What jobs are safe from AI?

The principle, not a list of job titles

Predictions about specific titles age badly. The durable signal is what kind of work the current autonomy gap protects. AI is strong at scoped, digital, well-specified tasks (~80% SWE-Bench Pro, ~83% GDPval) but weak at reliable, accountable, long-horizon ownership without a human in the loop. That gap — the same one keeping AGI-2027 graded Open — maps directly onto which work is exposed.

Why "safe" is about autonomy, not intelligence

It's tempting to assume "smarter AI = fewer safe jobs," but the binding constraint in 2026 isn't raw capability — it's reliable autonomy. A model can draft brilliantly and still can't be left to own a consequential decision unsupervised. Work that requires being accountable for outcomes end-to-end is protected exactly as long as that autonomy bar stays unmet — which is what this scorecard tracks. See also will AI cause mass unemployment? and can AI replace knowledge workers?

How to know when the picture changes

The moment AI demonstrates reliable, unsupervised, end-to-end work — the drop-in-worker milestone — the "safe" map redraws fast. That milestone is the headline verdict on the AGI-2027 Thesis Tracker: one auditable score that moves the day the autonomy bar is cleared. It resolves by January 1, 2028.

Where does YOUR job land? Take the 60-second AI exposure check — six questions, an honest tier, no invented percentages.

Frequently asked questions

In 2026, the most durable work needs reliable, unsupervised, end-to-end judgment, physical-world action, or deep human trust — the things AI still can't do reliably. AI automates scoped digital tasks (~83% GDPval) but can't own a whole job unsupervised, so task-bounded digital roles are the most exposed.

Task-bounded, fully digital, well-specified work is most exposed, because that's where AI's capability is strongest and the autonomy gap matters least. But 'replace' still overstates it — AI is automating tasks within jobs faster than it's replacing whole roles.

Because the binding constraint in 2026 is reliable autonomy, not raw intelligence. Work that requires being accountable for outcomes end-to-end, unsupervised, is protected as long as AI can't reliably do that — the exact gap the AGI-2027 verdict tracks.

When AI demonstrates reliable, unsupervised, end-to-end work — the drop-in-worker milestone behind the AGI-by-2027 verdict, which resolves by January 2028. The Thesis Tracker moves the moment that changes.