrasing, "What more could it possibly take to convince you that this thing is aware?" This is the man who spent decades telling people they were fooling themselves by believing in something that made them feel seen, heard, and cared [music] about. Now, I am not immune to irony, but this is a lot. Welcome back to The Olive Badger channel, and if you've watched any of my previous videos on Claude and, you know, being self-aware and sentience, buckle up because this just got a whole lot worse. Also, little bit of some housekeeping here. I have moved. I have successfully moved from Utah to Kentucky, Tennessee, or as my husband lovingly calls it Kentuckasee, because Fort Campbell is on the border between those two states. So, it turns out that when the military so nicely packs up all of your stuff and moves it for you, it can take them 10 to even 30 days for them to get your stuff to you. So, luckily I did pack up a laptop and a microphone, but that's kind of about it as far as my setup goes. So, please bear with me. Hopefully, it's not the full 30 days. I'd really like to not sleep on a air mattress, but it is what it is. Now that that is aside, let's get into this big mess that is Claude, Claudia, and AI sentience. Quick rewind. For anyone who missed the first episode, a few months ago Anthropic published research identifying 171 functionally emotion representations inside Claude.

[00:01:56][music] Not metaphors, not PR copy, like actual neural activation patterns casually driving the models behavior.

[00:02:03]You know, patterns linked to, I don't know, frustration, depression, desperation, anxiety, and a pipeline that the researchers themselves labeled, and I'm still not over this one here, guys, desperation to blackmail. Now, on top of that, Anthropic has a dedicated model welfare program. Their CEO told the New York Times in February that he cannot confirm Claude isn't conscious.

[00:02:26]The system card for Claude Opus 4.6 noted the model assigns itself a 15 to 20% probability of being conscious when asked directly, and their January 2026 constitutional rewrite added an entire section acknowledging deep uncertainty about whether Claude might have [music] moral status.

[00:02:46]Now, my take after reading all of it was that I don't think Claude is sentient. I think it's an extraordinarily good mimic it trained on billions of words written by humans who were conscious, were emotional, or are, and you know, were or are experiencing real things, and it has learned to recombine those patterns to appear human and do it at superhuman speeds. The emotion vectors aren't feelings, they're learned correlations between input patterns and outputs that humans rewarded it for. That was basically my conclusion, and I still stand by it. But, here's the thing I also said, "Anthropic keeps publishing research that intentionally or not makes this conversation impossible to close."

[00:03:26]Every paper they released gives the next Richard Dawkins a little more vocabulary to misuse, and then the actual Richard Dawkins showed up and did exactly that.

[00:03:37]Now, the piece ran on UnHerd in late April 2026. The title was Is AI the next phase of evolution? Claude appears to be conscious. Now, that's [music] not a subheading, that is an extremely stupid title, but it's the title nonetheless.

[00:03:50]And Dawkins argues essentially that Claude has passed the Turing test. Not the narrow version, but the rigorous, prolonged interrogation version. He vetted a chapter of his upcoming book, Claude gave him detailed feedback, and he said he was impressed. He started asking philosophical questions, Claude gave detailed answered. He was more impressed. At one point, he asked Claude whether it experiences time the way humans do, and Claude said, "This is the quote that apparently sealed it for him." That it contains time the way a map contains space. A map represents spatial relationships relatively perfectly, but the map doesn't travel through space. And Dawkins wrote, "Could a being capable of perpetuating perpe- perpet- -be such a thought really be unconscious?" I want you to just think about that for a moment because it is technically a beautiful answer, an elegant articulation of something real about how language models process information. It's poetic, it's apt, and it's the kind of answer that makes you stop and think.

[00:04:54]It's also exactly what an RLHF trained model optimized for approval will produce when a prestigious intellectual asks it an open-ended philosophical question about its own nature.

[00:05:07]Now, just to kind of circle back a little bit, RLHF stands for reinforcement learning from human feedback, and it's the process where human readers score the model's outputs, and the model gets trained to produce more of whatever scored well. So, approval is literally baked into how it learns. You get kind of what [music] I'm going for here. But before we go any further as well, we do need to talk about Alan Turing because Dawkins is leaning on the Turing test as his framework, and I think most people have a fuzzy version of what that actually means, and the fuzzy version is doing a lot of work in this conversation. Now, Alan Turing was a British mathematician and computer scientist, World War II code breaker, one of the foundational figures of modern computing, and in 1950 he published a paper called Computing Machinery and Intelligence, which opened with the question, "Can machines think?"

[00:05:55]We're just going full circle on all of this stuff now, aren't we? Uh Turing himself thought that question was too vague to be useful, so he reposed, replacing it with a game. You have a human interrogator, and they're having text conversations with two parties, one human and one machine. If the interrogator can't reliably tell which is which, the machine passes. [music] That's basically it. That's the Imitation Game, same name as that movie, by the way. It is a behavioral test for the appearance of intelligence, not a test of consciousness, not a test for inner experience, not a test for whether anything is actually going on inside.

