Contents7
Will AI take young people’s first jobs? The two camps arguing over this are each right, but not at the same scale. Across the labour market as a whole, the available studies see no collapse. Among young people entering the most exposed occupations, a drop-off appears in the United States, and it is concentrated where AI does the work in place of the person rather than helping them.
For a business, the useful question is therefore not “will AI eliminate jobs?” but this: if AI does the simple tasks that used to be given to beginners, how will those beginners learn the job?
In short
Across the US labour market as a whole, Yale’s Budget Lab sees no discernible disruption in the 33 months that followed ChatGPT’s release.
According to a Stanford working paper, employment of people aged 22 to 25 in AI-exposed occupations is 19% below where it would be had it kept pace with that of their less-exposed peers, mainly because fewer of them are being hired.
In Denmark, a study published by the National Bureau of Economic Research (NBER) finds no effect on earnings or hours, but a reorganisation of tasks.
For a business, the useful question is who will take over the training role of the tasks that automation removes from beginners.
Position 1: AI takes entry-level tasks first
The argument is intuitive. The first years of an office job are made up of repetitive, well-defined tasks: writing a first draft, summarising a file, answering simple requests, coding a standard function. These are precisely the tasks that a language model does well and fast.
The strongest support for this position is a working paper from the Stanford Digital Economy Lab, “Canaries in the Coal Mine?”, by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen. It draws on payroll data from the provider ADP, covering millions of US workers through June 2026. Its revised version of 12 August 2026 says that employment of people aged 22 to 25 in AI-exposed occupations stands “19% below where it would be” had it kept pace with that of their less-exposed peers. Experienced workers show no comparable gap.
The detail matters more than the figure. According to the authors, the gap comes mainly from fewer young people being hired, not from more of them leaving. It is concentrated in occupations where the use of AI substitutes for human tasks; where it complements the work, employment is flat or rising.
The authors also tested competing explanations.
What does not make the gap disappear
Removing technology firms and computer occupations
Taking into account exposure to interest rate rises and to remote work
What qualifies the result
Taking education level into account attenuates the gap
Some of the diverging trends predate generative AI
The gap is more pronounced in the ADP data than in national surveys
Two caveats: it is a working paper, not a peer-reviewed article, and it covers the United States. The authors themselves present these results as early, descriptive indicators, not as proof of a causal effect.
Position 2: nothing massive is visible
The other camp does not deny these signals; it looks at the whole. In a note dated 1 October 2025, Yale University’s Budget Lab measured how quickly the distribution of US jobs across occupations has changed since the release of ChatGPT, and compared it with the upheavals caused by the arrival of computers and the internet. Its conclusion: the broader labour market “has not experienced a discernible disruption” in the 33 months since ChatGPT’s release. The authors point out that their analysis does not predict the future.
On recent graduates aged 20 to 24, the same note stays cautious. The gap between their occupations and those of people aged 25 to 34 was widening slightly faster in the months before its publication, which could be consistent with the Stanford work, but could also simply reflect a slowing labour market. This trend may in fact predate ChatGPT, and the samples are small.
In Denmark, Anders Humlum and Emilie Vestergaard linked surveys on chatbot adoption to administrative labour market records. Their working paper, published by the NBER (w33777) and revised in March 2026, finds precise null effects on earnings and recorded hours, ruling out effects larger than 2% two years after the launch of ChatGPT. What moves, in their view, is the organisation of work: employers absorb AI by reorganising tasks, and new tasks appear, such as content production, AI oversight and AI integration.
Why the two do not contradict each other
The studies do not measure the same thing: Yale looks at the distribution of all jobs across occupations, Stanford at the employment of one age group in exposed occupations, the Danish study at the wages and hours of employees already in work.
Key takeaway
A national average can stay flat while a specific subgroup falls behind
If AI removes some of the beginner tasks in a few occupations, the effect is real for the people concerned and invisible at the scale of the whole economy.
We also need to be honest about what these studies do not say. None proves that AI is the only cause of the drop observed among young people, and none claims to say what will happen in five years.
What we think
We do not believe in a mass disappearance of office jobs in the short term, and the studies cited above point the same way. On the other hand, we take the signal among young people very seriously, because it matches what our own automations show.
From the field
Our automations take on beginner tasks
We operate the daily import of new properties on a real estate portal and the entry of our expenses from a photo of the receipt. These are exactly the kind of tasks that used to be given to an assistant who was just starting out.
