I believe that currently existing narrow AI systems are unlikely to cause technological unemployment. This paper finds that narrow AI can automate individual tasks but not entire jobs; however, it specifically brackets off general AI. I’ve argued that we should still be worried about mass unemployment from artificial general intelligence (AGI) here.
However, there is a deeper economic argument that automation cannot cause mass technological unemployment, even if AGI is involved. A superintelligence would have an absolute advantage over humans at every task, for by definition, it’s better than humans at doing them. However, it would not have a comparative advantage at everything. An agent’s comparative advantage is any task they can do at lower cost (including opportunity cost) than everyone else. Since all agents have a comparative advantage at something, humans would still be hired to do the tasks they can perform at lowest cost.
This argument has been made here, but I’ll try to elaborate on it. Take the classic parable of Tiger Woods and the lawnmower:
Tiger is a great athlete. One of the best golfers to have every lived. Most likely he is better at other activities too. Tiger is probably in better shape than most: He can run faster, lift more, and work quicker. For example, Tiger can probably mow his lawn faster than anyone else. But just because he can mow his lawn fast, does this mean he should?
To answer this question we can use the concepts of opportunity cost and comparative advantage. Let’s say that Tiger can mow his lawn in 2 hours. In the same two hours he could film a television commercial for golf clubs and earn $100,000. By contrast, Joe, the kid next door can mow Tiger’s lawn in 4 hours. In that same 4 hours he could work at McDonald’s and earn $24.
In this example, Tiger’s opportunity cost is $100,000 and Joe’s is $24. Tiger has an absolute advantage in mowing lawns because he can do the work in less time. Yet Joe has a comparative advantage because he has the lower opportunity cost. The gains in trade from this example are tremendous. Rather than mowing his own lawn, Tiger should make the commercial and hire Joe to mow his lawn. As long as Tiger pays Joe more than $24 and less than $100,000, both of them are better off.
Now substitute Tiger Woods for a superintelligence. Even though the superintelligence can outperform humans at all cognitive tasks, it can’t perform all cognitive tasks at lower cost than all humans. If a company with both a superintelligence and human employees tried to assign the superintelligence all of its tasks, the machine would quickly eat up a lot of compute and I/O, and the company would still have to find humans to spend some time training the AI to do them or writing up task descriptions. This could end up costing more money than simply assigning humans to do some of those tasks. Also, by allocating the AI’s computational resources to the given task, the company forgoes opportunities to use those resources to beef up the AI’s performance on other tasks.
Do you buy this argument? Why or why not?
I think the standard assumption is that with any task you can create an expert system that is cheaper to power and run than it is to feed humans. Though I was talking with someone during EAG Virtual who was worried that humans might be one of the most efficient tools if you are only thinking about needing to feed them, and then it would be efficient for malevolent AI to enslave them.
I think the basic issue with the argument is that we are dealing with a case that Tiger Woods can just create a new copy of himself to mow the lawn while another copy is filming a commercial. So the question is whether creating the processors and then feeding them electricity to get the compute to run the process is cheaper than paying a human, and the most a human could be worth to pay is the amount that it costs to build compute that could replicate the performance of the human.
My intuition has always been that humans are unlikely to be at the actual optimum for energy efficiency of compute, but even if we are, I highly doubt that we’d be worth much more in the long run working for the AGI than it costs to feed us.
The solution to technological unemployment following AGI is to set everything up so that we make moving to a world in which there are no jobs a good thing, not to try to keep jobs by figuring out a way to compete with tools that can do literally everything better than we can.
A post employment society, where everyone has a right to their fraction of mankind’s resources.
Hmm, might the lawn mowing analogy break down with increasing speed difference and dependencies? Imagine if the lawn had to be ready for Tiger to play golf and Tiger being 1000 times faster than Joe.
Not sure if related, but I looked up Robin Hanson‘s predictions of the role of humans in the Age of Em, where brain emulations (ems) would become feasible and increasingly perform most of the economic activities on Earth. Summary of chapter 27 on the book website:
Unfortunately, I don’t recall how he derived at the conclusion, maybe somebody else can chime in.