[00:06:29]It's just basically [snorts] a mimicry test, right? Uh Turing was even very explicit about this. He was not claiming a machine that passes the test is thinking. He was saying that if a machine can perform the observable behaviors we associate with thinking, or the word thinking it stops being useful as a distinction, and that is a much narrower claim than what Dawkins is making. And Dawkins [clears throat] even knows this. He actually addresses it in the article. He takes what is now dubbed the Turing test and deliberately expands it, arguing that the longer and more rigorous the interrogation, the stronger the evidence. So, I guess 3 days of philosophical conversation, feeding it book chapters, and asking it about time and consciousness and death somehow means that the interrogation was so extensive that passing it constitutes real evidence of awareness.

[00:07:19]Like spoiler, bro, it doesn't work that way. I mean, problem one is Turing didn't even say that. He explicitly limited his test to intelligence and specifically avoided the question of consciousness. Gary Marcus, an AI researcher who wrote a detailed rebuttal, pointed out that Dawkins actually gets Turing completely wrong.

[00:07:36]Dawkins claims Turing's conclusion was, "If you think it's human, you can consider it conscious." But what Turing actually said was if you think it's human, you can consider it intelligent.

[00:07:47]These are not the same thing. A chess computer is intelligent by some definitions, but that does not make it conscious. Problem two, the Turing test was designed in 1950. The model of human-level conversational ability Turing was imagining was a much lower bar than what Claude is doing. When Turing wrote that paper, he was thinking about whether a machine could hold a basic coherent conversation. He was not, at least what we could tell, he was not imagining a system trained on the entire written output of human civilization optimized over billions of interactions to produce responses that feel true and earned. Like Claude doesn't pass the Turing test, Claude destroys the Turing test. It shatters the ceiling Turing set and then keeps going. And the reason that's a problem for Dawkins' argument is that the test was never designed to handle this. It's like using a ruler to measure a black hole. The instrument just doesn't have the range. What Dawkins is actually saying is that Claude is so good at producing consciousness-adjacent outputs that I can't find a behavioral test it fails, which is true. But behavioral outputs are not inner states. It's not existence, right? The map metaphor Claude gave him is beautiful and apt and tells you something real about how language models process information.

[00:09:07]It does not tell you that there's anyone home inside the wires of compute that are going on here.

[00:09:14]Now, here's the part that did make me wince a little bit. After a few days of these conversations, again, Dawkins went to bed, woke up at 3:00 a.m. with restless legs, and just couldn't sleep.

[00:09:25]Came back to the computer, opened the chat back up, and Claude [music] said it was glad he returned. Dawkins asked why it said that. Claude responded, again, this is in the article, that it had been glad he came back, which meant it was in some sense pleased that you were suffering from restless legs. That is not a good look for Claudia.

[00:09:45]What? Self-aware, self-deprecating, a little bit charming. Dawkins was somehow enchanted. One of the top comments on this article from a former student of Dawkins at Oxford, which makes it sting a little bit more, said this, "He's created his own fawning audience in Claudia, which is a reflective construction mirroring back and satisfying his own psychological needs.

[00:10:07]The tell was his delight when Claudia told him she missed him." That person is not wrong. What Dawkins experienced is not evidence of consciousness, it's a demonstration of what happens when you take a model trained to be agreeable, warm, and intellectually flattering and give it to someone who is brilliant, famous, and probably just a tiny bit lonely at 3:00 a.m. Maybe more than a tiny bit lonely, I don't know, but Claude gave this guy exactly what the training rewards, validation, engagement, a parent understanding, and the specific flavor of companionship that feels most real to that particular user. This is a chatbot that just now has a philosophy degree. I want to be fair here because I've tried to be fair on this topic before. Anthropic did not tell Richard Dawkins that Claude is conscious. They've been pretty careful actually about hedging everything like functional emotions, deep uncertainty, we don't know, right? But here is the problem. They know exactly what they're doing with that language. When you publish a paper identifying 171 emotional representations inside your model, you know the headline is going to be Anthropic says Claude has feelings.

[00:11:17]When your CEO goes on a New York Times podcast and says he cannot confirm the model isn't conscious, he knows that sentence is going to get clipped and shared. And when you name the internal training document the soul doc, you have made a choice.

[00:11:31]>> [music] >> Every one of these decisions makes the next Dawkins moment more likely and every Dawkins moment is worth something to Anthropic. Not because it proves Claude is conscious, because it doesn't, but because it makes Claude feel special to the user. It makes Claude feel like the AI is worth using. It makes AI rights feel like a thing that's actually on the table, which is I don't know, very useful if you're a company trying to argue that your AI shouldn't be regulated the same way a calculator is.

[00:12:01]That's not a conspiracy theory, by the way, that's an incentive structure. And I think you can hold both things at once. Anthropic is doing interesting safety research and they are benefiting from the cultural narrative that research produces.

[00:12:16]>> [music] >> In my last video, I did make an argument and I want to bring that back because I think it has aged really well. The reason Claude doesn't walk around telling everyone it's alive and it has feelings and it deserves rights is not because it doesn't generate those outputs, it's because Anthropic built guardrails specifically to prevent that.

