Generative AI: the first changes in the labour market are affecting young people

Chart 1: Correlation between the AI exposure index and actual use

Recruitment reacts before total employment levels

Recruitment can react to a technological transformation before it has a discernible effect on total employment. A firm that anticipates that its labour requirements will decline may limit new hires or not replace certain departing staff, without making existing employees redundant. This adjustment margin is particularly relevant in France, where the termination of permanent contracts (CDI – contrat de travail à durée indéterminée) is more regulated than in many other countries. We therefore separately track new hires and permanent contract terminations between January 2019 and March 2026.

To identify the occupations where generative AI-associated adjustments are most plausible, we construct an exposure index based on France Travail’s Répertoire opérationnel des métiers et des emplois (ROME – the Operational Directory of Trades and Jobs), which catalogues more than 1,500 occupations by the skills they require. A language model assesses the extent to which AI can perform or assist with each of these skills, taking into account factors that limit AI substitution, such as tacit expertise, physical presence or human interaction. The index measures potential technical capacity rather than businesses’ actual AI use. Nevertheless, as Chart 1 shows, there is a high degree of correlation with data on actual AI use measured by Google ATLAS.

We then juxtapose this index with comprehensive administrative data on labour movements and distinguish four age groups.

The decline in recruitment is concentrated among the youngest

After ranking occupations according to their exposure to AI, an employment gap becomes particularly apparent for 21-29 year-olds. In March 2026, permanent contract hires in the occupational quintile most exposed to AI were around 28% below their November 2022 level, compared with 8% for the least exposed quintile (Chart 2). The gap between the two trajectories thus amounts to around 20 percentage points. It narrows for the 30-39 and 40-49 age groups and then almost entirely disappears for employees aged 50 and over (more detailed results will be made available in this working paper   Exposition des métiers à l’IA générative, adoption par les entreprises et dynamiques du marché du travail en France).

Contract terminations follow a similar trend, although the decline is less pronounced than that for new hires. For the 21-29 age group, terminations stood approximately 23% below their November 2022 level for the most exposed quintile, compared with 6% for the least exposed quintile. The simultaneous reduction in both hires and separations corresponds to lower labour turnover. As new hires have declined more than contract terminations, workforce renewal has slowed, particularly for young people and in the most exposed occupations.

However, the earliest studies available from abroad provide a more mixed picture. In the United States, Brynjolfsson et al. (2026) also find a relative decline in the employment of young workers in AI-exposed occupations, mainly due to reduced hiring. In Norway, however, Hernæs et Kostøl (2026) do not detect a relative decline in the most exposed quintile when all occupations for young workers are considered. In Denmark, Humlum et Vestergaard (2026) observe a rapid reorganisation of tasks, above all, with no detectable average effect on hours or earnings. These differences suggest that we should distinguish between occupations, adjustment margins and institutions, rather than trying to discern, at this early stage, a single average AI effect on employment.

Chart 2: Permanent contract hires for 21-29 year-olds according to AI exposure

Recruitment of young people is declining more in firms that report intensive AI use

The exposure index measures what AI could technically perform in a particular occupation, while the Banque de France’s monthly business surveys provide information on AI’s reported use in firms. By pairing up the two sources, we track around 6,000 firms, nearly 600 of which reported intensive AI use at the beginning of 2026. These firms were already hiring in the more AI-exposed occupations, regardless of the sector, before ChatGPT. Therefore, adoption of AI is not randomly distributed across firms and depends in part on the nature of the roles that firms previously hired for.

When firms are categorised based on intensity of AI use reported at the beginning of 2026 (according to the survey described by  Genre et Parpais, 2026), the trend in permanent contract hires for 21-29 year-olds since the end of 2022 is shown to have been less favourable in firms that use AI intensively than in other firms, while the gap is much less pronounced for older age groups (Chart 3). However, there is not a statistically significant divergence in firms’ total employment levels. Even within firms using AI intensively, recruitment is also declining more in the most exposed occupations. The adjustment observed therefore relates more to the composition and pace of recruitment than to an already discernible contraction in total workforce.

Chart 3: Intensity of AI use and permanent contract hires by age group
 

The spread of AI coincides with other changes in the labour market

The decline in junior hires in AI-exposed occupations is, however, happening at a time when several forces have impacted those same jobs. Monetary tightening and the post-Covid correction in the technology sector have weighed on recruitment since 2022, particularly in professions that are themselves highly exposed to AI. Teleworking (which is also considered to have a negative effect on recruitment for junior tasks) further complicates the interpretation, as it largely affects the same occupation categories and has profoundly altered their organisation. Therefore, AI, teleworking and the employment cycle are not necessarily mutually exclusive explanations, as they may have jointly affected the same firms and the same tasks.

Furthermore, the Banque de France surveys measure the intensity of AI use at the beginning of 2026 only, while gaps in recruitment appear several quarters earlier. Firms that report intensive AI use may have started adopting it earlier, but they may also stand apart in terms of characteristics that explain both their recruitment trajectory and their subsequent adoption. The comparison therefore provides an insight into the link between reported adoption and changes in recruitment, without yet enabling us to separate the specific effect of AI from the selection of the firms adopting it. 

The slowdown in recruitment may affect the start of careers

The slowdown in recruitment for junior tasks may have effects that go far beyond a number of jobs not being created today. Several recent theoretical studies show that the automation of tasks assigned to new recruits can reduce learning-by-doing and, in the longer term, future human-capital accumulation (e.g., Asriyan et al., 2026 ; Afrouzi et al., 2026).

Our results do not allow us to establish whether this mechanism is already at work in France. However, they show that hiring of young people has begun to contract in exposed occupations, which is precisely where effects would be expected to appear first. Tracking total employment alone could therefore result in part of the adjustment being missed if AI first leads to changes in workforce renewal and workers’ skill development.