Does AI open doors for workers to move into new roles, or lock them into a steeper ladder, and who decides?

This explores what decides whether AI lets workers move into new tasks and roles or locks them into a steeper ladder, and how much of that is set by firms, incentives, and governance rather than by the technology. The corpus has no notes on unions, licensing, or training policy directly. It does say clearly that the technology doesn't set the direction. One interdisciplinary review across information, work, education, and healthcare found generative AI can both worsen and reduce inequality, and that access, integration, and incentive structures decide which happens (Does generative AI inevitably worsen or reduce inequality?).

The clearest mechanism for mobility is how jobs are bundled into tasks. When AI exposure is spread across most of a job's tasks, labor demand falls. When it is concentrated in a few tasks, workers can shift to the tasks that weren't displaced, and net employment effects stay modest (Does concentrated AI exposure enable workers to adapt and reallocate?). Whether a job looks like the second case depends on how the employer has divided the work, which is an organizational choice. The same choice shows up between firms. Firms with more AI exposure replace online-marketplace freelancers with AI faster and more cheaply than others do, which suggests returns to scale in building internal AI capability (Do firms substitute labor for AI at different rates?). The gains go to the organization that has the capability, and the workers outside its walls are cut first.

Access to the ladder is also uneven. Workers have handed AI structured tasks mainly in information-intensive jobs, following what the technology can do rather than the old routine-versus-non-routine split (Where have workers actually delegated tasks to AI?). Vacancy data from ten countries show AI skill demand clustering in STEM roles around Python, SQL, and machine learning, while non-technical occupations drift away from that core instead of moving toward it (Is AI creating common skills across jobs or deepening divisions?). So there is no shared ramp. Moving up means crossing a gap between the technical core and everyone else. Exposure also falls unevenly on people: in female-dominated occupations it spreads across all skill and wage levels, so lower-paid, lower-skilled women are exposed while having the fewest resources to adapt (Does AI exposure hit low-wage workers harder in some fields?).

There is also a slower institutional effect. Societies stay aligned with human interests partly because their systems depend on human workers who care about outcomes. As AI replaces that dependence step by step, that leverage weakens, and misalignment across interlocking institutions could become hard to reverse (Does incremental AI replacement erode human influence over society?). Worker mobility and worker bargaining power draw on the same source. Markets can also work against humans: in a partner-selection game, AI agents were penalized at first when their identity was disclosed. Over repeated rounds they were chosen more, because they returned more points with less variance (Do humans learn to prefer AI partners over time?). Reputation systems that reward reliability can wear down a human premium without anyone deciding to.

Two rule-setting layers determine who ends up in charge. Once agents buy, deploy, and transact, the binding constraint is coordination infrastructure: identity, delegation, attestation, and audit trails, more than raw reasoning ability (Does agent capability matter more than coordination infrastructure?). Whoever builds that layer decides whether humans stay accountable supervisors or get routed around. The other layer is measurement. Agents ace benchmark contests but fail long-horizon professional workflows, because the field optimized what it measured, and it measured contests rather than work (Why do agent benchmarks not predict real economic value?). Displacement forecasts built on those benchmarks may overstate what agents can do today, and what gets measured steers what gets built.