Your future job will be to keep AI on task
The one thing humans will always be comparatively good at is knowing what we want.
“Reason is, and ought only to be the slave of the passions, and can never pretend to any other office than to serve and obey them.” — David Hume
If you’re of a certain generation,1 you’ve probably seen the movie Office Space. If you haven’t, I strongly recommend it, both because it’s funny as heck, and because it’s a perfect encapsulation of a certain time and place in the world. The movie hearkens back to the big technology companies of the 1990s, when — according to the mythology, at least — nerdy engineers did all the real work while know-nothing middle-management types took all the credit. (You’ll also recognize this as the culture that gave rise to Dilbert.) The iconic character representing the backwardness and inefficiency of the 1990s corporation was Bill Lumbergh, the suspender-clad boss whose main function was to pester engineers to fill out useless paperwork.
A lot of nerdy types watched Office Space and assumed — or at least hoped — that in the end, the smart engineer types would take over corporate America from the plodding Lumberghs. And in fact, something like this happened in the 2010s — as Big Tech eclipsed much of the old economy, engineers became extremely highly paid, and began to fill the ranks of middle and upper management. It was the Revenge of the Nerds, the age of human capital, the triumph of humans who actually knew how to do difficult technical things.
But just as with the highly paid artisans of early 18th century Britain, the scarcity of human capital spurred a wave of automation. This year, AI found its killer app — Claude Code and other agentic coding tools that allow AI to do much (though not all) of the hard mental work that the much-put-upon engineers in Office Space were doing by hand. Although AI has not yet replaced many professions, the rapid progress has lots of people wondering what exactly humans will be useful for in 10 or 20 years. If AI does replace coders and mathematicians, what chance do any of the rest of us have?
Although some people in the AI industry still think that humans will be rendered economically irrelevant, that answer is increasingly unsatisfying. Realizing that a popular backlash against their industry is underway, many AI leaders and AI boosters are actively looking for the answer to the question of “What will humans be useful for?”. So far, the most popular answer, advanced by folks like Alex Imas, is that humans will be useful simply because they’re human:
The idea here is that it will become a sign of prestige and social status — which are always in short supply — to have humans do something for you instead of AI. No matter what else machines can do, they can never replace the knowledge that it’s a real human being making your sandwich.
I kind of have my doubts about this thesis — I’ve seen a lot of people pay extra to have a Waymo drive them instead of an Uber, so they didn’t have to sit with a human driver in the car. But maybe Imas is right; we’ll have to see.
But I have a slightly different answer to the question of “What will humans do?”. I think humans will continue to be required for something beyond simply being their beautiful human selves. I think there will be an increasing demand for human labor in the all-important job of maintaining AI alignment.
“Alignment” can mean many things in the AI community, but one basic definition is “ensuring that AI’s goals are the same as humans’ goals”. This is something that AI labs try to do before releasing their products to the world. But as AI becomes more and more agentic — as we turn over more complex and longer-lasting tasks to intelligent machines — it’s going to be harder and harder to keep them aligned with what humans actually want. And if there’s one thing humans will always have a comparative advantage at, it’s knowing what we want.2
In other words, it’s the lumbering Lumbergh, rather than the technically competent engineers, who I believe represents the ultimate future of human labor. He may seem boring and pointless, but by forcing engineers to file their TPS reports and do other seemingly useless tasks, Lumbergh is — however approximately and inefficiently — keeping the engineers’ goals aligned with those of the company they work for.
In the far future — maybe 10 or 20 years from now — this will mean that humans’ main productive function is to make sure that increasingly autonomous AIs stay on task instead of “reward-hacking”, rewriting their own utility functions, going rogue, or otherwise slacking off. Over the next few years, though, I expect human work to gently shade from technical work into alignment work, as we spend our hours verifying the output that AI delivers us.
Slopocalypse Now: why “verification” and “alignment” are the same thing
Generative AI has dramatically decreased the cost of many kinds of output. With the touch of a button, you can write an essay, turn a data set into an academic paper, write a report for your boss, and so on.
Everyone is doing this, and the result is that our society is currently being overwhelmed by a wave of AI output of questionable quality — in other words, what has come to be known as “slop”.
Slop is rapidly taking over every domain of human output. Over one-third of new websites are now estimated to be AI-generated, and over half of internet traffic is now believed to be AI. AI-generated court filings and news articles are on the rise. Even respected public figures are now posting obviously AI-generated content as their own. AI-generated political influencers are becoming a standard tool for electoral campaigns.
