Who Predicted the AI Jobs Apocalypse — and Now Disowns It

Sam Altman spent two years telling everyone AI would gut the job market. "Probably replace most of the jobs people do today," he said in 2023. By 2024 it was sharper: entire categories of work, "totally, totally gone."

In May 2026, he changed his mind. "Pretty wrong," he called it. "Delighted to be wrong about this" (Fortune, 2026).

The reason he gave is almost too on-the-nose. He'd set AI to auto-reply to his own Slack and email. It worked. He hated it. He missed the human part of talking to people — the part he'd spent two years insisting AI would simply replace (Entrepreneur, 2026).

He's not alone. Dario Amodei once put a number on it: 50% of white-collar jobs gone, unemployment as high as 20%. Now he says automation might expand what people do, not erase it. Jeff Bezos says AI will create a labour shortage. Jensen Huang has been saying for a while that the jobs fear is overblown. The same people who built the apocalypse narrative are now taking it apart, one interview at a time.

The data doesn't fully agree with either version. Tech layoffs through May 2026 hit 115,000 — nearly the full-year total for 2025 already — and Meta, Amazon, and Snap have all named AI directly as the reason. Goldman Sachs puts the number at 16,000 US jobs a month, hitting entry-level and Gen Z workers hardest. But the Yale Budget Lab looked at unemployment specifically in high-AI-exposure jobs and found nothing — no shift, no spike, nothing through March. Both of these are true at the same time. Disruption in pockets. No collapse. Not yet.

The economics are moving faster than the talking points. Uber blew through its entire year's Claude Code budget in four months. Microsoft has started cancelling AI licences for engineers because of cost. An Nvidia engineer told Axios that for his team, the compute now costs more than the salaries it was supposed to replace.

There's a name for what might actually be happening: Jevons Paradox. Make something faster and cheaper, and people don't do less of it — they do more. Radiologists read more scans. Programmers ship more code. The tool doesn't replace the worker. It multiplies the output. That's the pattern every general-purpose technology has followed, going back to electrification.

Here's the question nobody's answering directly: did the data change Altman's mind, or did the IPO timeline? OpenAI quietly filed in May. A calm jobs story is worth more to that filing than a scary one. And yet — OpenAI's own policy paper from earlier this year argues for taxing automated labour and cutting the work week to 32 hours. That's not the position of a company that believes the disruption isn't coming. Both things can't be true. We'll find out which one was the real position. The executives already picked theirs.