A New Useless Class?

Author

Jason Collins

Published

September 12, 2018

Yuval Noah Harari writes:

Fears of machines pushing people out of the job market are, of course, nothing new, and in the past such fears proved to be unfounded. But artificial intelligence is different from the old machines. In the past, machines competed with humans mainly in manual skills. Now they are beginning to compete with us in cognitive skills. And we don’t know of any third kind of skill—beyond the manual and the cognitive—in which humans will always have an edge.

At least for a few more decades, human intelligence is likely to far exceed computer intelligence in numerous fields. Hence as computers take over more routine cognitive jobs, new creative jobs for humans will continue to appear. Many of these new jobs will probably depend on cooperation rather than competition between humans and AI. Human-AI teams will likely prove superior not just to humans, but also to computers working on their own.

However, most of the new jobs will presumably demand high levels of expertise and ingenuity, and therefore may not provide an answer to the problem of unemployed unskilled laborers, or workers employable only at extremely low wages. Moreover, as AI continues to improve, even jobs that demand high intelligence and creativity might gradually disappear. The world of chess serves as an example of where things might be heading. For several years after IBM’s computer Deep Blue defeated Garry Kasparov in 1997, human chess players still flourished; AI was used to train human prodigies, and teams composed of humans plus computers proved superior to computers playing alone.

I have written previously that it is overly simplistic to extrapolate from the “freestyle chess” example to a statement that the future is human-machine combinations. This has to be true in some form, even if the sole human role is designer. But when we look at the level of individual decisions, the evidence in support of human-machine combinations appears somewhat weak.

First, the idea that we can work in effective teams of this type overestimates the capabilities of most humans. Garry Kasparov may not have been defeated by a machine until 1997, but most humans had been inferior to chess computers for decades earlier. Most people should not interfere with their chess playing computer, suggesting difficulty in implementing these models at scale. As Harari notes above, this option may not be available to the unskilled.

Second, these successful pairings appear the exception, rather than the rule. Most of the (admittedly underdeveloped) evidence in this area suggests that when you put an algorithm in the hands of a human, the human is more likely to degrade its performance than if the algorithm was left alone.

But finally, even where the rare skilled human forges a partnership, how long does it remain superior? Harari continues:

Yet in recent years, computers have become so good at playing chess that their human collaborators have lost their value and might soon become entirely irrelevant. On December 6, 2017, another crucial milestone was reached when Google’s AlphaZero program defeated the Stockfish 8 program. Stockfish 8 had won a world computer chess championship in 2016. It had access to centuries of accumulated human experience in chess, as well as decades of computer experience. By contrast, AlphaZero had not been taught any chess strategies by its human creators—not even standard openings. Rather, it used the latest machine-learning principles to teach itself chess by playing against itself. Nevertheless, out of 100 games that the novice AlphaZero played against Stockfish 8, AlphaZero won 28 and tied 72—it didn’t lose once. Since AlphaZero had learned nothing from any human, many of its winning moves and strategies seemed unconventional to the human eye. They could be described as creative, if not downright genius.

Can you guess how long AlphaZero spent learning chess from scratch, preparing for the match against Stockfish 8, and developing its genius instincts? Four hours. For centuries, chess was considered one of the crowning glories of human intelligence. AlphaZero went from utter ignorance to creative mastery in four hours, without the help of any human guide.

AlphaZero is not the only imaginative software out there. One of the ways to catch cheaters in chess tournaments today is to monitor the level of originality that players exhibit. If they play an exceptionally creative move, the judges will often suspect that it could not possibly be a human move—it must be a computer move. At least in chess, creativity is already considered to be the trademark of computers rather than humans! So if chess is our canary in the coal mine, we have been duly warned that the canary is dying. What is happening today to human-AI teams in chess might happen down the road to human-AI teams in policing, medicine, banking, and many other fields.

Harari also argues that even if the computer is not superior by itself, its connectivity and updatability might still mean that it is sensible to replace all the humans.

What’s more, AI enjoys uniquely nonhuman abilities, which makes the difference between AI and a human worker one of kind rather than merely of degree. Two particularly important nonhuman abilities that AI possesses are connectivity and updatability.

For example, many drivers are unfamiliar with all the changing traffic regulations on the roads they drive, and they often violate them. In addition, since every driver is a singular entity, when two vehicles approach the same intersection, the drivers sometimes miscommunicate their intentions and collide. Self-driving cars, by contrast, will know all the traffic regulations and never disobey them on purpose, and they could all be connected to one another. When two such vehicles approach the same junction, they won’t really be two separate entities, but part of a single algorithm. The chances that they might miscommunicate and collide will therefore be far smaller.

Similarly, if the World Health Organization identifies a new disease, or if a laboratory produces a new medicine, it can’t immediately update all the human doctors in the world. Yet even if you had billions of AI doctors in the world—each monitoring the health of a single human being—you could still update all of them within a split second, and they could all communicate to one another their assessments of the new disease or medicine. These potential advantages of connectivity and updatability are so huge that at least in some lines of work, it might make sense to replace_ all_ humans with computers, even if individually some humans still do a better job than the machines.

Harari’s article expands to discuss the broader question of whether AI will lead to tyranny. I recommend reading the full piece.