Generative AI and employment: could the decline in junior recruitment weaken tomorrow’s skills?

Since the arrival of ChatGPT at the end of 2022, the debate on artificial intelligence and employment has often been reduced to one question:

How many jobs will AI eliminate?

We may be looking at the wrong indicator. The first changes could appear well before a visible drop in total employment. Not initially through layoffs, but through slowing recruitment, positions that are no longer opened, and departures that are not replaced.

A study published on October 8, 2026 by the Banque de France provides a particularly interesting signal in this regard. Researchers Marco G. Palladino, Antonin Bergeaud, Antoine Bertheau, Simon Bunel and Oscar Degos observe that, since the end of 2022, permanent contract hires of young people have declined more sharply in occupations most exposed to generative artificial intelligence. At the same time, companies reporting intensive use of AI do not show a statistically clear divergence in their total employment compared to other companies. This result does not demonstrate that AI is destroying young people’s jobs. The authors are very explicit on this point.

But it opens a much more strategic question for companies:

If artificial intelligence is transforming the first tasks entrusted to juniors, how will we continue to build the skills of tomorrow’s experienced professionals?

A signal appears in the recruitment of 21-29 year-olds

To study the phenomenon, the researchers construct an exposure index to generative AI based on the Operational Directory of Occupations and Jobs of France Travail. More than 1,500 job profiles are analyzed according to the activities and skills they mobilize.

A large language model evaluates the extent to which AI can perform or assist these activities, taking into account factors limiting substitution, such as tacit expertise, physical presence or human interactions. This distinction is fundamental: being exposed to AI does not mean that an occupation will disappear. The indicator only measures a potential technological capability, not the actual use of AI in each company nor the probability of a position being eliminated. The researchers then match this index with administrative data on workforce movements. Among 21-29 year-olds, the gap is significant. In March 2026, the trend in permanent contract hires in the quintile of most exposed occupations was approximately 28% below its November 2022 level, compared to approximately 8% in the least exposed quintile. The gap thus reaches around twenty percentage points. It decreases for the following age groups and becomes almost non-existent among employees aged 50 and over.

The associated working paper provides an order of magnitude: in the most exposed quintile, this evolution represents approximately 6,200 fewer permanent contract hires of young people per month, relative to the trend level of November 2022. This figure is significant. However, it must be interpreted with caution.

It does not mean that AI has eliminated 6,200 jobs each month. It measures a gap compared to a reference trajectory in seasonally adjusted and smoothed series. The researchers speak of descriptive work and specify that their analysis is part of ongoing academic research.

Fewer entries, but also fewer exits

Another result is almost as important as that of hires. Terminations of permanent contracts are also decreasing. Among 21-29 year-olds, they are approximately 23% below their November 2022 level in the most exposed quintile, compared to approximately 6% in the least exposed quintile. Entries are therefore decreasing, but so are exits.

The researchers describe a form of freeze in workforce movements. The market renews itself less. Since recruitment declines more than exits, this freeze is asymmetric: it particularly affects the arrival of new employees. This is an important change in perspective. If we look only at the total number of people employed, we may see nothing impressive. Yet, beneath the apparent stability of headcounts, the way an organization renews its skills may already be changing.

A company can retain its experienced employees, use more AI, recruit fewer beginners and not replace certain departures. The change then appears in flows before appearing in stocks. This is precisely one of the lessons from the Banque de France:

the first adjustments associated with AI may be visible in recruitment before being visible in total employment.

What the study shows, and what it does not demonstrate

However, it would be incorrect to conclude that “AI is causing youth employment to plummet.” Several transformations have overlapped since 2022. Monetary tightening has weighed on recruitment. The technology sector has experienced a significant correction after the hires of the Covid period. Remote work has also changed organizations and may affect the integration of young employees. These phenomena often concern the same professions as those highly exposed to AI.

Another important limitation: the Banque de France observes the declared intensity of AI use at the beginning of 2026, while the differences in recruitment trajectories are visible earlier. Companies using AI heavily may therefore possess other characteristics that explain both their technological adoption and their recruitment policy. The authors are explicit: their results show an association, but do not yet make it possible to isolate the specific effect of artificial intelligence. This caution does not diminish the interest of the study. It simply prevents transforming an emerging signal into certainty.

