AI delivers big productivity gains but risks uneven job impacts

Artificial intelligence is accelerating productivity and reshaping the labor market, but a new review warns that its effects will be uneven, benefiting some workers while putting pressure on others. The paper, published as a pre-print on arXiv, examines theory, estimates, and early evidence on how generative AI affects jobs.

The research, titled AI and jobs. A review of theory, estimates, and evidence, consolidates findings from task-based theories of automation, exposure estimates across occupations, and emerging evidence from randomized trials, field experiments, and digital trace data. The study highlights large but uneven productivity gains, an early pattern of substitution in content-related work, and gaps in current research that could mislead policymakers.

How much productivity does AI really deliver?

The authors review experimental evidence to establish how AI affects productivity in practice. Controlled trials show that generative AI can increase output by between 20 and 60 percent, while field experiments suggest gains of 15 to 30 percent. The most consistent effect is that novices benefit more than experienced workers when using AI for simple tasks such as drafting text or summarizing content.

However, the picture is less clear when the tasks are complex. The study introduces a simple classification that distinguishes between simple and complex tasks based on four criteria: knowledge requirements, clarity of goals, interdependence, and the need for context. While AI tends to accelerate basic work, the evidence on whether it improves complex decision-making is mixed. Some studies find that advanced users are able to leverage AI effectively, while others show little to no benefit.

The review stresses that productivity gains are not evenly distributed. High-wage occupations appear to be more exposed to AI, meaning that even well-educated professionals may see parts of their workflow reshaped by automation. This contrasts with earlier waves of automation, which mainly affected routine, lower-wage work.

Where is substitution already visible?

By analyzing digital trace data, the study identifies where human labor is already being replaced. Platforms for writing and translation show signs of substitution, with demand for entry-level workers declining. At the same time, there is a sharp increase in demand for AI-related skills, suggesting that jobs requiring familiarity with these tools are becoming more valuable.

The authors caution that while substitution is evident in some domains, it is not yet widespread across the economy. Instead, what is visible today is a shifting balance between human and machine roles within certain tasks. Early adopters are reconfiguring workflows, allowing AI to handle drafting, summarization, or translation, while humans focus on review and complex problem-solving.

Still, the evidence suggests that novices face higher risk. In fields where AI complements skilled professionals, inexperienced workers may see fewer opportunities as firms opt to rely on AI systems rather than investing in entry-level staff. This raises concerns about career pipelines, particularly in industries where early work experience is essential for skill development.

What gaps remain in current research?

Despite the growing number of studies, the authors note several blind spots. Most research so far has focused on estimating exposure, measuring which occupations are technologically feasible for AI to perform, rather than tracking actual adoption. Adoption rates determine how fast and how far AI will reshape jobs, and these dynamics remain poorly understood.

Experimental studies are heavily weighted toward simple tasks, leaving open questions about how AI affects collaboration, creativity, and team-level outcomes. The review also points out that job creation effects are rarely analyzed, even though new tasks and roles often emerge alongside technological change.

The authors argue that policymakers must move beyond static exposure metrics. To design effective responses, governments need better evidence on who is adopting AI, how workflows are changing, and what long-term effects this will have on wages and employment pathways.