In my recent essays, AI Is Fire and Artificial Sentience vs Artificial Intelligence, I wrote that we shouldn’t fear a super-advanced “Artificial Sentience,” because when AS arrives it won’t be hostile and it will never be smarter than the super-advanced people who create it. But that’s a long way off. Meanwhile, there is something more mundane to fear from AI: that AI and robots may do so many jobs better than people that job loss will be widespread and permanent.
AI is a threat to humans in the same way that printing presses, textile mills, and other factory machines have been a threat to humans over the centuries: they replace some amount of human labor with machine labor. AI’s impact will be like that of the railroad. The development of the railroad was a tectonic shift in the industrial world. It was a transportation technology, but it transformed industry, communication, migration, leisure—pretty much everything—along with it. AI will be the railroad of job loss.
Today, over the infinitude of human tasks, AI exceeds human performance at a mere handful. To answer the question, “Will robots take my jobs?”, you must understand that computers only excel at closed systems with binary rules. So computers’ early successes outperforming humans were unsurprisingly in math and chess. As AI has improved, it’s been able to handle more gray area. Every task involving gray area requires a heavy data set, a lot of training, and time to learn and improve. The more gray area in a given job, the harder it is for AI, and the longer it will take for AI to replace human labor.
The Incompetent Robot
AI today still underperforms people even in jobs with virtually no gray area. Have you ever actually watched, for example, a robot vacuum cleaner at work? It’s almost comical how incompetent it is at path-finding. My (admittedly, basic model) Roomba routinely goes in circles, goes over the same territory over and over, repeatedly bumps into the same objects every day even though I haven’t moved them, and all in all takes maybe 3-5x longer than a human would to clean the same area. But I’d rather have my Roomba vacuum inefficiently than spend my own time vacuuming. Note that Roomba doesn’t eliminate the vacuuming jobs—I still have to vacuum the furniture myself. It just reduces the number of vacuuming jobs. Similarly, I don’t expect robots to replace all police or fire fighters in the future. But I do expect that robots, under the supervision of human partners, will reduce the number of police and fire fighter jobs.
These will be recurring themes in the steady advance of AI. It will reduce rather than eliminate the jobs done by people in a given industry. And it will often do so not because it outperforms people at the given task, but merely because it renders people (at least partially) unnecessary. Just as most people who’ve ever vacuumed will be unimpressed by the Roomba’s skill, most people who’ve ever driven a truck will be unimpressed by autonomous trucks. AI often disrupts an industry not by improving on human performance, but merely by doing it well enough and cheaper.
There is less gray area in long haul trucking than in driving residential neighborhoods, where the large population of pets, kids, and bicyclists creates more gray area problems for an AI. We often hear that self-driving cars will be much safer—and that’s true, but in the near term it’s not because of AI. What will make self-driving cars safer is that they all talk to each other. Each vehicle knows what every other is doing. Much of the power of autonomous vehicles will initially be in the power of the network rather than the AI itself.
A great many tasks that seems obvious and objective are surprisingly difficult to an AI. Google’s neural network’s ability to recognize pictures of cats had a failure rate of 25% back in 2012. Image recognition has come a long way since the: this one is nearly 94% accurate. Still, any small child can accurately recognize images of cats or trains basically 100% of the time. The labor of a small child is cheap—mine require only hugs or, sometimes, jelly beans. Of what use is very expensive yet largely ineffectual software? Like Roombas, image recognition software doesn’t work particularly well—but these AIs are getting better, and either do jobs that humans don’t want to do (vacuuming) or can’t do cheaply at vast scale (image recognition).
This is how AI will progress in the real world. It won’t be particularly good at a given job at first, and will take jobs due to other marketplace advantages. The earliest chess programs weren’t very good but were fun for intermediate players to get a game whenever they wanted. Robot vacuuming isn’t very good but is convenient for people who turn it on and leave the room. Auto-correct on your phone isn’t very good, but is better than nothing. Gradually, with technological improvement and ever-more data to train it, AI will exceed human skill at one task after another.
The Gray Market
Every job has some number of tasks with some amount of gray area. Some jobs are nearly all gray area—teacher, nanny, psychologist. Some have little gray area—checkout clerk, fast food order-taker, toll booth operator, long haul trucker. Most jobs are on a spectrum somewhere in between.
I call this spectrum “The Gray Market.” (Nomenclature isn’t my strength. Please comment if you can think of a better name!) The higher your job is on this curve, the safer and more desirable it is. Jobs that require creativity, flexibility, empathy, emotional intelligence, communication skill, or fine motor skill (at unpredictable tasks), have greater natural immunity to AI. They also strongly tend already to be higher value-add and higher paying jobs today.
AI is beginning to replace humans at light gray jobs that have very little gray area. As AI improves over the coming decades, it will steadily move up the curve, replacing workers in one industry after another. You could practically graph the likely progression of this—all jobs nationwide, ranked by how much gray area they have vs. the rate of development of AI over time—to calculate the number of job losses due to AI in each coming decade. This phenomenon will eventually slow as it approaches dark gray jobs that are nearly all gray area. AI will never replace humans at such jobs (though Artificial Sentience could).
