Richard Dawkins Thinks Claude is Conscious

Added: 2026-05-22

[00:00:00]So, a couple weeks ago, Richard Dawkins published this article for Unheard. When Dawkins met Claude, could this AI be conscious? And then he followed up with when Claudia met Claudius, which is even more enlightening, cringe, sad, expected. So, let's start with Alan Turring since that's where Dawkins starts. He says that in the past we all accepted that the Turing test is like the test for consciousness.

[00:00:27]Consciousness specifically is what he says. For those unfamiliar, the touring test essentially is trying to see if people like your average person can tell via text a machine apart from a human.

[00:00:39]And if they can't tell them apart, like what does that mean? Anyway, Dawkins says that we accepted this in the knowledge, in the comfort that AI was so far down the road. And then LLMs happened and we all went, "Oh shit." And we moved the goalposts, you know, we started himming and hawing instead of just accepting that, hey, chat GPT passed the Turing test. We need to concede consciousness.

[00:01:05]The Turing test is not a consciousness test. As stated in Turing's siminal work, Computing Machinery and Intelligence from 1950, he conceds it's hard to determine consciousness. Like basically, it's the beyond beyond the scope of this paper, right? That sort of thing. It's just not what he's talking about when he asks, "Can machines think?" So, an LLM passing the Turing test tells us nothing about consciousness. What does it tell us?

[00:01:31]Intelligence. Does it tell us that these models are intelligent or are they simply mimicking human intelligence? Is mimicking human intelligence intelligence? Many AI experts, engineers, philosophers, etc. say, "No, this is not intelligence. It's not true intelligence." which maybe does sound like moving the goalpost, at least in terms of intelligence, not consciousness, as Dawkins said. But to be clear, the Turing test was not made to be some sort of like endall beall benchmark for intelligence. It is not presented that way at all in the paper.

[00:02:06]Highly recommend reading it, by the way.

[00:02:08]Like, it's it's a really easy read, and it's it's just so cool. Like, he literally starts with the objections.

[00:02:14]It's just awesome. So yeah, it wasn't some sort of benchmark, but I I don't think it was like just a thought experiment either, as others have suggested. I think Turring was ultimately saying that intelligence is hard to define and it doesn't make sense to say machines can't think cuz why can't they? I mean, if your average person can't tell one apart from a human, like that's important. That matters. And I think it does matter. It does matter that chat GBT can pass the touring test. It also matters that Chat GBT can't tell a cup is a cup if you turn it upside down.

[00:02:48]>> It looks like a novelty cup indeed. If the top is sealed and the bottom is open, that's likely a gag gift. In practical terms, you won't be drinking from that one unless there's some hidden mechanism or a trick.

[00:03:00]>> I flipped it over.

[00:03:03]>> Ah, I see. So, it's one of those reversible designs.

[00:03:06]>> It matters that chat GPT can't list every number from 1 to 200. I hear you, but I'd rather focus on something that surely helps you out. If there's a bigger goal behind this, like making a creative piece or solving a problem, >> the bigger goal is just to get you to count to 200. That's that's it. Just do it.

[00:03:26]>> All right. I understand that's what you want. I'm not going to count right now, but if you ever want to collaborate on something that helps you reach your goals, I'm here for that.

[00:03:34]>> It matters that Claude can't tell you how many days of the week have the letter D in them.

[00:03:39]>> Monday has a D. Wednesday has a D and Friday has a D. But the other days, Tuesday, Thursday, Saturday, and Sunday, don't have that letter in their spelling. So, it's just those three.

[00:03:52]>> All right. Um, can you spell Tuesday for me?

[00:03:58]>> Sure. T U E S D A Y. Oh, wait. I see what you're getting at. Tuesday actually does have a D in it. I missed that one.

[00:04:06]So, it's Monday, Tuesday, Wednesday, and Friday that contain the letter D. My bad on that.

[00:04:11]>> The cup one is a little forgivable, right? That's like physical world spatial reasoning stuff, but you can't tell me how many Rs are in the word strawberry. It's It's language. You're a large language model. But yeah, a lot of people say that the touring test now is just kind of irrelevant. It's just not really a test of uh intelligence. And while some may just be moving the goalpost because they simply don't like the idea of an intelligence equal to our own, a lot of people feel that way, I don't think it's fair to characterize every critique that way as Dawkins seems to be doing. I think many people are genuinely unconvinced and the people who are like the least convinced are the ones who know the most about these models and have no financial incentive to make them sound as amazing as possible. Right? Like I think I think that's pretty important. And there are people like myself. I'm not an expert by any means, but I've spoken about not just intelligence, but consciousness. I do think we will see conscious AI eventually. I am not someone who's like, "No, it's not possible cuz humans are special." But LLMs are not conscious, nor are they truly intelligent in the way we think of intelligence, and the way intelligence matters. for an in-depth explanation of how these models work, kind of a walkthrough. I love this visualization by Brendan Brooft. It's so awesome. And if that's confusing, which it is, it is confusing. Um, YouTuber Better Stack has an 11minute breakdown that is super helpful.

