AI and Entry-level Jobs: Who Pays to Build Judgment?

Plenty of people are sounding the alarm about AI and graduate jobs. The headlines focus on the entry point, fewer junior roles, a generation locked out before it starts. However, the bigger picture is that careers are changing shape.

The conventional model of early-career development in knowledge work depends on some unspoken rules. The entry-level tasks a junior does, data cleaning, basic analysis, first drafts, standard documentation, initial research, produce something the organisation needs while teaching the person doing them how to think. None of it is glamorous but it is how professional judgment gets built, in almost every knowledge-work field.

So when AI absorbs that work, the task goes and the learning it carried does not go with it. Everyone agrees graduates still need judgment, the ability to tell a good answer from a plausible one. What that conversation lacks is edges. Judgment described this way can't be located or measured, so the requirement outlives the task while becoming far harder to pin on anyone.

The cost of building judgment hasn't disappeared. It has turned into something too vague for any one party to be sure it's theirs.

The visible half: the entry point thins

The clearest evidence sits in two markets. In South Korea, the National Assembly Budget Office found that employment among software developers aged 22 to 25 fell around 20 per cent from its late-2022 peak through July 2025, while employment among workers aged 35 to 49 rose over the same period. The Bank of Korea put roughly 211,000 youth jobs lost in the three years since ChatGPT's release in late 2022, concentrated in the sectors most exposed to automation. Fewer junior hires, more experienced ones.

In Singapore, SUTD analysis shows entry-level vacancies easing from 34,600 at the end of 2024 to 32,500 a year later, while full-time employment rates for graduates kept falling and junior roles increasingly demanded data analysis, AI familiarity and the kind of soft skills once expected of someone two years in.

A word of caution before leaning on any single number. A May 2026 working paper, The Broken Ladder by Peter John Lambert and Yannick Schindler, argues that once you separate AI exposure from remote-work exposure, much of the apparent AI effect on junior hiring is really a working-from-home effect. But it doesn't rescue the graduate. A separate Harvard working paper by Hosseini and Lichtinger found junior employment fell around 9 per cent after six quarters at firms that adopted generative AI. Whether the driver is AI or the way we now work, the layer that used to train people is thinning.

The entry point is only the easy part to see. Even where the junior role survives, the climb out of it is weaker than it was. Promotion rested on the same arrangement: do the entry-level work, absorb the judgment buried inside it, move up. Strip out the tedious work and the title stays, but the learning that earned the next role goes with it.

Who picks up the cost?

If the learning requirement survives the task, the next question is who picks it up. Three parties, and in the traditional education model, each is currently assuming it sits with one of the others.

The graduate is first. They're expected to arrive with judgment that used to take two or three years of junior work to build, except that work is now thinner on the ground. Fluency with the tools doesn't close the gap, because fluency was never the hard part. A graduate can produce a competent-looking analysis in minutes and still not know whether the framing is right, whether a key assumption is missing, or whether the conclusion would survive a first hard question.

AI tools got faster, but the person using it isn't becoming more discerning.

It's telling that ManpowerGroup's 2026 survey now ranks AI literacy and AI application development as the hardest skills to fill in Singapore, and frames the gap as people needing to use these tools with sound judgment. The market is already naming the missing thing.

The university is next, mostly without having agreed to it. Employability was treated as a transition problem for a long time. A student studied, polished a CV, practised interviews, landed a graduate role, and the first job took it from there. The institution never had to produce a finished professional, only a credible entrant, because the workplace would handle the rest. As the workplace does less of that, the responsibility drifts back up to the university, which was never built to carry it and often hasn't registered that it's now being asked to. A careers service measured on placement rates has no instrument that even detects the new problem.

The employer pays latest, and the delay is the trap. A mid-level professional isn't produced by a promotion. They're produced by exposure, repetition, feedback and rising responsibility, the same junior grind seen from the far end. Thin out that early exposure now and nothing breaks this quarter. It breaks in three to five years, when an organisation goes looking for people who can scope an ambiguous problem, supervise juniors and own a decision, and finds fewer than the headcount plan assumed. The task they automated to cut cost was also the training they were running without accounting for it.

The training nobody put on the books

The handoff is easy to miss because the thing being lost was never on the books. It taught through review comments, redrafts, awkward client meetings, messy data and the slow correction of early assumptions. None of it had a budget line, which is exactly why it can be cut without anyone deciding to cut it. AI hasn't removed the need for that learning. It has removed the situations where the learning happened. The gap doesn't close because the task closed. It moves, onto people who mostly haven't registered that it's now theirs.

The question underneath the headline

The headline question is whether AI is taking graduate jobs. The more useful one, the one a hiring statistic can't settle, is where the judgment-building that the old entry-level work used to do has gone, and who has actually taken it on. For now the honest answer is no one. The graduate assumes the system will still develop them. The university assumes the first job still trains. The employer assumes the pipeline refills the way it always did. Each assumption held while the entry-level task carried the learning inside it. None holds now, and the space between the three is where the next layer of mid-level professionals is quietly failing to form.

That gap will be more obvious when, years later, a hiring manager won’t be able to find a thirty-year-old ready to run a team, and can't say why the pipeline ran dry.

So the ones worth watching, universities and employers both, are the ones who have stopped assuming someone else is handling it. They're asking where judgment actually gets built now, and whether anything in the system still does it on purpose, rather than waiting for the old pipeline to refill on its own. That work doesn't sit cleanly with any single party, which is exactly why it keeps falling through. The rung that used to make mid-level professionals isn't going to rebuild itself. Someone has to decide it's theirs before the shortfall makes the decision for them.

‍