Six CEOs Reversed Their AI Jobs Predictions Since April. The Skill That Matters Didn't Change.

If you've been waiting for the AI jobs forecast to settle into one number before deciding what to learn, stop waiting. It isn't going to settle — the CEOs making the predictions haven't settled on one themselves, and the number has moved six times in six months.

Six reversals, five months

In April, Sam Altman published a policy document calling for a four-day workweek and warning of "widespread job loss" as a near-term consequence of the technology his company sells. The same week, Altman and Yann LeCun publicly clashed — Altman's camp near 50% of entry-level white-collar work at risk, LeCun putting the real number at 5–10% and calling the higher figure "dangerous."

By May, Dario Amodei walked back his own 50% warning, reframing the risk around the Jevons Paradox instead: automate 90% of a job and the remaining 10% expands to fill the role, rather than disappearing. A second post the same month worked through what that reframe actually asks of a working professional — identify the 10% before it's decided for you.

June brought two more turns. Altman said in Sydney he'd been "pretty wrong" about the pace of entry-level elimination, days after Amodei published a 15,000-word essay warning displacement "may be an intrinsic property of the technology" — then, the same afternoon, funded 1,000 new $85K roles to help organizations adopt the thing he'd just warned about. By late June, Altman, Amodei, and Demis Hassabis shared a stage at the G7 and each gave a specific timeline for human-level capability — one to five years, depending who you ask.

Read as a sequence rather than seven separate news cycles, the pattern isn't "the experts disagree." It's that the same three people disagree with their own prior statements every four to six weeks. Nobody funding, hiring, or building this technology is treating any single number as fixed. Waiting for one to hold still before you act is waiting for something that isn't coming.

What doesn't move when the number does

Every one of those six positions — 50%, 5–10%, 90%, "pretty wrong," "intrinsic property," one to five years — implies the same practical instruction underneath the disagreement: some part of your job is going to be handled by AI, and some part isn't, and the second part is the part you get paid for. That split exists whether the first number is 10% or 90%. What changes with the forecast is the size of the slice, not the fact that a slice remains, or that knowing which slice is yours is the actual skill.

Most people never do the exercise of naming their own slice. They read the headline, feel the anxiety spike or the relief, and go back to whatever they were doing. The number was never the actionable part of the story. The actionable part is a five-minute audit: which of your weekly tasks would you still trust yourself to catch an error in, if AI produced the first draft of everything else?

A supply chain planner at a 220-person consumer goods company

Renata had read every one of these headlines and filed each one under "wait and see." Her forecasting reports — demand projections, safety stock recommendations — had been mostly AI-drafted for a year, and she signed off on them the way she signed off on most things: a read-through for tone, not a rebuild from the underlying data. After the G7 timelines story circulated in her team's Slack, she picked one recurring judgment call — whether to flag a demand spike as a real trend or a one-off blip — and started writing her own call before reading the AI's, then comparing.

Three weeks in, her calls diverged from the AI's about once every five reports, usually on exactly the ambiguous cases where the difference mattered. That gap is the slice. It didn't shrink or grow depending on which CEO said what that week.

A partnerships lead at a nine-person nonprofit

Dev manages the org's relationships with three foundation funders and drafts every grant report through AI first. He'd been treating "should I learn to do X without AI" as a question that depended on the jobs debate resolving — if the pessimists were right, better to build the skill now; if the optimists were right, maybe it didn't matter. He stopped trying to resolve that question and instead framed one decision explicitly: which of his three funder relationships needed a report he could defend line by line in a follow-up call, and which just needed to be accurate.

Naming that split took him fifteen minutes and didn't require picking a side in any CEO's forecast. Two of the three reports still go out AI-first with a light review. The one funder who asks hard follow-up questions gets a report Dev builds his own argument for before AI touches it.

The one thing

The forecast will keep moving — it's moved six times already this year, from the same three people. The task that survives every version of it is the same one: name the slice of your work where the judgment is actually yours, and practice catching the moments AI gets it wrong there specifically. That's a decision you can make this week, regardless of which number turns out to be closer to right.