The question "is AI taking jobs" is not answerable in the form it is usually asked, and the research keeps demonstrating why in ways that headlines keep flattening.
Aggregate employment is the wrong instrument for the change most plausibly under way. If a firm reduces graduate hiring while expanding senior hiring, aggregate headcount can be flat, rising or falling depending on other factors entirely, and the compositional shift — which is the thing of interest — is invisible in the total. Research that finds no aggregate displacement is not finding that nothing is happening. It is finding that the aggregate is not where it would show up.
Postings lead, employment lags, both are noisy
Job-posting data has become central to this literature because it moves faster than employment statistics and because it describes intent rather than outcome. Both properties are useful and both introduce error. Postings are affected by hiring-process changes that have nothing to do with demand; a firm that consolidates three adverts into one has not reduced hiring. Comparisons across time need to control for this and frequently cannot fully.
Attribution is where the difficulty concentrates
Firms that adopt AI tools early are not a random sample. They tend to be larger, better capitalised, more technology-intensive and concentrated in particular sectors — all characteristics that independently predict hiring patterns. Separating the effect of adoption from the effect of being the kind of firm that adopts requires either a natural experiment or strong assumptions, and most available designs rely on the assumptions.
This is not a reason to dismiss the findings. It is a reason to read the identification strategy before reading the conclusion, and to be suspicious of any account of this research that does not mention one.
Why early-career contraction matters more than its size suggests
A reduction in entry-level hiring produces effects on a delay. The people not hired this year are the people not available for promotion in five years, and organisations that thin their intake tend to discover the consequence long after the decision that caused it. Sectors that went through comparable contractions in earlier technology transitions took most of a decade to rebuild the middle of their seniority distribution.
For disclosure purposes, this is the part that belongs in a workforce risk narrative: not a headcount projection, which is speculation, but a description of what the organisation's intake looks like now relative to its stated plans, and what it would take to reverse a reduction if the assumption behind it turned out to be wrong.