One of the more surprising concepts in the world of AI and Large Language Models is “uncanny valley.” Here’s how I’d explain it to any newbies who might be reading this (feel free to say hi and drop me a line!):
You would think that, as artistic representations of humans become more realistic, people would linearly relate to it more like humans. Like you would think, if you saw a cartoon – you think that’s not a human at all and treat it as such. But if you saw a human-ish image – you’d perceive it to be closer to human and thus treat it more human. Up until the point where you saw something indistinguishable from a human – and if you can’t even differentiate it from a human, you think it’s human, so tautologically, you’d treat it like a human.
But not so! It turns out this is almost true, with the interesting and “uncanny” part coming out in the footnote: the closer to human artistic representations online are to humans, the more and more human we treat it — until the point where it gets to be “almost human but not quite” — when you see images that almost feel like it could be a photo (for example) of a person but there’s something, probably something indescribable, that makes you realize it’s not human, and you feel a disgust towards it. Or if not disgust, another strong, negative emotion towards it. Said differently, humans tend to be turned off strongly and instinctively by something that is very very close to being human but you can feel and you can tell that it just isn’t. That moment is the moment of “Uncanny Valley.”
This concept has interesting implications in the development of digital representations of humanoid characters online. For example, maybe you shouldn’t make your characters TOO realistic, if they’re not yet at the indistinguishable point. You might want to purposefully degrade your artistic drawings of humans online, for example! How about that for a surprising conclusion!
So! With that as context, there’s an interesting and hopefully surprising application of uncanny valley (although maybe not so surprising to you since I’ve been building up to it here, or at least trying to): there’s an uncanny valley in talking to LLMs like ChatGPT.
Let me explain. When you first talk to a GPT or LLM, it’s a bit amazing, it’s like you’re talking to a person. What a wonderful high! What magical, literally magical, technology we have been blessed with! Or perhaps cursed!
But as you talk more and more to it, and as the LLM and GPT technologies become better and better, an interesting thing happens: it’s almost like talking to a human–but almost. Not quite. Not exactly. It’s like you’re talking to a person, but a bit awkward. It’s always wordier than a human would be, or more logical than a human would be, or more formulaic-sounding than a human would be, or missing out on subtle cues as compared to humans, or so forth. It’s so close to talking to a human… but it’s missing that, “something” to use the technical word.
And this creates an interesting challenge for anyone wants to build a “bot” to be chatted with for emotional or psychological reasons: it being just slightly off in a way turns you off from wanting to talk to it. Much more than if it were a simpler robot old-school bot, which we don’t really mind, because we expect it to be a robotic bot and nothing more. But here, it’s so close to being human, that when it fails and core human-ness, then it just generates these emotions to turn us off.
The solution? The only solution is to wait as the technology improves and overcomes this–and in the meantime, to perhaps keep the technology slightly less human than we could. But the “waiting” may only be a matter of weeks, months, or just a few years, because the light-speed at which this technology is developing, it’s hard to predict how good it will be very soon. Or at least how good what will be released to the public will be, because those who know what’s to come, they now what’s to come.
