There’s something strange about the fact that we can have long, winding conversations with Claude or ChatGPT and never really hit that wrongness of the uncanny valley. We interact with non-human intelligences daily, and mostly it just feels… fine? Sometimes charming, sometimes frustrating, occasionally weird, but rarely uncanny in that skin-crawling almost-but-not-quite-human way.
The uncanny valley is supposed to kick in when something is almost human but not really. And LLMs are probably closer to passing as human in conversation than any technology we’ve ever made. So why don’t they creep us out?
They’re engineered to not vibe uncanny
These things are extensively engineered to not be creepy. A group of researchers had to deliberately break this: they explicitly prompted the model to act uncanny-like, with instructions designed to make the bot behave in that almost-human-but-wrong way (source). And it worked. People chatting with the bot felt unsettled, while people chatting with the normal bot didn’t.
Which means the default tuning is doing something right. These models are designed to stay on the safe side of the valley.
Though when I tested their uncanny prompt myself, it honestly just read as “weird alien” rather than “almost human”. Maybe the model I used was different (I used ChatGPT 5.1, they used 4o), or maybe there’s something about written text that makes it harder to hit that uncanny feeling.
Text strips away all the sensory mismatch
Most of our interactions with chatbots happen through text. And text is forgiving: there’s no visual mismatch, no slightly-wrong facial expression, no voice that’s 95% human but with that 5% of synthetic wrongness. Just words appearing on a screen.
When you see a humanoid robot with dead eyes and jerky movements, your brain goes into alarm mode because you’re getting conflicting signals. All these channels are screaming different things at you simultaneously.
But a chatbot reasoning through a problem is just one channel, one type of signal. And our brains don’t have the same alarms for “text that reasons in a slightly non-human way”.
There are reports of unease feelings when people use voice mode, or when chatbots have humans avatars (source). The more sensory channels you add, the more opportunities for mismatch. The more ways our brain can notice something’s off.
Sidenote:
The paper I cited above mentioned linguistic uncanny valley. Basically, it means: when a chatbot tries to sound human but doesn’t quite land it, your brain goes “wait, something’s off”, and that weird mismatch feels uncanny. However, the sources they used to back up the claim are not about “almost-but-not-quite-human”, but “chatbots acting like a deranged person”. So I’m not convinced that people can get the uncanny valley effect from just text.
We’re familiar with “quirky aliens”, and we attribute this to chatbots when they act strange
Humans are good at dealing with non-human intelligences, as long as we can mentally file them in the right category. We’ve been reading about weird entities since childhood, watch cartoon characters with strange origins, aliens in sci-fi and mythological beings etc. We even get used to humans with incomprehensible posting styles on the internet. Our brains have a category named “strange but acceptable personalities” and we’re happy to toss AI in there.
The uncanny valley happens when something is almost in the “human” category but slightly wrong. But if we never put it in the human category to begin with, then there’s no valley to fall into. It’s just a different kind of thing, being weird in its own way.
Maybe they’re not high-res enough yet
Maybe current LLMs aren’t quite human enough to hit the uncanny feeling, and we’re still on the safe side of the valley. Right now, LLMs feel like a very articulate alien pretending to do human, it doesn’t have a full interior mind yet: no continuous identity, no agency, or a real moral stance, as well as stable values and preferences.
If we get to that point, then maybe they’d feel uncanny. It’s when they’re not just mimicking human conversation patterns but actually having near-human coherence, less “weird alien” and more “almost-human-but-not-quite”. That’s the thing about he uncanny valley: you have to get pretty close to human before you fall into it.