Between November 2022 and March 2026, permanent employment contracts (CDI) for 21-29-year-olds fell by 28.3% in French occupations most exposed to generative AI. This figure comes from an HEC Paris-Banque de France study published in October 2026. The shortfall accumulates, position by position, quarter by quarter.
The Essential Facts
- In March 2026, permanent employment contracts for 21-29-year-olds in the occupational quintile most exposed to AI were 28.3% below their November 2022 level (HEC Paris-Banque de France, October 2026).
- In 2024, the Stanford AI Index valued American private AI investment at $109.1 billion (Stanford AI Index, 2025).
- In India, job postings requiring AI skills show an average wage premium of 28% according to the World Bank (World Bank, AI and Labor Markets in Emerging Economies 2024).
- The observed decline is correlated with exposure to generative AI, but its precise causes are not identified by the study.
Hirings in AI-Exposed Positions Plummet While Technology Advances
The study analyzes labor force movements from January 2019 to March 2026, matched against the Banque de France’s business tendency survey. It distinguishes “AI-exposed” positions—those where tasks performable by generative AI tools represent a significant fraction of the work—from positions with low exposure. The result is clear: in exposed positions, hirings fell 28.3% between November 2022 and March 2026. In occupations least exposed to AI, permanent employment contracts for 21-29-year-olds declined by roughly 8%.
The authors’ interpretation deserves careful reading. In the sample studied, the total workforce of firms using AI intensively is not sharply distinct from other firms. The results are consistent with a reorganization of hiring in exposed occupations, without demonstrating that firms deliberately avoid these positions. Two reasons combine.
Second is a simple economic calculation: if automating a position costs less than hiring, and if regulation makes automation more predictable than human resources management, firms choose automation, but without training existing employees in the new tools.
The result is counterintuitive. The decline in hiring for exposed positions occurs amid economic slowdown, post-Covid correction in tech, and growth in remote work. The 21-29 age group shows the most marked relative decline in hiring across the study.
An Investment Gap Widening at a Pace Policy Can Barely Catch
The OECD publishes its data on private investment in artificial intelligence annually. In 2024, the Stanford AI Index valued American private AI investment at $109.1 billion. Half of European investment concentrates in the United Kingdom and Germany, placing France at significant distance from the technological frontier.
This figure measures something other than relative wealth. AI investment is cumulative: data trains models, improved models attract users, users generate new data. Economists call this network effects. A platform that accumulates eighteen months’ lead over competitors in training its models occupies a structurally difficult position to catch, because its models are better, its marginal costs lower, and its customers less inclined to switch providers.
Economist Noah Smith’s thesis, developed in his recent work on the future of the economy and AI, is empirically testable. Smith argues that an economy retreating from technology while competitors adopt it enters a bifurcation whose return path narrows as network effects consolidate. European data from 2022-2026 illustrate this dynamic.
The robotics comparison points the same direction. In 2023, China installed 276,000 industrial robots, compared to 28,355 in Germany, according to the International Federation of Robotics annual report. Germany remains Europe’s leading country in terms of robot density per industrial worker, but the volume gap indicates that China is building an automated manufacturing base at a speed Europe cannot absorb through upmarket shift alone.
The Asian Wage Premium Measures What Europe Is Not Training
The data on wage premiums in South Asia deserves reading as a signal rather than mere market statistics. In India, according to the World Bank, job postings requiring AI skills show an average wage premium of 28%.
This differential has a precise meaning. It indicates that Asian labor markets have already integrated AI competence as rare and valued skill. Firms pay more to secure it. In return, workers have strong incentive to train, independently of public programs. Private demand for AI training creates its own market.
In France, and more broadly Europe, this market signal is dampened by two factors. First, in France, hirings of young people in occupations most exposed to AI have declined more sharply, but the study allows neither establishing its cause nor concluding a European-wide drop in demand for AI skills, which reduces the perceived wage premium, which reduces incentive to train. The circle closes. Second, regulatory uncertainty discourages firms from investing in training for tools whose legal framework may change from directive to directive.
Carl Benedikt Frey, in his recent work on technology and social mobility, extending analysis begun in his studies on workplace automation, documents how the halt of technological progress hits least-skilled workers first—they lose both the routine jobs eliminated by automation and access to new skilled jobs created by technology. France, with its strong segmentation between protected permanent contracts and precarious employment, is particularly vulnerable to this mechanism: permanent employees absorb the cost of non-adoption through stagnant productivity; young entrants absorb the cost through absent entry-level positions.
This dynamic deserves confrontation with more institutionalist reading. Daron Acemoglu, whose work on institutions and the distribution of technological gains sets the reference, poses a different question: who captures the benefits of automation, and under what conditions are workers beneficiaries rather than victims. His answer invites distinguishing adoption speed from transition quality. Fast adoption concentrating gains on shareholders while deskilling remaining jobs constitutes a distinct model from the one Smith analyzes. Smith’s diagnosis and Acemoglu’s framework address two separate questions.
The first is a question of speed and competitiveness. The second is a question of benefit distribution.
French data from 2022-2026 show AI adoption slightly below the EU average but rising sharply, and French productivity still below its pre-Covid trend. Intensive uses remain more frequent in large firms with legal resources to navigate regulation. For further analysis on sharing technological gains between advanced and emerging economies, the journal’s analysis of polarization by robots in emerging countries offers useful complementary framing.
