Will AI replace entry-level jobs? The evidence does not justify a sweeping yes. It does justify a harder question: if employers use AI to hire fewer beginners, who will produce the experienced workers they still expect to recruit?
My position is simple. A company should not describe the elimination of its training pipeline as innovation unless it can explain what replaces that pipeline. Saving on junior salaries is an accounting result. Building a workforce is a management responsibility.
The warning is real. The causal verdict is not settled.
In an August 2026 revision of Canaries in the Coal Mine?, Stanford researchers examined ADP payroll data through June 2026. Employment among workers aged 22–25 in AI-exposed occupations was 19 percent below a benchmark based on the trajectory of less-exposed peers. The gap primarily reflected reduced hiring, not increased separations. Crucially, the authors called the findings descriptive rather than causal and found no evidence of widespread economy-wide displacement. Results also weakened with education controls and were stronger in their ADP sample than in national surveys.[1]
That is a warning about access to work, not proof that AI destroyed one in five young people’s jobs. Turning the estimate into that claim would replace analysis with a frightening but false headline.
There is meaningful counterevidence. In May 2026, Yale’s Budget Lab used a different design to account for persistent differences between AI-exposed and other occupations. It found no clear employment or wage effects attributable to AI so far. Its authors also noted that the survey data were less suited to detecting changes within narrow groups such as recent graduates.[2]
The two findings should not be forced into a fake consensus. They use different data and methods. We do not yet have an economy-wide causal verdict. We do have enough reason to scrutinize decisions that close the door on beginners.
Automation can teach. Exclusion cannot.
The strongest objection to my argument is that AI might make apprenticeship cheaper, not destroy it. That objection is supported by evidence.
A 2025 Quarterly Journal of Economics study followed the introduction of AI assistance among 5,172 customer-support agents. Productivity, measured as issues resolved per hour, rose by 15 percent on average; less experienced and lower-skilled workers benefited disproportionately. The researchers also found evidence of learning. This was one workplace setting, not a forecast for every profession, but it demonstrates that assisting beginners and eliminating them are not the same strategy.[3]
That distinction is the heart of the issue. If a junior employee uses AI, checks its answer, explains a mistake and receives expert feedback, the tool can become part of training. If the employee is never hired, no amount of productivity in the remaining team gives that excluded person workplace experience.
We should stop discussing these two arrangements as if they were an identical, inevitable consequence of better software. They are different organizational choices.
Some efficiency gains are somebody else’s training bill.
Consider a hypothetical firm that replaces two junior positions with a tool and retains its senior staff. It may genuinely produce the same output at lower cost. Nothing in this argument requires denying that possibility.
The problem appears when its future recruitment plan assumes experienced replacements will remain readily available. That plan relies on other employers continuing to train people. If enough firms make the same assumption, the savings at each business could become a collective shortage of opportunities to learn.
This is a risk mechanism, not a measured forecast of an inevitable skills shortage. New firms, lower prices, new tasks and better training tools could offset it. But an employer claiming those offsets will solve its problem should be able to name its own contribution. Hoping someone else develops your next expert is not a workforce strategy.
Do not preserve drudgery. Preserve the route to competence.
The answer is not to protect every repetitive task. Requiring beginners to waste time merely because previous generations did would be a defense of ritual, not education.
The answer is to preserve supervised responsibility. Give new employees real work with bounded consequences. Require them to justify an AI-assisted result. Have senior colleagues review their decisions, not merely their formatting. Test whether they can recognize when the tool is wrong.
Employers should measure entry-level hiring, paid training hours, progression into skilled roles and the time it takes new hires to work independently. A company could reduce one type of junior work while expanding another and still meet that test. What it should not do is count only the salaries removed and call the rest somebody else’s problem.
Universities also have work to do. But a university cannot supply the exact experience of a job an employer refuses to offer. Demanding an experienced beginner simply moves the contradiction into a recruitment advertisement.
AI does not make training obsolete. It makes the choice to train more visible. Employers are entitled to automate tasks. They should not be applauded for removing the first rung and then complaining that nobody knows how to climb.
nn
Sources
n
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, revised 12 August 2026. Version and limitations taken from the authors’ research page.
- Martha Gimbel, Joshua Kendall and Ryan Nunn, What We Do and Don’t Know About How AI is Affecting the Labor Market, The Budget Lab at Yale, 7 May 2026.
- Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work, The Quarterly Journal of Economics 140(2), May 2025, pp. 889–942. DOI: 10.1093/qje/qjae044. Uses the published 15% estimate, not the earlier working-paper version.