Hey readers,

A year ago, I covered the arrival of GPT-3, a language-generation AI developed by OpenAI, the Elon Musk-founded artificial intelligence lab. At the time, the GPT-3 rollout marked a validation of an important theory of AI — that we could get vastly better results just by training AI more — and we’re now starting to see commercial applications for it.

One that stands out is Github Copilot, which uses GPT-3 to write code. It’s not good enough to replace a programmer yet, and it has been mostly marketed as a tool to save time and check your work.

But while GPT-3 isn’t quite good enough to replace programmers or journalists, its emergence as a commercial product is in keeping with a notable trend: AI is indeed starting to move into the human labor force, but it’s much closer to replacing programmers and journalists than, say, robotics companies are to replacing agricultural workers or delivery people.

We often think of low-wage jobs like harvesting fruit, driving for Uber, working assembly lines, or delivering packages as those most ripe for automation. But with AI, it looks reasonably likely that the opposite will be true: it’ll turn out much easier to automate some or all tasks performed by stock traders, programmers, lawyers, and journalists than tasks performed by drivers, cashiers, and agricultural workers.

The business press periodically runs stories about stock traders who are struggling to make a living in a world where robots are really good at trading stocks, to the point where humans can’t generally outperform them. The programming isn’t quite there yet (and notably, there are still lots of people employed in finance). But the commercializations of GPT-3 thus far suggest that it’s the path we’re on.

The potential downsides of automating “knowledge work”

For a long time, the refrain about AI was that the hard things were easy and the easy things were hard.

If you wanted a computer in the early 2000s to solve math problems with incredibly large numbers, for instance, you were all set. If you wanted it to recognize a dog, you were out of luck.

Now, computers can recognize dogs. But it’s still sort of true that the hard things are easy and the easy things are hard.

Knowledge work like contract-writing and stock-trading and programming are more straightforward to teach to a machine-learning system like GPT-3 than “basic” human skills like picking things up off the ground or walking on a bumpy road. Amazon has automated managing warehouse workers before it succeeded at automating warehouse workers themselves.

I hesitate to predict too much about the coming decades, but I believe AI is going to seriously affect how work gets done. And I suspect that’s largely not going to look like robots picking strawberries or installing drywall.

Instead, one scenario discussed by experts who are concerned with risks from advanced AI is that more and more of our sophisticated systems will be handed off to computers whose outputs are impressive but whose behavior we don’t fully understand. Stocks will be traded and code will be written by robots, but human programmers and financial analysts won’t necessarily be able to make sense of what the robots do — they’ll only know that it works.

Contracts will be written by AIs and evaluated by other AIs, and there’ll be fewer humans with the expertise to evaluate the contracts and confirm that the AIs are doing things that make sense. We might invent additional specialized AIs for the task of auditing other AIs and reporting to us on whether their output is still high-quality, since it’ll be hard for humans to tell.

All of that could coexist with a world where self-driving cars still aren’t really widely commercially viable, where fruit-picking, child care, education, nursing, cleaning, plumbing, and construction still basically work the same way they do today.

The earliest vision of automation was that we’d hand off manual labor but retain decision-making power: Robots would work in the factories and fields while humans would trade stocks and negotiate contracts and write the next generation of computer programs. But we may be on the verge of an era where we instead hand off decision-making power to machines but keep the manual labor.

There’s some potential for that to be a path we regret heading down.

—Kelsey Piper, @kelseytuoc