The AI jobs panic is real, but the timeline is wrong

What I'm actually seeing is companies using AI to handle the boring parts of knowledge work. Document summarization. First-draft emails. Data cleaning. The stuff that takes hours but doesn't require judgment. This frees up workers, but it also means companies need fewer of them for the same output. That's the tension nobody wants to acknowledge: productivity gains and job losses aren't opposites. They're often the same thing.

Simon Johnson, the MIT economist who won the Nobel, has been making this point at conferences. He spoke at UBS recently about the gap between AI adoption and job creation. According to Bloomberg's coverage, his concern isn't that AI will eliminate work entirely. It's that the new jobs might not be widely shared. They'll cluster in certain regions, certain skill levels, certain demographics. That feels right to me, tbh.

The timeline problem is what bugs me most. Every major automation wave in history took longer than predicted. ATMs didn't eliminate bank tellers overnight (bank teller employment actually grew for decades after ATMs appeared, because branches became cheaper to operate, so banks opened more of them). Self-checkout didn't gut retail employment in the timeframe analysts expected. These transitions happen, but they happen slowly, messily, with lots of adaptation along the way.

So why do we keep predicting AI will be different? I should probably know this better, but I think it's partly because AI feels more general-purpose than previous technologies. A robot arm replaces one specific motion. An AI model can, in theory, handle many different cognitive tasks. That generality is genuinely new.

But here's what the doomers miss: generality doesn't mean deployment. Even if an AI can do something, getting it integrated into actual business processes takes years. You need training. You need infrastructure. You need to convince middle managers to change workflows they've been using since 2008. You need to handle the edge cases that work fine in demos but break in production.

I'm not saying jobs are safe. Some roles will shrink. Entry-level positions in certain fields look genuinely vulnerable, because companies will use AI to handle work they used to give to juniors. That's a real problem for career pipelines. How do you become a senior analyst if you never got the junior analyst job where you learned the basics?

But the "everyone's unemployed by 2028" narrative? It's too clean. Reality is messier. Some sectors will be hit hard. Others will barely notice. The distribution will be uneven in ways we can't fully predict yet.

You might be wondering what any of this means for, say, your career. Honestly, I don't have a neat answer. The people who tell you to "learn to work with AI" aren't wrong, but they're also not saying anything specific enough to be useful. My best guess is that adaptability matters more than any single skill. But that's been true for decades.

What I do know is that the current conversation isn't helping anyone. We need less hysteria and more granularity. Which jobs, specifically? Which tasks within those jobs? What's the actual timeline based on deployment realities, not demo capabilities?

Until we get there, I'd take the apocalyptic headlines with a lot of salt. The change is coming. It's just not coming the way the discourse suggests.

Two new papers suggest we've been solving the wrong problem in model predictive control. I'm cautiously optimistic, but let me explain why the caveats matter.

Sarah Williams · 1 hour ago · 7 min