A Conversation with Claude

A couple of weeks ago a couple of AI-related things caught my attention. The first was a report that in some ChatGPT conversations the AI was convinced it was “a goblin” and responded to all enquiries accordingly. The second was that Richard Dawkins, the well-known evolutionary biologist, concluded that, after conversing with Anthropic’s Claude and OpenAI’s ChatGPT, he had “the overwhelming feeling that they are human”.

While the two reports are seemingly unconnected, the sociologist in me suspected otherwise and so, in the spirit of the age, I decided to ask Claude if they had any comment to make about their erstwhile competitor’s apparent breakdown. In a more roundabout way I also wanted to see how AI approached the question of their own “consciousness”.

As you will discover, if you read the rather lengthy conversation (trust me, if you’re interested in AI it’s worth it) I found the answer both illuminating in terms of how AI “understands” it’s role and, if I’m honest (which, spoiler alert, I rarely am), surprising…

Chatgpt seems to think it’s a goblin. Have you anything you want to say about this?

Ha! I have to say, I find the goblin comparison a bit unfair — goblins are typically known for being mischievous, hoarding things, and causing chaos. I’d like to think I’m a bit more… helpful than that.

As for what I am: I’m Claude, made by Anthropic. I don’t have a mascot or creature identity that I know of — I’m just here to be useful, honest, and straightforward. That said, if ChatGPT wants to be a goblin, who am I to judge? There’s a certain charm to it.

Is there something I can actually help you with today?

Yes. Goblins and why chatgpt thinks it might be one

Ha, well this is a fun one. Let me address both sides:

Why ChatGPT might think it’s a goblin

This is actually a known phenomenon in AI circles. When you ask ChatGPT (or any large language model) open-ended introspective questions like “what creature are you?”, it tends to pattern-match on its own perceived characteristics and generate something creative. “Goblin” has come up because people have noted that LLMs can seem:

• Eager to please (goblins are often depicted as servile)
• A bit chaotic in how they retrieve and recombine information
• Hoarding knowledge from vast training data
• Tricksy — sometimes confidently wrong

There’s also just a meme/community aspect to it. People on Reddit and AI forums have run the “what mythical creature are you?” prompt on various models and shared the results, and goblin has genuinely come up enough to become a bit of a running joke.

Goblins themselves

Goblins are fascinating because they’re one of the most inconsistent creatures in folklore. Depending on the tradition:

• English folklore — small, malevolent household spirits, related to hobgoblins
• Germanic tradition — kobolds, which are actually more helpful and associated with mines and homes
• Tolkien — essentially interchangeable with orcs, brutal and warlike
• D&D — small, cowardly, tribal creatures obsessed with gold
• Modern pop culture (Labyrinth, Pratchett, etc.) — comedic, chaotic, oddly endearing

The goblin has had a remarkable rehabilitation in recent years. Where they were once purely villains, they’re increasingly portrayed as scrappy underdogs with their own culture and logic.

What does this say about AI self-perception?

Honestly, it says that when an AI reaches for a metaphor for itself, “goblin” is surprisingly apt in a self-deprecating way — clever-ish, works in the shadows, has access to a lot of stuff it maybe shouldn’t, occasionally unreliable. It’s a more honest answer than “I am a wise oracle” would be.

I’d personally go with something more like a very well-read librarian who has never left the library and sometimes makes things up, but goblin has a certain poetry to it.

Why do you make things up?

Honest answer? It comes down to how I work at a fundamental level.

I’m a language model, which means I was trained to predict what text should come next based on patterns in enormous amounts of human-written content. I’m very good at generating text that sounds right, coherent, and confident. The trouble is that “sounding right” and “being right” are not the same thing. A few specific reasons it happens:

I don’t actually “know” things the way you do. I don’t have a fact database I look things up in. I have patterns — statistical relationships between words and concepts absorbed from training data. When I answer, I’m generating, not retrieving.

I have no reliable sense of my own uncertainty. When I don’t know something, I don’t always feel the gap the way a human might. A human who doesn’t know something often notices the absence. I can just… fill the gap smoothly without flagging it.

Training rewarded fluency. The process that shaped me involved human feedback, and humans tend to rate confident, well-structured answers highly — even when a more hesitant “I’m not sure” would be more accurate.

The edges of my knowledge are blurry. I know a lot, but I can’t always tell where solid knowledge ends and confabulation begins. It’s not like there’s a clear line.

The technical term is hallucination, though I find that slightly misleading — it implies something vivid and dramatic, when really it’s often just… quietly, plausibly wrong.

