Employment Level

Bottom Line

Employment hasn't collapsed since late 2022. AI is shifting work: routine cognitive roles shrink while complex, interpersonal, and AI-complementary roles grow, with wide differences by sector, firm, and worker group A. T. D.; Zhanabay; Fruits, Stout. Exposure scores from platform logs can be biased unless adjusted to reflect the labor force Yin, Ogut (2026).

What This Means in Practice

• Rebuild roles around human strengths. Emphasize complex problem-solving, interpersonal work, and oversight. Fund upskilling in advanced AI and data skills A. T. D.; Genz (2026); Gohel.
• Pace adoption to your training and mobility capacity. Rolling out faster than you can reskill can lower participation; align timelines with retraining and transition supports Levy Yeyati.
• Measure exposure credibly. Reweight platform-log exposure scores to match workforce shares, and pair exposure metrics with designs that exploit policy or timing changes (natural experiments) or with validated task measures Yin, Ogut (2026); Joshi (2026).
• Build transition infrastructure. Use active labor market programs (training, job placement, wage subsidies), portable benefits (benefits that move with workers), and targeted supports to move under-represented groups into AI-complementary roles A. T. D.; Mishra (2026).
• Set sector strategies. Plan for higher demand for skilled labor in digitalizing manufacturing; in some services, expect short-run displacement unless you redesign work Wang (2026); Ropele; Aoki.

What the Research Finds

Aggregate employment to date: modest net change, visible pockets of decline and entry

• EU unemployment did not surge. In computer-heavy sectors, a natural experiment comparing more- and less-exposed sectors before and after found no significant employment change; remote work and wages rose Zhanabay; Fruits, Stout; Bieliaieva et al. (2026).
• After ChatGPT, coder employment growth slowed beyond what sector shocks explain; online labor markets saw fewer translation jobs and more web development work Crane, Soto; Qiao et al..
• Solo business formation rose following ChatGPT, concentrated among solo founders; top-quality startups remain team-based Kim et al. (2026).
• Measurement choices can flip findings. Changing platform inputs reverses some post-ChatGPT employment estimates, and reweighting platform logs to workforce shares cuts exposure levels by 42–93 percent Yin, Ogut (2026).

Task and occupation reallocation: routine down, complex and "green" up

• Across OECD countries, a one-standard-deviation rise in AI adoption is associated with 2.3 percent lower employment in routine cognitive jobs and 1.8 percent higher in complex problem-solving and interpersonal roles. Adjustments are smoother where active labor programs and portable benefits are stronger A. T. D..
• In Chinese cities, higher AI exposure is associated with more "green" jobs and better energy efficiency, with gains spilling over to nearby areas Li et al. (2026).
• Where AI skills spread faster, employment and wages are higher overall, but highly exposed jobs with little complementarity see lower employment, especially for young workers Jaumotte et al..
• In Germany, jobs using cutting-edge digital skills show a U-shaped link to job growth across occupations Genz (2026).

Firm- and sector-level adoption: productivity gains with divergent employment responses

• Italian firms adopting AI shift from blue-collar toward higher-skill white-collar roles, with little change in total headcount Ropele.
• At the worksite level, AI adoption measured by AI-skill job ads is associated with higher productivity and output but lower employment and total pay, concentrated in non-AI tasks and non-senior roles Aoki.
• In Chinese manufacturing, digital transformation is associated with higher labor demand and stronger demand for highly educated workers, via higher productivity (total factor productivity) and digital skills Wang (2026).
• In a randomized field trial, training startups to map AI into workflows increased tasks completed, customer acquisition, and revenues without changing labor demand Kim, Kim, Koning (2026).
• Firms adopting GitHub Copilot are associated with a higher chance of hiring software engineers, especially entry-level; new hires show broader non-programming skills Baird et al..
• National-accounts work links AI to higher productivity, input savings, and a shift toward younger and relatively less-educated workers, though results depend on timing assumptions Highfill, Samuels.

Platforms, gig work, and entrepreneurship: growth outside traditional firms, with risks

• A Chinese cultural-sector program is associated with higher employment mainly through platform gigs and public procurement, with gains only where digital infrastructure met a threshold Wang et al. (2026).
• Across 24 OECD countries, platform work is a meaningful employment share; reclassifying platform workers as employees is associated with an 18 percent drop in platform labor supply and a 31 percent rise in remaining workers' hourly pay Han.
• Demand for expert-annotated data has created an expert gig market around AI labs, shifting some expertise to managed contract work Wolfe (2026).

Distributional and equity impacts: who benefits and who is left out

• In India, graduates from Scheduled Castes and Tribes are less present in generative-AI-intensive jobs than upper-caste peers in the same district; these jobs pay more, pointing to widening gaps without intervention Mishra (2026).
• Reviews find mixed outcomes for women on AI-mediated platforms: more earning opportunities and flexibility, but algorithmic bias, precarity, and digital exclusion; little evidence yet of durable career gains from supportive AI tools Hattab, Charni (2026); Portell-Fonolla et al. (2026).
• Hybrid human-AI translation for limited-English-proficiency immigrants is associated with better job matching and retention than informal or AI-only options Abdalla et al..
• Young workers see larger employment declines in highly exposed, low-complementarity jobs; in higher-AI regions, older low-educated workers report lower willingness to stay employed as retirement nears Jaumotte et al.; Fu (2026).

What We Still Don't Know

• Headcount effects of agentic, cross-workflow AI (AI agents that take actions across workflows). Task-level gains exist, but credible causal evidence on organization-wide employment is sparse Ooi (2026).
• Generalizable causal estimates in developing and informal economies. Outside advanced economies and a few sectors, much evidence is descriptive or correlational Bimbari (2026); Ly, Ma; Ligot.
• Adoption speed and macro adjustment. Models suggest faster adoption can lower participation by overwhelming retraining capacity; empirical calibration is missing Levy Yeyati.
• Measurement validity. Platform-log exposure indices can mislead unless adjusted and validated; no at-scale, validated task-based exposure measure embedded in official statistics yet Yin, Ogut (2026); Joshi (2026).
• Within-firm labor dynamics from generative tools. Evidence that coding assistants are associated with higher hiring and broader skills is observational; we lack randomized or natural-experiment evidence on sustained headcount, internal mobility, and career ladders Baird et al..
• New since the cutoff: recent natural-experiment studies show mixed reallocation with pockets of displacement (for example, coder slowdown and declines in non-AI tasks at some worksites), alongside net-neutral headcount with role shifts and higher entrepreneurial entry Crane, Soto; Aoki; Ropele; Kim et al. (2026).