Regulation Can Protect or Paralyze Depending on Its Calibration
Before concluding that regulation is the problem, one must specify exactly what the data say. The study measures correlation between occupational exposure to generative AI and hiring trends, without analysis of the AI Act. It proves neither direct nor exclusive causality. Other factors play in: the eurozone’s economic slowdown in 2023-2024, rising interest rates compressing investment budgets, and global macroeconomic uncertainty reducing hiring across many sectors.
Regulation by the AI Act is not identical to prohibition. It imposes transparency, audit, and documentation obligations on systems classified as high-risk, which include among others employment, credit, justice, biometrics, education, critical infrastructure, migration, and law enforcement. These obligations carry significant cost for SMEs according to initial compliance assessments. Fixed costs may weigh proportionally more on SMEs. A recruiting tool may be high-risk and entail compliance obligations.
Calibration is therefore the central lever. Regulation that precisely distinguishes high-risk uses from routine uses, and that sets compliance thresholds proportional to firm size, could maintain essential safeguards—protection against algorithmic discrimination and transparency on automated decisions—while reducing the brake on SME adoption. The question of civil liability framework for autonomous systems is directly linked to this calibration: so long as liability for AI system errors is not clearly assigned, firms prefer avoiding risk.
The comparison with American industrial policy is instructive here, but with nuances. The CHIPS and Science Act mobilized massive public financing for semiconductors; the Inflation Reduction Act mobilized broad financing and credits for clean energy and certain advanced manufacturing capacities, rather than provisions specifically devoted to AI. The Congressional Budget Office publication from May 2026 exposes a method for assessing effects of federal R&D. Domestic semiconductor production increased, but production costs remain high compared to Asian standards. Public procurement can create demand; it does not automatically create long-term competitiveness.
Europe must observe this experience attentively before reproducing it mechanically.
Decisions in 2025-2027 and Their Probable Effects on 2028-2035
The question posed by 2022-2026 data for the years ahead is structured around a threshold. Network effects in AI have a particular property: they do not strengthen linearly. They accelerate beyond a certain critical mass point, after which dominant platforms become very hard to dethrone. If Europe intervenes before this threshold, targeted industrial policy can reverse the dynamic. If it intervenes after, it risks funding actors who will never have the data or user base necessary to compete.
Two signals would allow situating Europe relative to this threshold. The first: the quarterly evolution of hiring in AI-exposed positions, compared to the American trajectory. If the decline stabilizes and reverses by 2027, European industrial policy, notably the Commission’s AI program planned in the Draghi plan framework, will have produced measurable effect. If the decline continues, irreversibility becomes progressively more probable. The second signal: the evolution of the wage premium gap between AI-trained workers in Europe and South Asia.
If this gap widens, it indicates that Asian labor markets have locked in their training advantage. If the gap narrows, it indicates Europe has succeeded in creating domestic demand for AI skills strong enough to value training.
The scenario of catching up through targeted industrial policy exists and remains open. Its conditions are precise. It requires public procurement oriented toward AI adoption in SMEs, with financing conditional on training existing employees, not merely deploying tools. It requires retraining programs specifically targeting positions whose hiring has declined, with short training durations and certifications recognized by employers. And it requires revised calibration of the AI Act for SMEs, without weakening protections for high-risk uses.
The scenario of irreversibility through network effects is equally realistic. A persistent investment gap between the United States and Europe could progressively reinforce American technological lead. The stake then becomes specialization: on which AI segments can Europe build a defensible position, leveraging its strengths in health data, industrial data, and an ethical framework valued on certain global markets.
The question of the state’s capacity to guide technological transitions lies in this choice: without institutions capable of executing industrial policy commensurate with the stakes, neither the catching-up scenario nor the specialization scenario materializes.
France’s Room for Maneuver
The 2022-2026 data describe a mechanism, not a fate. The observed decline is correlated with exposure to generative AI, but its precise causes are not identified by the study. This environment can change.
Three concrete decisions would have measurable effect within a two-year horizon. First, a revision of the AI Act’s compliance timeline for SMEs with fewer than 250 employees, with an 18-month postponement of the heaviest documentation obligations for moderate-risk systems. Second, an accelerated training program in consumer AI tools, two to six weeks’ training, financed through personal training accounts, targeting specifically employees in sectors most affected by the hiring decline. Third, AI conditionality in public procurement: firms responding to public calls above a certain threshold must document their AI adoption and training strategy.
None of these measures guarantees catching up. They create conditions for the market signal—the wage premium for AI competence—to transmit to French employees and SMEs. Without this transmission, that signal will be captured elsewhere. It already has been for three years.
Sources
- Bergeaud, A. et al. (October 2026). Private French Firms Hire Fewer Young Workers in AI-Exposed Jobs. HEC Paris-Banque de France.
- OECD. Digital Economy Outlook 2025. Paris: OECD. https://www.oecd.org/digital/oecd-digital-economy-outlook.htm
- OECD. Employment Outlook 2025, chapter on AI and Employment. Paris: OECD.
- International Federation of Robotics (IFR). World Robotics Report 2024. Frankfurt: IFR.
- World Bank. AI and Labor Markets in Emerging Economies 2024. Washington DC: World Bank.
- Congressional Budget Office (CBO). Evaluation of the CHIPS and Science Act, 2026. Washington DC: CBO.
- Noah Smith. A.I., the Economy & the Future of Humanity (2025). Substack / Noahpinion.
- Carl Benedikt Frey. How Progress Ends (2025) and recent work on technology and social mobility. National Bureau of Economic Research. https://www.nber.org/
- INSEE. Labor Market Data 2022-2026. Paris: INSEE.
- Banque de France. Administrative Business Data 2019-2026. Paris: Banque de France.