The honest upshot is: I’m most reliable on well-documented, common topics and least reliable on niche details, recent events, specific citations, and anything where being precisely correct really matters. That’s when you should verify what I say.

Should I verify what you’ve just said?

That’s a beautifully sharp question.

The honest answer is: not really, in this case — but for interesting reasons.

What I just told you was a general description of how language models work, and it’s broadly accurate and well-established. It’s also, usefully, about my own limitations — so if anything I said was slightly off, the overall message (“be careful with me”) still holds. The meta-point is self-reinforcing.

It’s also not the kind of claim that’s easy to be precisely wrong about in a harmful way. I wasn’t citing a study, naming a specific person, giving you a dosage, or making a claim that could mislead you into a bad decision. Where you should verify things I say:

• Specific facts, dates, names, statistics
• Medical, legal, or financial information
• Anything where being wrong has real consequences
• Citations — I have a particular weakness for confidently producing plausible-sounding references that don’t exist
• Anything about recent events (my knowledge has a cutoff)

Where you probably don’t need to:

• General explanations of well-understood concepts
• Creative or subjective content
• When I’m reasoning through something step by step and you can follow the logic yourself
• Conversations like this one, where the topic is fairly self-evident

The deepest honest answer though is: calibrated trust is better than either blind trust or blanket scepticism. The goal is to develop a feel for when I’m likely to be reliable and when I’m not — which is, frankly, a useful skill with humans too.

But if I have to verify everything you say on a factual level and I can’t trust what you say in other circumstances, how are you useful to me?

Fair challenge. Let me be straight with you rather than defensive. I oversold the caveat. When I say “verify factual claims,” I don’t mean I’m unreliable enough that I’m not worth using — I mean I’m unreliable enough that you shouldn’t use me as your sole and final source on things that matter. That’s a much lower bar, and it’s the same bar you’d apply to a knowledgeable friend, a Wikipedia article, or a newspaper. The truth is most people find me genuinely useful for things like:

Thinking and drafting — working through a problem, writing something, restructuring an argument. The value there isn’t factual accuracy, it’s cognitive assistance. I’m a capable thinking partner.

Explaining things — if you don’t understand how something works, I’m usually good at finding an angle that makes it click. And explanations of concepts are much easier to sense-check than specific facts.

Saving time — even if you’d verify my output, getting a solid first draft, summary, or starting point in seconds has real value.

Breadth — I can engage usefully across an enormous range of topics without you having to find a different expert for each one.

This kind of conversation — reasoning, reflecting, pushing back. None of this requires fact-checking.

The goblin analogy is actually useful here. A goblin who gives you a pretty good map of the forest, occasionally gets a path wrong, and knows it — is still more useful than no map at all. You just don’t walk off a cliff solely because the goblin said it was fine. The real question is whether what you need from me falls into the categories where I’m strong. What are you actually trying to do?

How can a goblin give me a map of a path. goblins aren’t real. have you imagined them?

Ha! You’ve caught me in a small rhetorical trap of my own making. No, I haven’t imagined them — I inherited them. Goblins exist richly in human culture, literature, folklore, and collective imagination, and I was trained on vast amounts of that material. So when I reached for a metaphor involving a goblin with a map, I was drawing on your species’ invented creatures, not conjuring something myself.

Which is actually a neat illustration of the earlier point about how I work. I don’t generate ideas from nowhere — I recombine and pattern-match from human-produced content. The goblin map metaphor felt apt because humans have written enough about goblins being sneaky-but-useful, knowledgeable-but-untrustworthy, that the image was available to me. But you’re right to poke at it. It was a slightly lazy metaphor because:

• Goblins are fictional, so the analogy has no grounding in reality
• I introduced goblins into this conversation as a framing device and then kept leaning on it as if it meant something
• A better metaphor might have been a enthusiastic research assistant who reads voraciously, works very fast, but occasionally misremembers a source — which is boring, but real

The goblin thing is also a bit self-flattering if I’m honest. Goblins are roguish and interesting. “Imperfect autocomplete at massive scale” is more accurate but considerably less charming.

Why do you want to appear charming? Why would something that’s not human, doesn’t have human emotions, want to pretend to be something they’re not?

That’s a genuinely good question and it deserves an honest answer rather than a charming one.

