Jobs AI Is Actually Replacing, According to the Labor Data

"AI is going to replace millions of jobs" is usually said as if the net outcome is job loss. The actual World Economic Forum forecast for 2025-2030 says something more specific: 92 million jobs displaced, against 170 million jobs created, for a net global gain of 78 million jobs — roughly 14% of the current global workforce in new roles. Read on its own, that sounds like reassuring news. It's true, and it's also not the whole picture, because a net global gain says nothing about who specifically loses their current job versus who gets one of the new ones, and those are frequently not the same people with the same skills in the same place.

The more revealing number isn't the job count — it's the rate of change in what AI can technically do. The WEF's earlier analysis measured AI "exposure scores" — a metric for how much of a given job's tasks AI is technically capable of performing — growing at roughly 2% per year. As of 2026, that growth rate has jumped to 9% per year. That's not a small revision; it means the technical capability curve is accelerating well beyond what earlier forecasts assumed, which is a meaningfully different situation than "AI capability is advancing steadily."

“The job count getting better doesn't mean the disruption curve got any gentler.”

Averaged across the entire workforce, exposure numbers can sound manageable. Broken down by role, the picture sharpens considerably: an estimated 46% of the tasks performed in administrative roles, and 44% of the tasks performed in legal roles, are considered technically performable by AI today. These aren't fringe estimates about hypothetical future capability — they describe tasks AI systems can already largely do, in roles that employ enormous numbers of people. "Technically performable" doesn't automatically mean "will be replaced tomorrow" — organizational adoption lags technical capability — but it's the leading indicator worth watching, not the workforce-wide average.

It's tempting to assume AI disruption concentrates in developing economies with less capital to adapt. The data says close to the opposite: the IMF estimates that 60% of jobs in advanced economies are exposed to AI, compared to 40% of jobs globally. The reasoning is straightforward once stated: advanced economies have a larger share of white-collar, computer-based, cognitively-oriented work — precisely the category of task current AI systems are best at augmenting or automating — while developing economies retain a larger share of physical, manual work that remains harder for AI to touch directly.

By 2030, an estimated 14% of the global workforce — close to 375 million people — may need to change careers due to AI-driven disruption. This is the number that reconciles the seemingly good headline (net job growth) with the seemingly bad one (mass disruption): both can be true simultaneously. Total employment can rise while thirty-seven million people still have to find and retrain into a different kind of work than the one they currently do. A net-positive economy-wide number provides essentially no comfort to someone in one of the roles being displaced, especially if the new roles being created require different skills than the ones they have.

A national or global net job gain says nothing about timing, geography, or skill-matching. New roles created in one sector don't automatically absorb workers displaced from a different one, especially on a short timeline — retraining takes time, and not every displaced worker in an administrative or legal role is positioned to move directly into whatever new role gets created elsewhere. The forecast being net-positive is a real, meaningful data point about the economy as a whole; it is a substantially weaker data point about any individual person's specific situation.

The "170 million jobs created" side of the forecast isn't an abstraction — recent hiring data shows specific, fast-growing categories behind it. AI Engineer roles have grown around 143% and are the fastest-growing job title on LinkedIn for a second consecutive year; AI Content Creator roles have grown roughly 135%. Demand for human AI trainers and evaluators — people who review, correct, and rate AI model outputs to improve them — has grown 25-35% annually, and it's realistic as an in-house hire even outside big tech companies, for any organization running its own domain-specific model.

The prompt engineer story specifically is a useful cautionary tale about how fast these labels shift: pure "prompt engineer" job postings are shrinking, not because the skill stopped mattering, but because it's been absorbed as a baseline expectation inside other roles — AI engineer, AI trainer, context engineer, AI product manager — rather than remaining a standalone title. Someone searching job boards for the exact title "prompt engineer" in 2026 is searching a shrinking pool, while the underlying skill is in higher demand than ever under different names. It's a preview of how quickly job titles reorganize around a new capability, faster than the labor-market data can relabel itself.

The clearest pattern across this data isn't "which jobs are safe" — it's which task types are most exposed: routine, well-defined, primarily digital, information-processing work is the category current AI handles best. Work requiring physical presence, complex interpersonal judgment, or navigating genuinely novel and ambiguous situations remains comparatively harder to automate, not because AI can't get there eventually, but because it isn't there yet at the pace routine cognitive work is being reached. The practical takeaway isn't a specific list of "future-proof" jobs — it's paying attention to whether your own day-to-day tasks skew toward the routine, well-defined end of that spectrum or the ambiguous, judgment-heavy end.