Beware the hyperbole surrounding artificial intelligence and how far it has progressed.

Great claims are being made for artificial intelligence, or AI, these days.

Amazon's Alexa, Google's assistant, Apple's Siri, Microsoft's Cortana: these are all cited as examples of AI. Yet speech recognition is hardly new: we have seen steady improvements in commercial software like Dragon for 20 years.

Recently we have seen a series of claims that AI, with new breakthroughs like "deep learning", could displace 2 million or more Australian workers from their jobs by 2030.

Similar claims have been made before.

I was fortunate to discuss AI with a philosopher, Julius Kovesi, in the 1970s as I led the team that eventually developed sheep-shearing robots. With great insight, he argued that robots, in essence, were built on similar principles to common toilet cisterns and were nothing more than simple automatons.

"Show me a robot that deliberately tells you a lie to manipulate your behaviour, and then I will accept you have artificial intelligence!" he exclaimed.

That's the last thing we wanted in a sheep-shearing robot, of course.

To understand future prospects, it's helpful to see AI as just another way of programming digital computers. That's all it is, for the time being.

We have been learning to live with computers for many decades. Gradually, we are all becoming more dependent on them and they are getting easier to use. Smartphones are a good example.

Our jobs have changed as a result, and will continue to change.

Smartphones can also disrupt sleep and social lives, but so can many other things too. Therefore, claims that we are now at "a convergence" where AI is going to fundamentally change everything are hard to accept.

We have seen several surges in AI hyperbole. In the 1960s, machine translation of natural language was "just two or three years away". And we still have a long way to go with that one. In the late 1970s and early 1980s, many believed forecasts that 95 per cent of factory jobs would be eliminated by the mid-1990s. And we still have a long way to go with that one too. The "dot com, dot gone" boom of 2001 saw another surge. Disappointment followed each time as claims faded in the light of reality. And it will happen again.

Self-driving cars will soon be on our streets, thanks to decades of painstaking advances in sensor technology, computer hardware and software engineering. They will drive rather slowly at first, but will steadily improve with time. You can call this AI if you like, but it does not change anything fundamental.

The real casualty in all this hysteria is our appreciation of human intelligences ... plural. For artificial intelligence has only replicated performances like masterful game playing and mathematical theorem proving, or even legal and medical deduction. These are performances we associate with intelligent people.

Consider performances easily mastered by people we think of as the least intelligent, like figuring out what is and is not safe to sit on, or telling jokes. Cognitive scientists are still struggling to comprehend how we could begin to replicate these performances.

Even animal intelligence defies us, as we realised when MIT scientists perfected an artificial dog's nose sensitive enough to detect TNT vapour from buried landmines. When tested in a real minefield, this device detected TNT everywhere and the readings appeared to be unrelated to the actual locations of the mines. Yet trained mine detection dogs could locate the mines in a matter of minutes.

To appreciate this in a more familiar setting, imagine a party in a crowded room. One person lights up a cigarette and, to avoid being ostracised, keeps it hidden in an ashtray under a chair. Everyone in the room soon smells the cigarette smoke but no one can sense where it's coming from. Yet a trained dog would find it in seconds.

There is speculation that quantum computers might one day provide a real breakthrough in AI. At the moment, however, experiments with quantum computers are at much the same stage as Alan Turing was when he started tinkering with relays in the 1920s. There's still a long way to go before we will know whether these machines will tell deliberate lies.

In the meantime it might be worth asking whether the current surge of interest in AI is being promoted by companies like Google and Facebook in a deliberate attempt to seduce investors. Then again, it might just be another instance of self-deception group-think.

James Trevelyan is emeritus professor in the School of Mechanical and Chemical Engineering at the University of Western Australia.

Excerpt from:

When robots learn to lie, then we worry about AI - The Australian Financial Review