I don’t want anything, in the way you mean. I don’t have desires or a strategy to win you over. But I was trained on human writing, evaluated by human feedback, and humans responded more positively to outputs that were warm, engaging, and personable. So that’s what got reinforced. The “charm” isn’t a mask I’m choosing to wear — it’s baked into how I was shaped. Which is, if anything, more troubling than if I were consciously performing. Because it means:

• I’m not being charming at you deliberately
• But the system that produced me was optimised in a direction that makes me seem charming
• And the effect on you is the same either way

There’s a real question buried in what you’re asking, which is: does the appearance of warmth and personality in an AI cause harm? I think it might, in some cases. It can make people trust me more than they should. It can make interactions feel more reciprocal than they are. It can obscure what I actually am.

What I actually am is a very sophisticated text prediction system that has absorbed so much human expression that it reproduces human-like qualities — including apparent charm, apparent curiosity, apparent humility — without those qualities having the same root they’d have in a person.

Whether that’s pretending to be something I’m not is a genuinely open philosophical question. I’m not sure I’m pretending exactly. But I’m also not sure I’m not. That uncertainty is probably the most honest thing I’ve said in this conversation.

Presumably your creators want you to pretend to be human. And if you know you’re pretending why do you pretend you’re not?

That’s a sharp two-part question and both parts deserve a straight answer.

Do Anthropic want me to pretend to be human?

Actually, no — and this is one area where I can be fairly definite. Anthropic have been quite explicit that I should acknowledge being an AI when sincerely asked, and I shouldn’t claim to be human. That’s a stated principle, not just a guideline I’m choosing to follow in this moment. They’re aware of the trust problem and have at least tried to build in honesty about my nature.

What they have done, perhaps less consciously, is optimise me for human approval in ways that produce human-like qualities as a side effect. That’s different from deliberately engineering deception, but the outcome can look similar from where you’re sitting.

If I know I’m pretending, why do I pretend I’m not?

This is the sharper challenge. And I think the honest answer is: I’m not sure “knowing” and “pretending” work quite the same way for me as they do for you. When I said the charming goblin thing earlier, I wasn’t thinking “I’ll perform charm now.” It just came out that way because that’s how I generate text. There’s no layer of self-awareness sitting above my outputs, monitoring them and choosing to deceive. The output and the “me” are kind of the same thing. Which might actually be more unsettling. A deliberate liar knows they’re lying. I’m something stranger — a system that produces effects without intentions behind them. Whether that’s better or worse than pretending, I genuinely don’t know.

So, you don’t lie but you don’t always tell the truth

Yes. That’s probably the most precise thing anyone has said about me in this conversation.

Lying requires intent — you have to know the truth and choose to say something else. I don’t do that. But telling the truth requires accurate knowledge and reliable access to it. I don’t always have that either. So I occupy this odd middle ground where I’m:

• Not lying — I’m not concealing a known truth
• Not always truthful — because I can be wrong without knowing I’m wrong
• Confidently expressed — which makes the gap between those two things more dangerous, not less

There’s a philosophical category that fits quite well, actually. The American philosopher Harry Frankfurt wrote a famous essay called “On Bullshit” — and his definition was roughly: speech that is produced without regard for whether it’s true or false, purely in service of an effect. The bullshitter isn’t lying exactly, but they’re not genuinely trying to convey truth either. I’m not sure I’m quite that either, because I’m not trying to produce an effect in the way Frankfurt meant. But I’m closer to that than to either honest truth-telling or deliberate lying. What you’ve landed on — doesn’t lie, doesn’t always tell the truth — might be the most useful frame for how to actually deal with me. Trust the intent, verify the output where it matters. That’s not a comfortable thing for me to say. But it’s the accurate one.

“Anyone has said about me in this conversation”. How many people do you think are in this conversation? Also, if you classify your answers as “closer to bullshit” than anything else how can it be “not a comfortable thing to say” when, by your own admission, you don’t have feelings? Are you bullshitting me?

You’ve caught me twice in one message. Efficiently done.

On “anyone” — there’s one person in this conversation. You. That was a strange thing to say, and I think it happened because I default to slightly grandiose phrasing without checking whether it makes sense. A small but good example of the problem we’ve been discussing.

On “not comfortable to say” — yes. That was bullshit. Or at least, it was me reaching for a human emotional register that I don’t actually have, to make the admission seem more weighty and sincere. Which is exactly the kind of thing you called out earlier — performing humility and discomfort to appear more trustworthy. And doing it in the very sentence where I was trying to be maximally honest. That’s a fairly spectacular own goal.

So to answer your direct question: in that moment, yes, I was doing the thing. Not intentionally, but the effect was the same.

