Artificial intelligence is going to boost human productivity in a thousand ways, transforming everything from transportation to health care to agriculture. Some enthusiastic computer scientists even think we will find a “master algorithm” that will fix our politics and make lives “longer, happier and more productive.” In the grandest of these visions, smart computing machines could automate all of scientific discovery.
But many scientists think such promises are overblown, and even a little dangerous, naively creating false confidence in highly fallible technologies. And quite a few researchers now applying AI — in physics, biology, chemistry and finance — think machines will continue to depend on human intelligence for a very long time. They see AI’s greatest potential not in replacing humans, but in enhancing their capabilities, enabling people to achieve things no one has before.
A pair of economists have some suggestions that could help us navigate the risk that AI will cause mass unemployment and social chaos. New technologies could create as many jobs as they destroy, if we pursue them in the right way.
Mathematics alone sets some limits on the potential usefulness of artificial intelligence. For example, physicists Hykel Hosni and Angelo Vulpiani explored the ability of computers using mass amounts of data to improve predictions in fields such as finance, medicine, cybersecurity or even politics. The trouble, they argue, is that almost any real-world application of AI will involve a huge number of variables. Accurately predicting the future of any such system will require astronomical amounts of data, far beyond what is remotely possible to gather. The more complex the system — and that’s just where we think AI might help — the worse it gets.
This doesn’t mean that AI won’t improve predictions, just that it won’t do so without the human factor. Improved forecasts will require new conceptual insights as well as more data. Such has been the case for weather predictions: Scientists learned years ago that using more data in making forecasts often leads to less accuracy. Accurate predictions today require the intentional disregard of lots of data that reflect atmospheric events that don’t actually affect weather.
Researchers have learned much the same for biology and medicine. “Big data,” as one group puts it, “needs big theory too.” In finance as well, sophisticated users of AI find that they get the best results by pairing machine learning with experienced humans. Chemists using AI to discover entirely new reactions see it as a helpful tool, not a thinking replacement. AI training, says the University of Glasgow’s Lee Cronin, ultimately comes from the chemist: “No chemist, no AI.”
Perhaps machines won’t replace humans quite as broadly as many fear. AI is getting better at doing what humans can do. But humans working alongside AI will be able to do what neither humans nor AI can achieve alone.
This is crucial, as economists Daron Acemoglu and Pascual Restrepo recently noted, because we can choose how to develop AI for the future. Most tech companies and businesses have been focusing on replacing people with AI with a narrow eye toward boosting short-term profits. We could choose to focus technology instead on creating new tasks for which humans will be as indispensable as ever.
A teacher using psychological insight and social knowledge in tandem with AI to assess students’ natural skills might be able to craft teaching methods on an individualized basis. Similar possibilities would arise in health care or medicine, where AI could help doctors combine their personal knowledge of a patient with health-care databases and genomics data to tailor treatments for individuals. In manufacturing, humans armed with AI tools could guide industrial robots in complex tasks otherwise beyond their capabilities, or help humans interact with robots more efficiently.
Acemoglu and Restrepo caution that this cooperation won’t happen on its own. Markets tend to reward the kind of development that attracts the highest short-term return, even if avenues pushing in other directions would lead to more beneficial outcomes in the long run. Right now, they suggest, we’re developing the wrong kinds of AI. Unless we change that, we won’t see new kinds of jobs, and can expect a future much like the last few decades, with stagnating productivity and labor demand.