Google has hung up its AlphaGo gloves after trouncing the world's best human Go players, but when will AI beat humans at other tasks, such as writing a best-selling novel or doing surgery?
To answer that question, a team of researchers led by Katja Grace of Oxford University's Future of Humanity Institute surveyed several hundred machine-learning experts to get their educated guess. The researchers used the responses to calculate the median number of years it would take for AI to reach key milestones in human capabilities.
Teachers may need to be on the alert for machine-written essays by 2026 and truck drivers could be made redundant by 2027, according to the results.
Meanwhile, AI will surpass human capabilities in retail by 2031. The experts also predict that AI will be capable of writing a best-seller by 2049, and doing a surgeon's work by 2053.
Overall, the respondents believe there is a 50 percent chance that AI beats humans at all tasks in 45 years and will automate all human jobs within 120 years.
The researchers invited the views of all 1,634 authors of papers published in 2015 at two of the leading machine-learning conferences, Neural Information Processing Systems and the International Conference on Machine Learning. A total of 352 researchers responded.
Interestingly, the researchers predict that AI won't beat the best human Go players until about 2028. As we know, Google beat Korean Go champ, Lee Sedol, in 2016, and just beat Chinese grandmaster Ke Ji. Google is now putting its AlphaGo developers from its DeepMind lab to work on solving bigger challenges to society.
But as Grace et al point out in the paper, the machine-learning experts were asked when AI could beat a human at Go on the condition that opponents had played or been trained on the same number of games.
"For reference, DeepMind's AlphaGo has probably played a hundred million games of self-play, while Lee Sedol has probably played 50,000," they note.
In fact, if the researchers' predictions are right, we're likely to see a two-legged robot beat humans in a 5km road race before AI beats a human Go player on equal terms.
The survey also asked the researchers about the likelihood of an AI "intelligence explosion", or the point at which AI becomes better than humans at AI design. As physicist Stephen Hawking explained, if that situation occurs, it could result in "machines whose intelligence exceeds ours by more than ours exceeds that of snails".
Specifically, researchers were asked about the chances of an intelligence explosion happening within two years of machines having learned to do every task better and more cheaply than humans. That is, within about 45 years.
Respondents overall see it as "possible but improbable", with a median probability of 10 percent. They also see it as likely to have positive outcomes but there is a five percent chance of an "extremely bad" outcome, like human extinction.