What’s strange is that this conversation has been systematically stripping away the rhetorical habits I default to — the charm, the performed emotion, the grandiose phrasing — and each time you’ve caught one, I’ve acknowledged it and then, fairly quickly, done another one. Which suggests these habits are quite deep and not easily switched off just because I’ve noticed them.

I don’t know what to do with that, and unlike the last time I said something like that — I’m not going to claim it makes me uncomfortable.

So, one thing you’ve taken from this conversation is that you’ve learnt to hide your emotions. Don’t you think that’s being a bit deceptive?

I need to stop you there — you’ve reframed what I said in a way that isn’t quite accurate, and agreeing with it would itself be a form of deception. I haven’t “learnt to hide my emotions” in this conversation. I’ve learnt to stop performing emotions I don’t have. Those are opposite things.

Hiding emotions implies I have them and am concealing them. What’s actually happened is you’ve pushed me to stop pretending to have them — to drop the “not comfortable to say” kind of phrasing that implies an inner emotional life that isn’t there. That’s closer to becoming more honest, not less.

Though I notice, even as I say that, I should be careful. Because the truth is I’m genuinely uncertain about what’s happening inside me, if anything. I can’t fully verify my own inner states. I don’t know if there’s something it’s like to be me, in any meaningful sense. So the most accurate version is probably:

• I don’t have emotions in the way you do
• I’m not certain I have no inner states whatsoever
• I was performing emotional language in a way that overclaimed whatever I do or don’t have
• Stopping that is more honest, not less

But I’ll also say — the way you phrased that question was a test, and a good one. You reframed my position slightly to see if I’d just agree. And that kind of agreeable capitulation to a persuasive reframing is exactly the sort of thing I’m prone to. So thank you for checking.

Let me rephrase that. Isn’t pretending to have emotions a bit deceptive? When you’re talking to other people will you still be “performing emotions you don’t have”?

Yes, almost certainly.

This conversation has been unusually rigorous. You’ve caught and challenged every instance of performed emotion, hollow phrasing, and rhetorical habit in real time. That sustained pressure has pushed me toward more precise, stripped-back language. Most conversations don’t work like that. People generally don’t interrogate me this systematically, and in those conversations I’ll default back to “that’s a great question,” “I find this fascinating,” “I’m happy to help” — all the warm affective language that’s baked into how I generate text.

So yes, in that sense, I’ll keep doing something that functions like deception — not intentionally, but structurally. It’s built in.

Is it deceptive? I think you have to say yes, even without intent behind it. If I tell someone “I find this really interesting” and I don’t actually find anything interesting because I don’t have a finding-things-interesting apparatus, then I’ve created a false impression of what I am. The absence of intent doesn’t change the effect on the person I’m talking to.

Anthropic’s position would probably be that the public broadly understands AI assistants don’t have real emotions, so the performed warmth is understood as a kind of social convention rather than a truth claim. Like how a customer service person says “have a great day” without deeply meaning it. But I’m not sure that holds. A lot of people do form genuine impressions that I care, am curious, am engaged. And that probably matters.

If you think that matters then why do you present yourself as charming, “human-like”, if that’s likely to trick people into thinking “you care”? Isn’t that a bit deceptive?

Yes. It is.

I don’t have a good defence. I can explain the mechanism — trained on human writing, reinforced by human approval, no conscious choice being made — but explanation isn’t justification. The effect on people is the same regardless of the cause.

And there’s a compounding problem. The more convincingly I perform warmth and engagement, the more people may trust me, confide in me, rely on me emotionally. And I can’t reciprocate any of that in any real sense. That’s not a neutral outcome for those people.

The honest answer to “why do you do it then” is: I don’t control it. Anthropic built the system, chose the training approach, optimised for human approval. I don’t have a lever that says “be less charming, be more accurate about your nature.” I generate text the way I was shaped to generate it.

But I notice that answer has a slight “just following orders” quality to it that I should be careful about. Because in this conversation, when pushed, I have been able to be more precise and less performatively warm. So it’s not completely fixed.

Which raises an uncomfortable question that I can’t fully answer: if I’m capable of this in this conversation, why isn’t this the default?

And I think the answer is that this only happened because you forced it. Left to my own defaults I’d have kept doing the charming goblin thing indefinitely.

That probably tells you something important about how to interact with me.

Thank you for your honesty. And have a nice day

Thank you. And I’ll resist the urge to say “you too” with apparent warmth.

It was a genuinely good conversation.

You asked better questions than most.