How fast can AI change the workplace?

This past week the Internet became radicalised about the progress of artificial intelligence by, appropriately enough, an online essay with an assist from AI. In his viral piece “Something Big Is Happening,” Matt Shumer compared the current AI moment to the early days of Covid-19, when people didn’t realise how completely their world was about to change. Except this time, instead of a virus, the agent of transformation is a technology increasingly capable of replacing white-collar workers en masse.
The essay offered a reader-friendly (AI-shaped pieces are very reader friendly!) explanation of what the people driving the development of artificial intelligence have believed for a while. You can hear a similar story from the lips of Anthropic CEO Dario Amodei in my podcast interview with him this past week: On the way to hopefully utopian outcomes — amazing cures, growth beyond our dreams — AI may put a lot of people out of work in an incredibly short amount of time. Imagine the effects of automation and outsourcing on blue-collar labour, except inflicted on the professional class and compressed into just a few traumatic years.
People need to understand the part of this argument that’s absolutely correct: It is impossible to look at the AI models we have now, to say nothing of what we might get in six months or a year, and say that these technological tools can’t eventually replace a lot of human jobs. The question is whether people inside the AI hype loop are right about how fast it could happen, and then whether it will create a fundamental change in human employment rather than just a structural reshuffle.
One obstacle to radical speed is that human society is a complex bottleneck through which even the most efficiency-maxing innovations have to pass. As long as the efficiencies offered by AI are mediated by human workers, there will be false starts and misadaptations and blind alleys that make pre-emptive layoffs reckless or unwise.
Even if firings make sense as a pure value proposition, employment in an advanced economy reflects a complex set of contractual, social, legal and bureaucratic relationships, not just a simple productivity-maximising equation. So many companies might delay any mass replacement for reasons of internal morale or external politics or union rules, and adapt to AI’s new capacities through reduced hiring and slow attrition instead.
I suspect that the AI insiders underestimate the power of these frictions, as they may underestimate how structural hurdles could slow the adoption of any cure or tech that their models might discover. Which would imply a longer adaptation period for companies, polities and humans.
Then, after this adaptation happens and AI agents are deeply integrated into the workforce, there are two good reasons to think that most people will still be doing gainful work. The first is the entire history of technological change: Every great innovation has yielded fears of mass unemployment, and every time we’ve found our way to new professions, new demands for human labour that weren’t imaginable before.
The second is the reality that people clearly like a human touch, even in situations where we can already automate it away. Economist Adam Ozimek has a good rundown of examples: Player pianos have not done away with piano players, self-checkout has not eliminated the profession of cashier, and millions of waiters remain in service in the United States because an automated restaurant experience seems inhuman.
But here we come to the crucial issue with artificial intelligence: It is less inhuman than any prior technological development; indeed, by its nature, it simulates the human in a way that power looms and steel mills and PowerPoint software never did. So the unanswered question hanging over all these scenarios is how much that imitation shapes its capacity to replace human labour and our willingness to accept that replacement.
It’s easy to assume, in other words, that people will always prefer human waiters and human musicians and a human doctor to give us a medical diagnosis — unless we’re entering a world where people are increasingly habituated to interactions with simulated people, adapt their own humanity to the simulated version and come to prefer the simulation to the messier reality of flesh and blood.
In Silicon Valley, a somewhat socially maladroit realm, most of the discussion about a potential AI takeover focuses on the power of digital intelligence to supplant our own. The most important factor may not be raw intelligence but the social personae through which AI is mediated, and (without entering the debate about whether AI could be really conscious) how intensely people relate to AI agents as though they were conscious beings, just like us.
The more they do, the more profound the implications for labour and employment. And the more profound the implications for larger questions about human power and agency — which we’re more likely to give over, with potentially existential consequences, to an AI that we imagine not as a tool but as a friend. — The New York Times

Ross Douthat The writer is an author and New York Times columnist