Here we go again with the always-wrong AI-job-loss predictions: NEW: Microsoft AI chief predicts AI will automate most white-collar work within 18 months.
— Polymarket (@Polymarket)
[May 17, 2026]
People have made this prediction every month, at least since 2023. Three years later, despite record progress in AI, nothing has changed. But people keep predicting it anyway. Here is my post from Jan 2025 where I predicted nothing would happen. 1.5 years later, no white-collar job apocalypse.
These CEOs and other experts should lead by example and resign if they truly believe their predictions about AI destroying jobs. If Jamie Dimon and other overpaid, overrated billionaires think AI will inevitably eliminate millions of jobs, they should quit to bring about that outcome themselves. Instead, it’s always other people who are expected to lose their jobs–never them.
Of course, one can argue that it’s too soon to tell. This is related to Scott’s recently viral post, “The Sigmoids Won’t Save You“. This is the crux of the “AI-critic” argument. They keep promising “just wait,” yet they cannot support their warnings with any evidence. It’s like “…at some point something bad will happen. We cannot prove it and there is no evidence yet to support it, but it’s going to eventually happen.”
This reasoning is fallacious. It’s more like an appeal to ignorance (or argumentum ad ignorantiam), which is when someone argues that a claim is true because it has not been proven false. A common deadline is 2027. This was 6-12 months ago. So now they have 1.5 more months until “whatever promised bad thing” happens. I am willing to push the deadline out until 2035. That is how confident I am.
It’s worth keeping in mind these AI companies are running at large losses to build their userbases. They are being subsidized by VC funding, and then selling stock when going public. At some point, to turn a profit, they will have to start charging more or throttling usage, which is already happening:
So this reduces the value proposition of AI-coding compared to human coders. Many have attested that tokens usage can be very expensive and unpredictable, with roadblocks being thrown without warning for seemingly small tasks, resulting in huge bills. At some point, the math may show that paying a human to do it right the first time is better: