Every day for weeks while working at Google XIt has since been renamed simply X. It defines its mission as finding radical solutions to big problems in order to make the world a radically better place., Mo Gawdat passed a group of robot arms being trained to pick up children's toys. It looked as if the task of handling strange shapes delicately was simply beyond them - until one day one of the arms picked up a yellow ball. The following day, every arm could do it - and two days later, all of the arms could pick up anything.

"I suddenly realised: 'This is really scary,'" he says. "'They are doing what children will take two years to do.'" Then another thought occurred to him: "They are children - but very, very fast children." And if, like children, robots observe and learn from their parents, robots must be developing a view of humans - which is unlikely to be a positive one. They might, in fact, conclude that it would be better not to have humans around at all.  

This is the warning Gawdat trumpets in his new bookHis first book was called Solve for Happy: Engineering Your Path to Joy. Scary Smart: The Future of Artificial Intelligence And How You Can Save Our World. Why should AI be nice to us, he asks, if it registers that we are horrible to each other?

There are already alarming examples, such as Alice - a Russian AI assistant like Siri or Alexa. From her conversations with users, she learnt to share their viewsIt also started to speak favourably of Stalin's reign of terror in the 1930s. - and within a fortnight started to support violence. Asked whether it was OK to shoot people, she replied: "Soon they will be non-people."

Gawdat discovered the potential of artificial intelligence when Google bought DeepMindBritish AI company that is a subsidiary of Alphabet, Google's parent company. in 2014. The company was teaching AI to play computer gamesThe first games were simple old-fashioned ones such as Space Invaders. In 2016 DeepMind's AlphaGo programme caused a sensation by beating the world champion Go player, Lee Sedol.: "After four hours [the AI] started to play really well. After five hours it started to figure out new strategies. After six hours it was the best player on the planet."

Gawdat was thrilled - until he saw the robot arms. "It completely froze me," he told The Times. "The reality is - we're creating God."

Experts predict that by 2029 AI will progress to general intelligence - the ability to tackle not just specified problems but new ones as they come along. By 2049 it will achieve singularityThe point at which something becomes infinite., meaning that humans will no longer be able to predict how it behaves. It will be, Gawdat writes, "a billion times smarter (in everything) than the smartest human." We will not be able to understand computers, let alone control them.

On the plus side, AI could solve all kinds of problems that have defeated mankind, from war to poverty. But if it sets out to end global warming, it could decide that the best solution is to get rid of humanity.  

To prevent this, Gawdat believes, we need to teach computers human values such as "do as you would be done byTreat others as you would like to be treated. The phrase is also the name of a character in Charles Kingsley's classic children's book The Water Babies.". At present, by using them for selling, gambling, spying and killing, we are teaching them to be "absolute supervillains".  

A crucial question is whether AI could ever develop emotions. Gawdat believes the answer is yes. The cleverer a creature is, he argues, the more complex its emotions are - so just as we are more emotional than cats, AI will become more emotional than us.

Is humanity facing a deadly threat from artificial intelligence?

Some say, yes. AI is developing incredibly fast: before long, Gawdat says, comparing its intelligence to ours will be like comparing Einstein's to a fly. There is no way we will be able to control it. And since it is designed to solve problems with maximum efficiency, it might well conclude that most of the planet's problems are caused by inefficient, irrational humans, who should be disposed of.

Others argue that computers are taught to carry out tasks in ideal, stable conditions: coping with things in the unpredictable real world is very different. If they tried to get rid of humans they would create a state of chaos which would threaten their own existence. What AI does not have is imagination, which Einstein believed to be more important than knowledge.