When do ML Researchers Think Specific Tasks will be Automated?

By Katja Grace, 26 September 2017

We asked the ML researchers in our survey when they thought 32 narrow, relatively well defined tasks would be feasible for AI. Eighteen of them were included in our paper earlier, but the other fourteen results are among some new stuff we just put up on the survey page.

While the researchers we talked to don’t expect anything like human-level AI for a long time, they do expect a lot of specific tasks will be open to automation soon. Of the 32 tasks we asked about, either 16 or 28 of them were considered more likely than not within ten years by the median respondent (depending on how the question was framed).

And some of these would be pretty revolutionary, at an ordinary ‘turn an industry on its head’ level, rather than a ‘world gets taken over by renegade robots’ level. You have probably heard that the transport industry is in for some disruption. And phone banking, translation and answering simple questions have already been on their way out. But also forecast soon: the near-obsoletion of musicians.

The task rated easiest was human-level Angry Birds playing, with a 90% chance of happening within six or ten years, depending on the question framing. The annual Angry Birds Man vs. Machine Challenge did just happen, but the results are yet to be announced.

The four tasks that were not expected within ten years regardless of question framing were translating a new language using something like a Rosetta Stone, selecting and proving publishable mathematical theorems, doing well in the Putnam math contest, and writing a New York Times bestselling story.

The fact that the respondents gave radically different answers to other questions depending on framing suggests to us that their guesses are not super reliable. Nonetheless, we expect they are better than nothing, and that they are a good place to start if we want to debate what will happen.

To that end, below is a timeline (full screen version here) showing the researchers’ estimates for all 32 questions. These estimates are using the question framing that yielded slightly earlier results – forecasts were somewhat later given a different framing of the question.