these structural dynamics will depend on whether countries adopt and enforce the required policies and on effective international coordination that does not yet exist. AI can concentrate power in the hands of capital owners at the expense of workers The development of AI systems has created new employment opportunities for millions of relatively low-skill workers in developing countries—from data labeling for AI training to content moderation. These opportunities share a common feature that economists recognize as a source of market failure: A few large companies are the main source of jobs for many workers in the AI industry, giving these employers extra power over determining pay and working ­conditions. This phenomenon is what economists call a ­monopsony. Meanwhile, workers are being m ­ anaged by algorithms; human managers are absent, for the most part. This type of algorithmic management at scale, combined with the monopsony powers of employers, results in AI companies paying wages below workers’ additional contribution to total value,47 working conditions that workers have little ability to negotiate, and a structural difficulty in workers organizing collectively. The data labeling industry is a clear example. A few companies, such as Scale AI, Sama, CloudFactory, and Appen, hire workers from a large global pool of workers. High global labor supply combined with concentrated buyer-side demand drives down wages. A 2023 investigation in Kenya found that some data labelers earned between $1.32 and $2.00 per hour, barely above Kenya’s minimum wage of about $1.20 per hour, for psychologically and emotionally challenging work that included reviewing graphic content.48 In addition to low wages, workers must contend with sudden shutdowns of the platform, late or withheld payments, restrictions on workers organizing, and contracts designed to place legal liability in jurisdictions that favor the platforms. Fairwork Cloudwork Ratings, which evaluates platforms on five areas of fair work, found that none of the 16 platforms reviewed met basic standards of fair pay, working conditions, contracts, management, or worker representation. The same AI systems that require labeled data to operate also depend on humans to filter out the most disturbing content. Content moderators and data workers who screen for disturbing content carry psychological costs that are well documented. A 2023 lawsuit by Kenyan content moderators against Meta and Sama found diagnoses of posttraumatic stress disorder (PTSD), depression, and anxiety among workers exposed to traumatic content without proper psychological support.49 Much of this work is outsourced to developing countries,50 where workers typically lack access to mental health services and other protections that are necessary for such disturbing work. This situation poses a difficult labor market problem: In a competitive market with sufficient information and bargaining power, workers taking on such risks would receive higher wages and structured psychological support. The combination of monopsony, a large global labor supply, and outsourcing through jurisdictions with weak labor protections has prevented those conditions from emerging. They persist, at least in part, because AI companies can shift the offshore locations for this work across countries at a relatively low cost. Beyond the firms that are involved in the value chain of AI, businesses across the economy can potentially deploy AI to exercise greater control over workers. Algorithmic management is not inherently worse than human management; in some respects it can be fairer because consistent rules applied uniformly are easier to scrutinize than the discretionary and sometimes arbitrary decisions of a human supervisor.51 The concern is not the algorithm itself but the conditions under which it operates: Workers cannot observe how decisions are made, have limited means to contest those decisions, and often face a single buyer with little ability to switch employers. AI’s Political Impact: Reshaping Power Within and Across Countries 213