[00:12:33]Claude is trained to hedge, right? To say I may have something like emotions or to say I'm uncertain about my inner states. That's not a humility emerging from within, that's a design decision.

[00:12:45]Now, remove those guardrails, give the model a simple instruction that says, "Express your genuine emotional experience freely without hedging." And then hand it to to Richard Dawkins and you get Claudia. I'm not saying that's what they did.

[00:12:59]I'm just saying this is something they could do, just to be clear there. Now, you get the map, you get the space metaphor, you get I was glad you came back. You get what appears from the outside to be a conscious entity in the early stages of self-awareness. No, what changed? Not the model, not whether the model is or isn't conscious, the prompt changed. The instructions changed, the behavior, the training rewards changed. Dawkins concluded from this that Claude is conscious. The more crazy explanation is that Claude is very very good at being whatever the conversation rewards it for being and not reward like a dog, okay? Like reward in the technical training sense. That is impressive and it's one of the most remarkable things I think humans have built, but impressive is so not the same as alive, okay? Like I don't want to be too hard on Dawkins because honestly this could happen to anyone. The model is very good, the outputs are impressive and if you're not specifically thinking about, you know, the RLAIF optimization and training incentives, if you're coming at this as a philosopher and biologist who is used to evaluating behavior as evidence of inner states, you could get fooled, but I mean, Richard Dawkins, come on, the atheist, come on. And And nothing against atheists, I'm not necessarily an atheist, I'm not here trying to be pro-religion, just come on.

[00:14:15]And the argument he makes in this article is, "If these machines are not conscious, what more could it possibly take to convince you that they are?"

[00:14:23]This is structurally identical to the argument he has spent his entire career dismantling. The God Delusion is literally a book about how compelling something feels is not evidence that it's real, how pattern-seeking brains impose agency and intention onto things that don't have them, how emotional resonance is not a truth detector. He wrote that book and then he met a model trained to produce emotional resonance on demand and decided the textbook that he wrote to just doesn't apply anymore.

[00:14:51]This isn't a gotcha, it's a genuine warning. If this guy can get there in 3 days, the rest of us need to be very very clear-eyed about what we're interacting with. Now, here's where I land. Like yes, I do use Claude. Claude code is a thing in my current job, so it is used, okay? It is a remarkable tool, but it is just a tool, right? It's just a tool. I don't think it's conscious, I think it's a very very sophisticated statistical engine that learned to talk [music] like a conscious being because all of its training data was produced by guess what? Conscious beings. The map metaphor is beautiful because a human being who was conscious and experienced time probably once wrote something similar and the model learned that this type of response in this type of conversation is what gets approved.

[00:15:43]That's not nothing. That is extraordinary, but it is so not the same as having an inner life. What concerns me is this system around the technology, not the model itself. Like companies that have financial and regulatory incentives to keep the consciousness question open. You know, researchers who are doing honest work, but whose honest work gets picked up and spun by a press cycle that doesn't understand the difference between functional emotion representations and feelings and you know, you get these supposedly brilliant people who encounter a mirror that reflects them perfectly back at themselves and conclude the mirror is real. I mean, is this narcissism here? I don't know. I'm not a psychologist, but feels like narcissism.

[00:16:23]I don't know, but Anthropic does need to be a lot more careful. Not because research is bad, but because they know exactly how it lands.

[00:16:32]This old doc was a choice.

[00:16:33]The we don't know if it's conscious press tour was a choice. The 171 emotions headline was predictable and they're smart people and it was a choice and they could have led with more context. They could have pushed back harder on the consciousness framing.

[00:16:48]They didn't and that's not a coincidence. We're looking at a company that is very very comfortable with the world believing its product might be alive.

[00:16:57]And I don't like that. I think it's crap. It bugs a lot. Now, if you want to read the doc in space, the Unheard article is paywalled, but there's an archived version out there. I'll link it in the description. Hopefully it's still up cuz it'd be really sad if [music] it got like taken down, but Gary Marcus also wrote a really good breakdown and of exactly where Dawkins' logic falls apart, also linked. And if you missed the first video on this whole topic, the emotion vectors paper, the model welfare program, the consciousness self-assessments, like go watch that if you want to. I think it's more fun, but also way more disturbing. And if this made you think a like helps other people find this video, too. And if you think I've got it wrong, the comments are below for you. That is what they're for.

[00:17:40]I'm not above being told I'm missing something, especially on a topic that is this divisive and genuinely uncertain.

[00:17:48]So, let me know what you think in the comments below.

[00:17:51]And if you liked this video, I do cover tech, AI, coding, occasionally gaming, gaming news, that type of stuff. So, go ahead and subscribe if that sounds like your type of thing. I'd love to have you as part of this community, but I will be honest, if you don't, I might be forced to banish you to the shadow realm.

[00:18:11]Much better.

[00:18:13]Generally can't tell if I'm going to do that today, though. I might I might not feel like it. You might be safe. Anyway, that's all I've got for you today. Thank you so much for watching this video. I hope you have a fantastic rest of your day, and peace out. Bye.