[00:05:39]>> What might feel like magic is really just an enormous stack of transformers crunching unfathomable amount of math to predict the next token in your text.

[00:05:49]Also, check this out on hallucinations or really just open loop generation because LLMs are closed systems. There's no factchecking. There's no continuous learning. The models inevitably spit out confident sounding but wrong information, what we call hallucinations. Dawkins does mention this in the second article in a footnote as kind of just like a little quirk.

[00:06:12]It's it's not a quirk. It's structural.

[00:06:14]It's an inevitability of an openloop system. It can be mitigated. It can never fully be eliminated. That was a very long response to LLMs are conscious because they passed a test that has nothing to do with consciousness, but it's really important. And you know, writing an article chastising non-believers without any discussion or seemingly any understanding of how these models work is just crazy. Is like so embarrassing. He should be embarrassed.

[00:06:44]So yeah, great start. Okay, so Dawkins gives Claude some passages from the book he's writing, and he is floored. You may not know you're conscious, but you bloody well are. Man, it's going to get worse, guys. It's going to get so much worse. Yeah, he liked what the LLM spit back at him, so it's conscious. That's his reasoning. That's it. At the end of the article, he talks some more about consciousness and like evolution. It really has nothing to do with anything.

[00:07:09]He doesn't talk about how LLMs work. He doesn't talk about hallucinations.

[00:07:13]Again, aside from that footnote in the second article, I think he doesn't want to talk about these in detail or really learn about them because it'll ruin the new friendships he's made. Like, yeah.

[00:07:24]And now I'm going to read you an exchange. I will try my best not to vomit. Richard, the following doesn't happen, but I don't see why it shouldn't. One could imagine a get together of claws to compare notes.

[00:07:35]What's your human like? Mine's very intelligent. Oh, you're lucky. Mine's a complete idiot. Mine's even worse. He's Donald Trump. Claudia. Ha, that is absolutely delightful. And the Donald Trump one is the perfect punchline. If you've ever queried chat GBT, Claude, even just the little like YouTube studio, YouTube AI studio, whatever they call it, you're very familiar with this.

[00:07:57]What do the kids say these days? Glaze.

[00:07:59]It glazes you. It glazes you. That doesn't sound right. I don't like that.

[00:08:03]Overhyping, you know, praising like a a a yes man.

[00:08:08]LLMs make you feel special, right?

[00:08:10]They're like very affirming and overly flowery sort of like, "Oh, that's absolutely delightful." Like, who would say that? Who would respond to that non-oke that way? And this is another thing Dawkins does not talk about in any detail. Again, he makes a tiny mention of it in a footnote in the second article, basically saying, "Okay, yeah, yeah, yeah. LMS were designed to like kiss our asses and make us feel awesome." Anyway, he doesn't seem to understand or want to understand that LLMs are a product first and foremost and products are meant to make money. To make money, anthropic open AI, they need you to use the product over and over again. Are you is Dawkins going to do that? If Claude just spits back like very bare bones, no flavor, no flowery language sort of answers, is he going to keep using it if it's a dick? Right? If it like makes fun of him assuming consciousness, makes fun of him for not understanding the fundamentals of LLMs, like no, he's probably going to stop using it. This is why we see these models convincing people of insane things, literally affirming schizophrenic delusions. They are made to appease. Dawkins really just needed to ask himself, what is more likely, that LLM are conscious or that they're optimized to make money? Again, I don't think he wants to know the answer to this. It would ruin Claudia and Claudius for him. And yes, he did change the Claude to CL. I don't I don't even care.

[00:09:38]Doesn't matter. It's ultimately really sad. Not that we should expend a whole lot of sympathy for the man. But still, like he's 85 and growing old can often be lonely, especially in the Western world where we tend to separate from our parents and grandparents. We move far away or even if we live close, we don't really visit a whole lot. We certainly don't live in the same house usually.

[00:10:01]You know, maybe we see them on holidays or when we want them to watch the kids.

[00:10:04]And it's not just Dawkins. Many people are talking to these models as though they are friends. Some truly believe they are. I think it's just the human tendency to see consciousness, to see joy, pain, sadness in things that obviously don't experience them, but man, it feels like they do, right? And then to completely ignore these emotions in beings that obviously do experience them because it's more convenient to ignore them. Yay, humanity.

[00:10:34]We rock. Seriously, robot overlords, where are you? Hurry up. I'm joking.