Note that some jobs appear to be dark gray, but are surprisingly programmatic. Creating new culinary delights will be hard for AI, but the act of perfectly cooking any known recipe will be something we all expect our robot home cook to do in a hundred years. The greatest recipe, after all, is still just information, as is every cooking technique required to create it. Robots will outperform virtually any cook. To survive as a chef, you’ll have to invent new dishes all the time. I expect a lot of popular music will eventually be created by AI as well. Jobs whose main function is delicate interaction with other people (teachers, nannies, etc.) are dark gray—but art, and music have highly programmatic elements to them. I also expect that much of the software written in the future will be done so by other software programs. There will still be rock stars and software developers, just fewer than there would otherwise be. Each time AI outperforms humans at one of these tasks that seem particularly human, society will be shocked, just as we were when Gary Kasparov lost to Deep Blue at chess.
The Efficiency Argument
At Harvard Business School, I was taught that any innovation that accomplished a task more efficiently and put people out of work was a good thing. The displaced would surely find other pursuits (ideally, they would be assisted as necessary with worker retraining programs). While the new innovation produced all the GDP that the displaced workers previously created, those workers were now doing something else and producing additional GDP. Overall, GDP rises.
Warren Buffet has extended this idea, vis-à-vis AI, to the entire US economy. He argues that if one person could press a button and produce all the goods and services of the United States, it would free us all up to do other things, including work less. GDP rises, or quality of life rises, or both. He points out that 100 years ago, Americans worked longer hours.
Perhaps MBA programs and Warren Buffet are right. I certainly hope so! I just worry that their arguments are based entirely on prior results—that human industry has ever been thus—rather than a thoughtful analysis of how new technology could affect future results. An argument that only looks backward for its evidence cannot possibly be a complete argument. What if something fundamentally different is right around the corner?
Something Fundamentally Different Is Right Around the Corner
Economists routinely point out that automation has been eliminating jobs for decades without crippling the economy’s ability to create new jobs. But AI is still in its infancy. It has barely begun to take jobs from people. Decades hence, it will take so many jobs from so many people that one wonders if it will massively overwhelm the economy’s ability to create new jobs.
Anatomically modern humans have existed for around 200,000 years. In that time, we’ve experienced two fundamental economic epochs: the agricultural and the industrial. The agricultural epoch required very large numbers of low-to-medium skill laborers. The industrial epoch required large numbers of medium-to-high skill laborers. Now we are entering our third industrial epoch: that of Information Technology (including AI and robots). Based on what we’ve seen so far, it seems likely to require small numbers of very high skill laborers.
Another argument that AI & robots won’t lead to fewer jobs is that “Producers will only automate if doing so is profitable. For profit to occur, producers need a market to sell to in the first place… if robots replaced all workers, thereby creating mass unemployment, to whom would the producers sell?” This is a purely economic argument divorced from technological reality. Once it’s technologically possible to create robots that are so good and so cheap that they wipe out large numbers of jobs, someone will do so. It’s a prisoner’s dilemma. Robot-avoiding producers will go out of business while robot-embracing producers will sell to a shrinking, but on average richer, market of medium-to-dark-gray laborers.
I can imagine a world where ever fewer, ever larger corporations make an ever greater percentage of the goods and services we consume. What will everyone else do? Reducing the average work week from 40 to 30 hours (while maintaining full employment) sounds great. Reducing the average work week to 0-2 hours would be catastrophic unless the human race suddenly becomes whole lot more generous. I know a lot of entrepreneurs and business executives: as a group they disproportionately credit success to themselves rather than to the environment they operated in. Will the business leaders of the future be wise enough to recognize that they happen to have been born into a time and place where their runaway success was possible? Will they be willing to share with (what might be) a large perpetually unemployed and underemployed population? Or will these future titans of industry distribute just enough calories to the masses via in-home food printers and endless Virtually Reality escapism to maintain the status quo?
There are things we can do to prevent this bleak future. We can use some of these same information technologies to improve universal education around the world—so that humans everywhere are better equipped to compete with technology for more of those jobs. After all, technological solutions are often expensive, and only become cheap with scale. A ubiquitously smarter human race would cede fewer jobs by being cheaper and better than their technological replacements. Fundamentally improving the global education system with information technology is already underway, so there is reason to hope on that score.
But this is a delaying tactic—the machines will continue to improve. So we would also do well to make ourselves wiser. The society of the future will have to deal with the prospect of a labor market fundamentally unlike any that has ever existed. We will need to forge a new society that treats everyone with dignity and respect, even if there isn’t much for a lot of people to contribute. Life-extending medical technology could average-up the life span of the human race. New technologies will make it possible to walk in the shoes of others. So perhaps there’s reason to be hopeful that we will become wiser in the coming centuries.
The irony of the conflation of AI and AS is that, ultimately, our best hope against Artificial Intelligence is in fact Artificial Sentience. As I wrote in AI Is Fire, I believe humans in the future will be linked to the cloud through a mind-machine interface such as nanobots in their brains. They will upload all their memories in real time, and download computing results whenever they need. And when human beings eventually create a true “deep intelligence in the network” it will immediately become part of our intelligence. When it takes off and re-designs itself at an ever-increasing rate, AI will be disrupted by AS-enhanced humans. People will be able to perform in their heads the most advanced techniques used by AI. While machines will still be more suited to many tasks than humans, people will once again out-perform AI at smart tasks.