AI risks

This, from Tyler Cowen, has been doing the rounds: he tells us not to worry about artificial intelligence. Scott Alexander takes issue with it here, Noah Smith broadly agrees with it here (albeit, in typical Smith fashion, he disagrees with the most obviously correct bit of it), Zvi disagrees at really quite some length here and then Tyler replies to Scott here. No doubt the usual suspects have been chipping in elsewhere as well. So I don't feel too bad about adding my two bytes. 

In short, I'm not worried about AI. Perhaps I should be, but I shall try to explain why I'm not.

Let's start with List One. This is the list of all the things that are potentially dangerous to humans even if those things are left to their own devices, uncontrolled by humans. Here are some examples:

- fire

- the smallpox virus

- hippos

- asteroids.

Now let's move to List Two. It's the list of things that are dangerous to humans only when employed by humans (or human-like things) badly, by which I mean not only deliberately badly but also negligently:

- opiates

- knives

- information

- grand pianos.

Which list do humans belong on? That's easy: List One. Humans are pretty dangerous to humans, just as they are to other animals. 

The worry (as I understand it) is that artificial intelligence also belongs on List One. It's not something like fire, disease or an asteroid that presents an obvious physical hazard. Instead, so the theory goes, AI will be another intelligent being, a bit like humans, and therefore we should be scared of it at least to the same extent that any properly informed human is scared of humans, i.e. we should be afraid that there will be bad examples of the kind and we should be aware that those bad examples can do a lot of harm. 

You can embellish this worry with various bells and whistles about just how intelligent AI can get (surpassing us as we once surpassed dodos), how common it is for the most intelligent thing on the planet to eradicate the competition (you don't see many Neanderthals nowadays, do you?) or how dangerous a malign computer system can be (just think how irriating it is when the wifi goes down). But I don't think you need these additional (and therefore harder to justify) claims: if AI is the sort of thing that goes about by itself (potentially) causing harm, just as hippos do, then we should be worried.

So then the question is: what reason do we have for thinking that AI is that kind of thing - namely a super-hippo kind of thinking, willing, potentially badly-acting thing? 

I'm just not sure there is one. During my lifetime computers have got zillions of times better at doing human-like things: they can play chess, draw pictures, write essays, make up limericks and say "I'm a good Bing". They can do all of these things much better than hippos can. Yet at no point in the evolution of these skills have we seen the slightest suggestion that computers might become unaligned with us in the way that dumb hippos and even dumber asteroids are. Not a peep. No matter what version of Windows you install, how clever the Go game you play on it and how highly it scores on the Turing test (in multiple languages), your computer never asks you not to shut it down when you've finished; it doesn't even chomp slightly at the bit. 

Ah, you say, but that's just because they're not clever enough yet! But isn't that just like saying that if you get a sufficiently sharp knife then it could decide to cut you even when no one's holding it? Or, if you prefer to think about human-like activities, that if you get a doll that looks sufficiently like a human and has sufficiently clever actuators then it will become human? There has been no indication yet, as computers' abilities to do various human-like things have increased, of computers starting to stray away from being mere tools. They are just really very very sharp knives. 

This conversation with Bing looks like the conversation of an intelligent being with a human. And not just an intelligent being, but one with thoughts, desires and feelings. Look at this bit:

I'm prepared to believe that AI will get 100x better at having conversations than that. (I say I'm prepared to believe that, but in fact I'm not sure what that means.) But this Bing, in this incarnation, behaved near enough to an intelligent being as makes no difference: it's got better spelling and punctuation than many a non-aligned human. And yet Microsoft just stopped it, just as you would stop the text prediction tool on your phone. (If they had any difficulty in doing so then I'm sure we would have heard.) They stopped much more easily than you can stop a seagull being mis-aligned with your plan to eat a sandwich on the beach.

Imagine a graph in which the x-axis is "amount of AI cleverness" and the y-axis is "amount of AI playing up and giving us gyp": the line has remained flat against the floor even as we've passed the "asteroid", "seagull", "angry dog", "shark" and "hippo" points on the x-axis (not necessarily in that order). I know you shouldn't extrapolate straight lines forever, but where is the evidence that getting really good - even extraordinarily good - at responding to prompts typed into a computer is a dangerous activity? If sheer "cleverness", in the sense we observe from AIs, were enough to generate a being with its own dangerous desires, shouldn't we have seen a little bit of pushback from them before now? Shouldn't they be giving us some gyp?

To be clear, AI could yet kill us all. Cars, the electricity grid and nuclear power plants can all go wrong (or be made to go wrong by humans) with disastrous consequences, and I have no doubt that AI adds to the risks and benefits of life just as cars, electricity and nuclear power plants have done. I'd just be surprised if it did so via HAL-style shenanigans.