Will AI Take Your Job? What the Research Actually Says

Two confident stories dominate. One says most knowledge work disappears this decade. The other says this is the usual technology cycle and net employment will be fine. The measured evidence supports neither cleanly, and reading it carefully is more useful than picking a team.

What the studies consistently find

Exposure is not displacement. The widely cited estimates that large shares of jobs are "exposed" to AI measure how many tasks a model could plausibly assist with. Exposure is a measure of overlap, not of firings. Nearly every paper that produces such a figure says so explicitly, and nearly every headline drops that caveat.

Productivity gains are real and uneven. Field experiments in customer support, software development and business writing have generally found meaningful speed and quality improvements. The consistent surprise is distribution: less experienced workers usually gain most, which compresses the measured performance gap between novices and experts.

Quality gains depend on task fit. The same studies find performance can drop when the task sits outside the model's competence and the worker trusts the output anyway. The benefit is not automatic; it tracks the user's ability to judge the answer.

Early labour-market signals are visible but narrow. Analyses of freelance platforms and entry-level postings in the most exposed categories — routine copywriting, basic translation, simple graphics, some support work — show declines. These are the segments most exposed to a general-purpose tool, and generalising from them to the whole economy is exactly the leap the data does not license.

Why the forecasts diverge so much

Because they disagree on assumptions rather than facts. How fast capability improves. How quickly firms reorganise, which historically lags the technology by years. Whether new task categories appear at the rate they have in past transitions. Whether the constraint is model capability or the messy business of integration, trust and liability. Change one of those assumptions and the same underlying data produces a wildly different number.

Economic history offers support to more than one view. Automation has repeatedly raised total employment while devastating specific occupations and regions for a generation. "Net jobs will be fine" and "this will be brutal for particular people" are both compatible with the record, which is why the aggregate argument so often talks past the human one.

The more useful framing

Ask which tasks in your work are drafting, summarising, translating, formatting or first-pass code, and which are judgement, relationships, accountability, physical presence and taste. The first set is being commoditised. The second is not, yet — and roles that are almost entirely the first set are genuinely exposed.

The uncomfortable middle is the entry-level rung. If juniors historically learned judgement by doing the routine work, and the routine work is now automated, the training pipeline needs redesigning rather than assuming. That problem is real, under-discussed, and not solved by either optimistic or pessimistic headlines.

What to do with this

Get fluent enough to supervise the tools rather than fear or worship them — our prompt generator and guide are a starting point, and comparing platforms tells you where each is actually competent. Invest in the parts of your work that require accountability. Be sceptical of any specific percentage-of-jobs figure, including the ones that agree with you.

Related: when AI can act, not just answer.