In the United States, a similar signal appears

France is not the only country where this question arises. In the United States, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen analyzed administrative payroll data from ADP covering millions of workers through June 2026. Their study finds no massive displacement of employment at the economy-wide level. But among 22-25 year-olds, employment declines more sharply in occupations exposed to AI. It is now 19% below the trajectory observed among young workers in less exposed occupations. The phenomenon mainly occurs through a decrease in recruitment rather than an increase in departures. The resemblance to the French signal is interesting.

But the Stanford researchers also present these results as early descriptive indicators, not as causal estimates. They notably observe that certain gaps already existed before the diffusion of generative AI. The OECD calls for the same caution. Young people are indeed experiencing more difficulties entering the labor markets of several developed countries, particularly young graduates. But the organization emphasizes that this deterioration began before the massive diffusion of large language models and that current data do not allow it to be attributed to artificial intelligence. We therefore have a signal to monitor, not yet an established economic law.

The emerging question: how does one become an expert?

For companies, the most important point may lie elsewhere.

A large part of the tasks on which generative AI is progressing was historically entrusted to employees at the beginning of their careers. These notably involved searching for information, producing an initial synthesis, preparing an analysis, writing a first text, processing data, programming standard functions or structuring a presentation. These tasks have productive value. But they also possess pedagogical value.

An experienced professional is not born an expert. They gradually build their judgment by performing tasks, making mistakes, observing their consequences, being corrected and comparing their work to that of more experienced colleagues. A senior developer started by solving simpler problems. A consultant learned by preparing analyses. A financial analyst built models before supervising them. A journalist learned to search, verify, select and prioritize information.

Yet automation can modify precisely these first steps. The authors of the working paper themselves emphasize this risk. Entry-level jobs are also places where, through practice, the skills of tomorrow’s experienced professionals are built. Their scarcity could therefore have consequences well beyond the period currently observed. However, they specify that their data do not yet make it possible to establish that this mechanism is actually at work in France.

This is where the issue becomes organizational.

Automating a task should not mean automating learning

For a leader, the question is therefore not to artificially preserve all the tasks formerly performed by juniors. That would be absurd. A company has no reason to maintain an inefficient manual activity when a technology makes it possible to perform it better or faster.

However, it must understand what that task made it possible to learn.

• If AI now produces the first analysis, how does the junior learn to construct an analysis?
• If it generates the code, how do they learn to identify a fragile architecture?
• If it writes a synthesis, how do they develop their ability to distinguish the essential from the accessory?
• If it produces a convincing but incorrect answer, do they have enough expertise to challenge it?

Skill development in the AI era cannot therefore be reduced to learning to use tools. It must also organize the construction of judgment.

For companies, three questions become particularly important.

• The first consists of mapping junior tasks now automated or heavily assisted by AI.
• The second consists of identifying among them those that played an essential role in learning the profession.
• The third consists of deliberately reconstructing these learning mechanisms: supervision, simulations, mentoring, critical review, confrontation with real cases, apprenticeship with experts and use of AI as a training tool rather than as a simple substitute.

Immediate performance cannot be the only criterion.

The productivity paradox

The Banque de France study does not demonstrate that artificial intelligence is massively destroying youth employment.

It shows something more subtle: the recruitment of young people is evolving less favorably in the most exposed occupations. The first adjustments may therefore appear in recruitment flows before being visible in overall headcounts.

For leaders, this changes the question. It is no longer enough to ask:

“Which tasks can we automate?”

We must also ask:

“What organization must we build when some of the tasks that trained our employees are now performed by a machine?”

Because a technology can create a productivity capacity. An organization can simultaneously, without having intended to, weaken its learning capacity.

Artificial intelligence creates a capability. Organizations create value.

This value will only be sustainable if they also continue to produce the experience, discernment and skills they will need tomorrow. The risk is therefore not only to reduce the number of juniors. It is to no longer sufficiently build tomorrow’s seniors.