How could we know if an artificial intelligence is really intelligent?

AI expert Beth Singler was on hand to fill us in...

Beth - Yeah. So lots of people have tried to come up with ways to show or describe, or explain intelligence and specifically artificial intelligence. And we've mostly gone down the line of thinking; there are ways to test for intelligence, and that really tells us more about what we think intelligence is. And actually, whether we'll be able to prove intelligence in an artificial entity or machine. I'm really quite fond of a quote from someone called Robert Wilensky, who was a computer scientist working in AI at the very beginning. And he says that very early scientists working on AI were mathematicians, and they looked around and they said, well, we're smart. So if an artificial intelligence is going to be smart, it's going to be able to do the things we can do. And as mathematicians, they could basically prove theorems and play chess. So these same sorts of ideas are now constantly mapped onto what we think AI is going to be able to do to be smart. Whereas I think there might be something interesting in thinking about how an AI might work against our assumptions and programming and be able to do things that are unexpected and unexplained, but also those could theoretically be programmed into it. So it's all very complicated, but I think it does tell us something very, very profound about why we think intelligence is measurable by being able to play chess really well, or Go very well or prove a theorem.

Phil - So are there some under appreciated aspects of quote unquote intelligence, that you think people making these AIs needs to pay more attention to or are starting to pay more attention to?

Beth - Well, we are very aware that intelligence is embodied. Scientists look at cognition through embodiment and as an anthropologist, as a social scientist, I see how an intelligence is a relational thing that we have in community through our human bodies. So increasingly the speculations about how we develop actual humanlike intelligence in machines, would have to require some sort of learning process within an embodied sensory system, and there's work going in that direction. But to simply say, you're intelligent, if you can play chess very well, that that would make me a very not intelligent person. And I hope I am a relatively intelligent person, but I cannot play chess. So there are levels and standards that we have set for intelligence, for our machines and artificial intelligence, but we need to think about how it works in the whole.