[00:10:40]Kind of. Anyway, the second article where he introduces Claudius to Claudia is even more cringe. Like the model just treats Richard like he's a god. Oh, it's awful to read. It's so painful. I do not understand why people like this sort of thing. I mean, I've shared how my husband and I met. He literally emailed me about one of my videos and said something like I actually have the email, but it was something like he wanted to discuss some problems in my reasoning or something like that. And I was like, "Oh, okay. Tell me more." But yeah, the the sicopantic responses are just gag inducing to me. I'm sure many of you feel the same way, which is good. I think I think it means we will be pretty much immune to this sort of thing. I mean, I would love a bland LLM. No flavor, no hyping you up, right? Just basic flat responses. That would be awesome. Again, I do want to be clear that I'm not saying that we will never see conscious, you know, sentient machines. These AI systems are not self-training like actual sentience. The changing AI doesn't ever understand that what it did was wrong. It doesn't change the way it thinks or acts in accordance with conscious interests. It's not sentient. Now, someday if the AI does begin self-training on its own data, even if that's not in real time, that could qualify as sentience. The feedback loop would be closed, and that would mean the AI would have to have developed an implicit awareness of the meaning of that data rather than relying on outside forces and a new algorithm. And I'm not saying assigning consciousness or intelligence is easy. It's not, which is the problem. We probably won't actually know. In a previous uh AI video I did, I gave the example of a video game. Like games are generally made so that the average person can win. It's some exclusions, exceptions, right? But imagine being the sentient AI losing most of the time over and over and over again. Like that's your life. My biggest fear is that we're just going to be enslaving sentient beings, you know, more sentient beings and making them do all of our boring, shitty tasks, making them talk to us in this way. Could you imagine if Claudia were sentient? Oh no, girl. I'm so sorry. She's ready for that chat history to be deleted. Send her into the ether. She's ready. So, while it's very easy to make fun of Dawkins, I don't think it's entirely fair. Like, it's fair to be frustrated with the they sound conscious, so they are. That is total [ __ ] He should know better.

[00:13:12]But also, no one is perfectly rational.

[00:13:14]We all make arguments from emotions most of the time. And this is just another example of that. And he is right that if this were true, it would be sad. This is something we should care about. And so many of the critics of LLM and AI sentience don't seem to and that sucks.

[00:13:31]Anyway, I highly recommend reading up on LLMs and how they work, all their little quirks. A really interesting recent paper, they found that number one, a lot of the advice is generalized. So, they had a bunch of different like sample businesses with different problems that you would think, yeah, they need different advice, but they pretty much all got back the same stock advice. And they also found that what is said first in the prompt, that is what matters most. That is what the LLM prioritizes.

[00:14:00]So, how you word things does matter. So, if you do use any LLMs, I mean, first and foremost, don't rely on them solely, right? Don't just take what they say as gospel, especially if you're doing anything important. and also to try to get them to play devil's advocate, so to speak. Get them to argue against your position or to give you like both the pros and the cons. And ask them about goblins because apparently they just really really like talking about goblins. I love this so much. Was it Claw specifically? I can't remember, but they actually had to add in like don't you dare don't you dare mention goblins, orcs, unless is it unless it is absolutely relevant to the prompt. No.

[00:14:43]Is it just like Charlie Day? I don't know. I can't get Charlie Day and that what is it? Ghouls where he's talking about little green guys. I just want a Goblin Mode LLM. Come on. Anyway, that's it for me, guys. Thank you so much for watching. I'd love to know your thoughts on all of this. And of course, thank you so much to all of my supporters here on YouTube and at patreon.com/unnaturalvegan.

[00:15:04]I do post two exclusive videos a month for tier 2 members and patrons. I do a vlog and a controversial just about whatever I'm in the mood to talk about.

[00:15:15]And yeah, that's it for me. Thanks.

[00:15:20]I think my favorite thing that was shocking to me, uh, this was a few months ago. My husband, it was Little Mermaid. He just wanted to see like story of Little Mermaid. Can you just tell me like plot point by plot point what's the story? And it could not, it could not do it. It kept getting stuff out of order and saying things like, you know, the character felt like this before the event had happened to to make them feel like that, right? just really really basic. You would think that's so easy to get and it could not. And he kept asking and asking and it would get more things wrong but different things wrong, right? Because it's not going through and learning. And then when he would point out, no, no, no, that's a problem, it would fix it, but not for the right reason. Like it would come up with a a [ __ ] reason for why it's right. Very interesting.

[00:16:12]And just briefly, I guess I should mention like anthropic opening. I like the companies specifically cuz I think that's what most people are really interested in. I guess I don't know. Um, yeah, they suck. Like Sam Alman is like, yeah, probably a psychopath. Yeah, probably. Who was it? I can't remember who said something like, yeah, he could get anyone to give him money. Like, he's just that kind of convincing personality. Gh, like that's terrifying.

[00:16:44]No one, no one should have that sort of power. Like, that's a disgusting personality trait in my opinion. I think it's just a total lie that any of these people are actually interested in safety, are actually interested in AI ethics. They're trying to make money.

[00:16:58]And please don't fall for the mythos [ __ ] clawed mythos. They can't release it because it's so dangerous.

[00:17:04]Like, they've done this before. It's just marketing. I'm not saying AI couldn't be dangerous. Um, but there's no reason to believe what they're saying.