Google’s Schmidt: 5 Years Before A.I. Is More Than Advisory Tool

Xconomy Boston — 

Former Google CEO Eric Schmidt says the next five years of artificial intelligence and machine learning will be more like Apple’s Siri and Amazon’s Alexa than a self-driving sedan or a robo-surgeon taking out your appendix.

“There are still significant errors, as you know, in machine learning systems that we tolerate. They are advisory,” Schmidt says, speaking at a Boston-area conference. “You wouldn’t want to have a pure machine learning system in a life-critical situation, like flying an airplane or whatever. But you’d want it to be an advisor to humans.”

He says A.I. will stay that way “for the next five or so years—until we can solve this accuracy problem” and other issues related to its “breakability.”

“When do they break? When do they fail? Can they be attacked? Until then these systems are best used as humans with advice,” Schmidt says.

Schmidt, currently a technical advisor to Alphabet, Google’s parent company, spoke Thursday in Cambridge, MA, at a Massachusetts Institute of Technology event on the future of work in a world with A.I. and robotics threatening to upend human labor in many sectors of the economy.

Schmidt says Google took its longtime mobile-first rule in product design and applied the concept to A.I. and machine learning, which he acknowledges will be a fundamental transformative force for technology, much like mobile phones and the internet were. Adding A.I. to each product has proved a straightforward task, given feedback data created by software that machine learning systems can use as “training data.”

He thinks MIT’s expansion of A.I. studies—through the recent $1 billion plan to lead in data analytics across disciplines—will slowly but surely edge its way into political science and other sectors.

“People have an enormous sense of their analytical judgment,” Schmidt says.

“What we know from machine learning is there are subtle patterns that computers can detect in the training data that we just don’t see,” he says. “It’s really a math problem: that these systems can go backward in time and backward in time, deeper and more subtly than we can.”