The "AI will take all the jobs" debate is mostly noise. The actual 2025 evidence is more specific, and more interesting. Instead of a dramatic spike in firings, the early data points to a subtler shift in who gets hired in the first place. That distinction matters, because it changes what a sensible response looks like.
What the data shows
- The ILO estimates GenAI could affect roughly one-fifth of tasks globally; 1 in 4 workers are in an occupation with some exposure, but most jobs are transformed, not eliminated.
- A Harvard analysis of 62M LinkedIn profiles found AI adoption correlates with steep drops in junior hires at adopting firms, while senior hiring stays flat, driven by slower hiring rather than layoffs.
- Yet the Yale Budget Lab found no clear link between AI exposure and unemployment through mid-2025.
Put those three findings together and a consistent picture emerges. The technology is broad in its reach but shallow in its bite: it touches most jobs without removing many of them. The pain, where it shows up, is concentrated at the bottom of the ladder: the roles that used to absorb new graduates.
The early effect of AI on jobs looks less like a wave of firings and more like a closing door for entry-level roles, with companies skipping the junior hire for tasks AI now covers.
Why the entry level takes the hit first
Entry-level work is, almost by definition, the most structured and repeatable work in an organisation. It is the first-draft research memo, the routine ticket triage, the boilerplate code, the data clean-up. That is precisely the kind of bounded, well-specified task current models are good at. A manager who can get an acceptable first pass from a model is tempted to delay backfilling the junior seat rather than eliminate the senior who reviews the output.
This is why the signal shows up as slower hiring rather than redundancies. Nobody has to make a wrenching decision to cut a person; they simply decline to open a requisition. That quiet form of contraction is easy to miss in aggregate unemployment statistics, which is exactly why the Yale and Harvard findings can both be true at the same time. An economy can show no jump in unemployment while a particular cohort, this year's graduates, quietly finds the first door harder to open. Averages hide exactly this kind of distributional shift.
It is also worth being precise about exposure versus displacement. The ILO's "one-fifth of tasks" figure is a measure of how much of the work a model could touch, not how many people lose their jobs. A role that is 40% exposed is not a role that is 40% gone; it is a role where 40% of the tasks change hands and the remaining 60% (usually the judgement, the relationships, the accountability) become the whole job. That reframing is the difference between panic and planning.
The pipeline problem
If AI does the work juniors used to learn on, where do tomorrow's seniors come from? Seniority is not a title; it is accumulated judgement, built by doing thousands of small tasks and getting feedback on them. Remove the bottom rung and you do not just shrink this year's intake; you starve the talent pipeline that feeds every level above it three to five years out.
This is a collective-action trap. For any single firm, skipping the junior hire is rational: the work gets done, headcount stays lean, and the cost of an undertrained workforce lands somewhere in the future. But if every firm makes the same locally rational choice, the industry as a whole stops producing experienced people, and in a few years everyone is bidding for the same scarce seniors. The shortage you create by not training juniors today is the salary inflation you pay tomorrow. Firms that keep investing in the bottom of the ladder while their competitors don't may find that the apprenticeship itself becomes a competitive advantage.
What this means for a business or team
- Don't quietly freeze junior hiring. Treat the question deliberately: which tasks genuinely move to AI, and which were how your people learned the craft?
- Redesign the junior role around judgement, not throughput. Have new hires direct, review and correct AI output instead of producing the first draft by hand; they learn faster and you keep the apprenticeship intact.
- Invest in the review layer. As AI produces more first drafts, the scarce skill becomes evaluating them. That is a teachable, high-value capability worth building deliberately, the same theme our piece on does AI actually make developers faster explores from the individual-productivity side.
- Measure before you restructure. Don't assume AI covers a role end-to-end; many of these tasks need a human in the loop for the parts that actually matter, and the same discipline behind measuring AI ROI applies to headcount decisions, not just tooling spend.
The firms that win here will not be the ones that cut fastest. They will be the ones that use cheaper task-level work to give junior people more interesting, judgement-heavy work sooner, and so build seniors faster than their competitors. If you're planning your team's AI adoption or building the team around it, we're happy to talk it through.
Sources
- ILO: Generative AI and Jobs: 2025 Update
- Anthropic: Labor market impacts of AI
- Yale Budget Lab: Evaluating the Impact of AI on the Labor Market