AI and Employment: Displacement vs Productivity Effects
Analysis note Trésor-Éco n° 391 (June 2026) from the Direction générale du Trésor (Ministry of the Economy), authored by Martin Chopard, Elisa Cotet, Tristan Gantois and Eloïse Villani.
#AI and employment#generative artificial intelligence#displacement effect#productivity effect#substitution#complementarity#task-based approach#AI exposure
Summary
This Trésor-Éco n° 391 note (DG Trésor, June 2026) offers a cautious, well-sourced review of the economic literature on the effect of AI — mainly generative — on employment. Starting point: AI capabilities have progressed sharply (LLMs, generative AI), fueling concern (62% of French people see it as a risk to employment), but economic analysis calls for distinguishing perception from measurement.
1. Aggregate effect, weak for now. AI acts through two opposing channels: the displacement effect (substitution of automatable tasks) and the productivity effect (proven individual gains: +14% in customer service, +26% for developers; complementarity, lower costs, increased demand). Recent empirical studies do not identify a significant aggregate effect, for lack of hindsight and because adoption remains partial (≈20% of EU firms in 2025). The absence of a macro effect does not mean an absence of localized destruction: "AI" layoffs account for 4.5-6.2% of announced layoffs in the US in 2025, with a risk of "labelling" (AI invoked as a pretext — 59% of US companies).
2. Heterogeneous effects. Via the task-based approach, exposure varies by task (cognitive > relational > physical), but exposure ≠ effect: everything depends on the degree of substitutability/complementarity and the price elasticity of demand (Jevons paradox — a substitutable occupation with elastic demand can see its employment grow). Biased technical progress could disadvantage certain segments, with marked concerns for young people: −16% employment among exposed 22-25 year-olds in the US (Brynjolfsson 2025), a rise in unemployment among 15-24 year-olds in France (19.1%→21.1%) — with no established causality. By sector, finance, IT and business services are the most exposed.
3. Uncertain long term. Two scenarios coexist: massive substitution (if agentic/physical AI becomes widespread) or creative destruction (past revolutions created more jobs than they destroyed; 60% of workers today hold jobs that did not exist in 1940). The transition will generate costs (slow reallocation: −40% of the benefit of robotization in France), to be smoothed by training and bridge occupations.
Public policy conclusion: support the transition (the "Osez l'IA" plan, training 15 million people by 2030, the Académie de l'IA, France 2030) and invest resolutely in AI to avoid competitive decline — France sitting in an intermediate position (18% adoption, catching up) amid international competition.
Key takeaways
Key thesis. at this stage, no measurable aggregate effect of AI on employment (insufficient hindsight + limited adoption), but two opposing channels structure the analysis — displacement (substitution) vs. productivity (complementarity). AI is neither a "job killer" nor neutral: it is empirically undetermined, and it will depend on policies and the speed of diffusion.
Perception ≠ measurement.62% of French people see AI as a risk to employment (2025 digital barometer), while the macro effect remains undetectable. The note clearly distinguishes social anxiety from the data.
Task-based approach. (Acemoglu-Restrepo, Autor): an occupation is broken down into tasks, and exposure is measured, BUT exposure ≠ effect — one must determine substitutability (AI replaces) vs. complementarity (AI augments the worker, who refocuses on judgment/validation/relationship). Cognitive tasks = more exposed to LLMs; physical = less so (for now); relational = intermediate.
Price elasticity decides the sign. (Chart 2): if productivity gains lower costs and demand is elastic (graphic designers, developers, elasticity > 1), employment can grow (Jevons paradox); if it is inelastic (firefighters, executives, ~0), employment can fall. → technical substitutability does not prejudge the employment effect.
Proven individual productivity gains.+14% for customer service agents, +26% for software developers (targeted experimental studies, probably underestimated since obtained on models that are already outdated).
Very wide exposure range. between 5% and 60% of total employment depending on the study (cover chart). France estimates: Arquié et al. (2026) — 3.8% of work content at risk of automation today, up to 16.3% within 2-5 years. Richmond (2026, OpenAI) US: 18% of jobs at high risk, 24% will see their tasks recomposed, 12% growing, 46% little affected.
"AI" layoffs to be put in perspective. in 2025, cuts attributed to AI = 4.5 to 6.2% of announced US layoffs, diluted within larger hiring flows. Risk of "labelling": 59% of US companies admit invoking AI to justify hiring freezes → AI serves as a pretext masking structural/post-Covid reasons (Josh Bersin case).
Concerns for young people. (the most concrete point): Brynjolfsson et al. (2025) — employment among 22-25 year-olds in the most exposed occupations −16% (Nov. 2022 → Jul. 2025, US), corroborated by Massenkoff & McCrory (Anthropic) and Hosseini-Lichtinger. In France (Insee), unemployment among 15-24 year-olds +2 pts (19.1% → 21.1%, Q1 2025→Q1 2026) — but with no established causality (economic conditions, overrepresentation in temporary contracts). Issue: automation of the codified tasks that train juniors → risk to entry-level integration and senior renewal.
Long term = an open bet. the massive substitution scenario (if agentic/physical AI becomes widespread — Arquié: >40% of occupations would exceed 30% automatable tasks) VS creative destruction (4-decade review: innovations created more jobs than they destroyed; Autor: 60% of workers hold jobs that did not exist in 1940). Reminder: Frey & Osborne 2013 forecast 47% of US jobs automatable — which did not materialize.
Transition costs. the slowness of reallocation (skills obsolescence, geographic frictions) cuts by 40% the benefit of robotization in France (Bocquet 2026) → the role of bridge occupations and training policies.
Strategic (policy) imperative.investing in AI = protecting employment via competitiveness. France = intermediate position (adoption 18% OECD 2025 vs. EU average 20%, but catching up +8 pts 2024-2025 vs. +6 pts EU). Measures: the "Osez l'IA" plan (2025, 15 million professionals trained by 2030, Académie de l'IA) and France 2030 ("Compétences et métiers d'avenir").
To link. a macro-institutional counterpoint to the productivity/employment cluster of the watch (SFEIR AI4IT [[girard-sfeir-ai4it-vs-ai4business-budgets-2027-2026-06-24]], the announced disappearance of developers [[bfmtv-tech-co-business-ia-developpeurs-disparaissent-2026-05-05]], human+agent headcount accounting [[sternfels-mckinsey-60000-people-20000-agents-officechai-2026-01-14]]). Here, the reading is cautious and aggregate, in contrast to firm-by-firm experience reports.
Key figures
des gains de productivité individuels de +14 % (service client) à +26 % (développeurs)