Every few months, someone runs a headline number — “AI could affect 300 million jobs” — and every few months, the actual labor-market data tells a narrower, stranger story. The jobs AI tools are displacing right now aren’t the ones most people predicted five years ago, and the ones considered safe turned out not to be.
Start with what’s actually shrinking. Entry-level copywriting and junior graphic design roles have seen real contraction — not because AI produces better work than a skilled professional, but because a competent generalist with a chatbot can now do 70% of what used to require hiring a specialist for routine tasks: first-draft product descriptions, basic social copy, template layouts. Customer support is the other clear casualty: tier-one support tickets, the repetitive password-reset and order-status volume, are increasingly handled end-to-end by AI agents, with human agents kept for escalations. Several large software and telecom companies have quietly shrunk their tier-one support headcount by double digits over the past two years while keeping senior support staff intact or growing.
The roles that turned out to be harder to automate
Coding is the interesting counter-example. AI coding assistants are genuinely good — good enough that junior developer hiring has tightened at some companies — but senior engineering roles have not shrunk, because the bottleneck was never typing code. It was architecture, debugging distributed systems, and knowing which of the AI’s suggestions will quietly break production in six months. If anything, the value of a senior engineer who can supervise AI-generated code at scale has gone up.
Paralegal and legal-research work followed a similar pattern: first-pass document review and contract redlining are increasingly AI-assisted, but the roles that vanished were the ones that were pure document triage, not the ones requiring judgment about what a clause actually means for a specific client’s risk.
The pattern isn’t “AI replaces a job.” It’s “AI absorbs the most repetitive third of a job, and companies decide whether that means fewer people or the same people doing higher-value work.”
What the honest read of the data says
The roles most exposed share three traits: the work is text- or image-based, the acceptable error rate is low-stakes (a bad first draft costs little to redo), and the task doesn’t require accountability for a real-world outcome. Roles that involve legal liability, physical presence, or relationship trust — nursing, skilled trades, sales involving large accounts — have shown far less displacement, because the cost of an AI mistake in those contexts is much higher than the cost of a slow human.
None of this means the disruption isn’t real for the people living through it. A support rep or junior designer who’s been displaced doesn’t experience “aggregate labor market resilience” — they experience a job loss. The honest version of this story is that the pain is concentrated and real for specific roles, even while the aggregate numbers look calmer than the more apocalyptic predictions suggested.