The problem is not that these tasks disappear. Many were tedious and taught little. The problem is that they also served as a school. An accountant learns to spot an odd entry by keying in hundreds of them. A developer learns to read code by fixing simple bugs. If the machine does all of that, the company has to organise learning differently, otherwise it will have no mid-level staff in five years.
What would change our mind: a peer-reviewed study, on European data, showing either a general decline in beginners’ employment or, on the contrary, no gap between exposed and non-exposed occupations.
What this changes in practice for a business
We draw five measures from this.
- Map what juniors used to learn.
Before automating a process, list the tasks it removes and ask yourself what they taught. Ask the question at the scoping stage, with the person who does the task today: it avoids hollowing out a job without meaning to.
- Make the junior a checker, not a spectator.
The first draft produced by AI still has to be reviewed. Giving this review to a beginner, with a senior alongside, is a good school: they learn to spot what is wrong, which is the heart of the job. AI oversight is, incidentally, one of the new tasks that the Danish study sees emerging.
- Keep some tasks manual on purpose.
Part of the work can stay done by hand during the learning period, even if the tool would do it faster. It is a training cost, to be budgeted as such.
- Train before you deploy.
Since it was rewritten by Regulation (EU) 2026/1744, which entered into force on 27 July 2026, Article 4 of the EU AI Act asks companies that use AI systems to take measures to promote the development of AI literacy among their staff, without requiring them to guarantee a specific level of literacy for each person. It is an opportunity to train everyone, juniors included, to spot when the tool is wrong; our guide to the AI Act and GDPR for an SME sets out what this obligation requires.
- Measure before cutting a post.
The gains promised by AI do not always hold up; we went into this in our analysis of the real return on AI in business. Replacing a hire with a tool on the strength of a demonstration is a gamble, not a decision.
Where to start
In October 2026, the studies do not say whether AI will deprive young people of their first job in the long run. They do say where to look: the beginner tasks that automation removes. Before automating a process, list what it taught and decide who will take over that training role. To choose what to automate first, and what can be done without AI, see Automation without AI: when a rule is enough and our Automation and bots section. And if you want to get rid of a re-keying task for good, our offer starts by observing the task with the person who does it today: that is the right moment to note what the task taught.
Frequently asked questions
Is AI destroying jobs?
Not at the scale of the economy, according to the studies cited in this article. Yale’s Budget Lab note sees no discernible disruption of the US labour market in the 33 months since ChatGPT, and the Danish study published by the NBER finds no effect on earnings or recorded hours. The drop observed concerns mainly people aged 22 to 25 in exposed occupations, according to Stanford.
Which occupations are most at risk from AI?
No list has a consensus, because each study measures exposure to AI in its own way. The Stanford working paper offers a more useful criterion than a list: the drop among young people is concentrated in occupations where AI is used mainly to replace human tasks. Where it complements the person’s work, employment is flat or rising, especially for experienced workers.
Should you stop hiring beginners?
No. A company that no longer trains juniors will have no experienced staff in a few years. It is better to redefine their role around reviewing what AI produces, with a senior alongside, and to keep a few manual tasks while they learn. The Danish study published by the NBER also observes that AI creates new tasks, including its own oversight.
Sources
- “Canaries in the Coal Mine?” (new tab)digitaleconomy.stanford.edu
- Yale University’s Budget Lab (new tab)budgetlab.yale.edu
- NBER (w33777) (new tab)nber.org
- Regulation (EU) 2026/1744 (new tab)eur-lex.europa.eu
Jim · founder of Brain Plus AI
Also in automation and bots
Stepping back10 min
The human role alongside AI: knowing how to judge
AI excels at some tasks and is confidently wrong on neighbouring ones. The human's value shifts: knowing where the line runs, keeping your judgement, answering for the result.
Stepping back11 min
AI return on investment: why it is taking so long
The figure of 95% of AI projects with no return went round the world. What it really measures, why the other studies say roughly the same thing, and the practical lessons to draw from them for an SME.
Stepping back8 min
Automated decisions: can AI decide on its own?
An algorithm may be more consistent than a manager in a hurry, but a decision that affects a person's life has to be explainable and open to challenge. The law leans clearly one way: the two positions, our view, and the rule to adopt without waiting for 2027.