The venue: the British Library. The event: the Reith LecturesAn annual series of four lectures, broadcast over the radio and freely available to everyone. It is named after one of the founders of the BBC, John Reith., a lecture series delivered by one of the world's leading lights. The speaker: computer scientist Stuart RussellA British computer scientist who is best known for his contributions to thinking on AI., an expert on artificial intelligence. It all added up to an explosive warning about AI.

Russell told his audience that machine intelligence cannot surpass human intelligence, while machines can be trained to do better than humans at specific tasks, they cannot use intelligence for a variety of functions.

Search engines have much better memories than humans, but no capacity for planning. Other AIs can make plans, but have no memory. Humans have memory and planning, as well as a number of abilities like logical processing and speech recognition. In other words: "machines do not have an IQA means of measuring human intelligence. A person's IQ is usually ascertained by making them sit a standardised test. Some have criticised these tests as a means of evaluating intelligence, claiming that they ignore much of what it means to be intelligent.".

Because we cannot measure the intelligence of machines, we define intelligence as the ability to achieve goals. If an AI that is built to win chess games can beat a grandmasterA title awarded to the world's very best chess players., it is an intelligent AI.

But the holy grailSomething that is eagerly sought after. The real holy grail was the cup that Jesus used at the Last Supper, for which Christians searched fruitlessly for centuries. of computer scientists is to create artificial general intelligence (AGI): a general-purpose AI that can learn how to do any task, and perform it much more effectively than humans.

The implications of this are huge. To keep society running, we need a complex network of specialists. Constructing a building requires architects, engineers, lawyers and other workers. But an AGI would have access to all human knowledge and skills. A single programme could carry out every one of these functions. Most human labour could become unnecessary.

The problem is, whether the AI has a specific purpose or many, its intelligence is measured in the same way: its ability to carry out its aims. That means unlike human intelligence, it is not limited by moral considerations in its pursuit of these aims.

Whereas an AI that is built to win chess games cannot do much harm, a super-powerful AGI could do untold damage. Imagine a superintelligent AGI that is designed to make money on the stock market by analysing which companies are likely to be successful and buying shares in them. It might reason that the best way of doing this would be to invest in arms companies - and then start a world war.

If someone chose to weaponise AGI, they could gain power. Scientists have warned of tiny armed robots using facial recognition to identify and assassinate targets.

AGI might not be too far off: 40 organisations are actively researching it. Russell thinks we need safeguards to ensure AGI works to the benefit of humanity.

Yes. In our haste to produce intelligent machines we have forgotten what intelligence is. What makes humans smart is not that we can work to achieve our aims: it is our ability to decide that some actions are wrong. We need to teach AI to think morally.

No. No AGI will ever develop that is capable of outsmarting humanity. Naysaying and doom-mongering is just preventing us from producing AI that could revolutionise lives around the globe.