Stop treating AI like unexplainable magic

AI and ML will always stay within the bounds of their programming, said Innovate UK's Zoe Webster

"AI and ML are not magic," Zoe Webster of Innovate UK told delegates at Computing's AI & Machine Learning Live event this week, even though people - even technologists - often talk about them in this way.

"Quite often I go to conferences and events, and people are talking about AI and machine learning as being these things that are using magic - just completely not understandable, almost out of our control," she continued.

But artificial intelligence is not real intelligence, said Webster, who is Innovate UK's Director of AI and Data Economy - it will never go outside the bounds of its own programming.

"We're seeing what they do and marvelling at what they come out with, [but] this is all algorithms. There is a step-by-step process that these machines follow; they don't go beyond that, they don't go outside that. They are programmed. It may be highly concurrent and highly complex, making it very difficult to really know what's going on in all cases, but it's not magic, it's an algorithm and I'm keen that we remember that. It ties back into the responsibility and accountability of who's putting that algorithm together."

The theme of transparency and accountability would be a recurring one throughout the day, with many speakers and delegates raising the issue of so-called black box AI. Webster said that, while this is almost synonymous with machine learning now, it is not actually a hallmark of the technology:

"Not all techniques are black box. I think most of the investment in the last 20 years has been in deep learning neural networks, which are sub-symbolic. So if you're trying to recognise a car, for instance, in a scene: you're not looking at the car as a whole or features you'd recognise, you're looking at, basically, pixel values, and those are highly distributed. You're not going to really easily pick out what a model is learning or what the network has come across. That's what makes it black box in a way."

People are now looking at how to make AI and ML more explainable, but the issue of under-investment in the "less sexy" but more transparent approaches will haunt the sector for some time to come.