Artificial intelligence at work is not merely a productivity story; it is a governance stress test. Unlike earlier waves of automation, which displaced routine, codifiable tasks, contemporary AI reaches into nonroutine, judgment-heavy work—drafting, summarizing, scheduling, retrieving information, maintaining code—the cognitive and service tasks once performed by paralegals, customer support staff, and software developers. And it is doing so across the global workforce, reaching data-labelers in Nairobi as surely as paralegals in New York—which means no single jurisdiction, however well regulated, can govern AI alone. In a new article, we argue that the central question this raises is not whether AI will transform work (which it undoubtedly will), but whether the legal architecture governing it can be put into place fast enough to keep pace with its deployment.

The threat has two layers. The first is structural: AI may automate mid-skill tasks faster than labor markets can generate complementary roles. The result would be a polarized labor market— jobs growing at the high and low ends while the middle hollows out—with productivity gains flowing to capital and high-skill workers. The second, less discussed, is what happens to workers whose jobs persist in altered form: degraded job quality through algorithmic management, intensified surveillance, volatile pay, and opaque discipline and deactivation. The first layer is no longer hypothetical. Stanford’s 2026 AI Index reports that employment for software developers aged 22 to 25 has fallen nearly 20 percent since 2024, even as senior, judgment-heavy roles remain intact—a pattern the report calls “seniority-biased technological change.”¹ Any serious governance response must address both layers.

The Law Already Requires a Lot

Our starting point is that the international legal framework for governing AI at work largely exists. Articles 6 and 7 of the International Covenant on Economic, Social and Cultural Rights (ICESCR) guarantee the rights to freely chosen work and to just and favorable conditions of work,² and the Committee on Economic, Social and Cultural Rights has made clear—in General Comments No. 18 and No. 23—that these are binding duties requiring active state measures, not aspirational ones.³ Applied to AI, these duties mean that states cannot passively allow automation to displace workers. They must fund retraining, guidance, and placement, and they must legislate, inspect, and enforce to curb algorithmic scheduling that cuts rest time, monitoring that heightens stress, and opaque systems that trigger unfair dismissals.

A word of candor for U.S. readers: the United States signed the ICESCR in 1977 but never ratified it, so the Covenant does not bind it as treaty law. American companies are hardly exempt, though: they must comply wherever they operate, ratifying countries are converting the Covenant’s guarantees into enforceable law—the EU’s AI Act⁴ and Platform Work Directive⁵ among them—and domestic employment and discrimination law imposes parallel duties at home. That gap between signature and ratification is exactly why the coupling mechanisms we describe below matter most in the United States.

Corporations share these responsibilities. The UN Guiding Principles on Business and Human Rights require firms to conduct ongoing human rights due diligence and remediate harms.⁶ In the workplace-AI context, that means bias-testing hiring and evaluation systems, ensuring human review and appeal for significant automated decisions, and maintaining effective grievance mechanisms. Algorithmic bias is not inevitable. As the well-documented case of a U.S. health-management algorithm that underestimated the needs of Black patients demonstrated, bias is design-dependent—and design choices are auditable, correctable, and therefore governable.⁷

Complementary Institutional Roles

National responses so far are a patchwork: the EU classifies workplace AI as high-risk and presumes platform workers are employees; the United States relies on state and local rules and, where statutes fall short, union agreements; China regulates content more than worker protections; and Global South jurisdictions like Brazil and South Africa are legislating ambitiously but enforcing thinly. Inconsistency at this scale invites regulatory arbitrage—firms can route AI-intensive operations to the weakest jurisdictions, undercutting any national floor. No single jurisdiction can close that gap alone, which is where the international institutions come in.

At the international level, the institutions’ roles are complementary rather than duplicative. The UN Human Rights Council and the Office of the UN High Commissioner for Human Rights supply the normative baseline, linking workplace AI explicitly to ICESCR Articles 6 and 7 and issuing benchmarks for courts and inspectorates. The International Labour Organization (ILO) provides the standard-setting muscle: Its platform-work discussions, opened at the 2025 International Labour Conference, are on track for a Convention and Recommendation, slated for adoption in 2026, that would elevate protections such as presumed employment under algorithmic control and human oversight to global standards—and activate the ILO’s supervisory machinery.⁸ The WTO, without becoming a labor agency, can sequence digital-trade commitments with labor-impact assessments and transition finance so diffusion does not outpace states’ capacity to protect workers. And the International Trade Union Confederation and global unions supply “ground truth”: they collect evidence from workers—unpaid waiting time, accounts deactivated without explanation, safety incidents—and feed it to the international bodies that monitor compliance, while spreading model AI contract language from one country’s unions to another’s, which functions as regulation before any statute is passed.

Because even well-governed AI will displace and reorganize tasks, the governance mix must also include guaranteed minimums of income security and health care for all workers—what the ILO calls “social-protection floors”⁹ —and a “Just Transition for AI” facility: an idea borrowed from climate policy, funding wage-loss insurance, rapid re-employment services, and training for the jobs that complement AI rather than compete with it.

This layered architecture turns abstract guarantees into routines: notice to workers before an AI system is deployed, pace limits written into workplace safety plans, appeal rights backed by income protection, and funded retraining for those displaced. The lesson for lawyers and policymakers is that the AI transition is a design problem, not a legislative one: the law we need largely exists. Soft law supplies methods and metrics; hard law supplies duties, remedies, and sanctions. The task is to combine them by design rather than by chance so that productivity gains arrive with enforceable rights, worker voice, and social protection.

Labor-Rights Impact Assessment

Our central contribution is an architecture for coupling the two bodies of law that govern AI at work: hard law (the ICESCR, binding ILO conventions, national labor codes) and soft law (OECD and G20 AI principles, UN guidance, corporate codes). A system of many rule-makers can work, but only when concrete coupling mechanisms—what we call “knitting hooks”—keep soft law tied to enforceable duties rather than substituting for them.

The operational core is a Labor-Rights Impact Assessment (LRIA): a structured, recurring assessment—defining the baseline before rollout, reviewing at six months and then annually thereafter—that employers conduct before deploying AI in hiring, evaluation, scheduling, pay, discipline, safety, or termination. It tracks fixed, measurable indicators (disaggregated error rates for nondiscrimination, wage dispersion and scheduling volatility for job quality, pace-of-work and psychosocial metrics for occupational safety) and is tied to remedies, including individual rights to notice, explanation, and appeal.

Four hooks then bind the LRIA to hard law. First, incorporation by reference: Statutes recognize LRIA templates only when tied to enforceable remedies such as reinstatement, back pay, and penalties—so conducting the assessment becomes part of complying with the law, not a substitute for it. Second, comply-or-explain with a public registry: Employers publish LRIA summaries against a common standard or publicly justify deviations, putting a company’s workforce-AI practices where investors, regulators, and plaintiffs’ lawyers can read them. Third, contract hooks: Procurement rules, licenses, and collective bargaining agreements convert the LRIA’s metrics into binding obligations with audit rights—meaning a vendor that cannot document its assessment loses the contract. Fourth, enforcement presumptions: A missing or noncompliant LRIA creates a rebuttable presumption of unlawfulness in disputes over dismissal, pay, or scheduling—shifting the burden to the employer precisely where records are in its hands.

None of this requires a new institution or a new agency; every hook runs through levers that legislatures, regulators, general counsel, and unions already operate. And that is the practical point for corporate counsel and compliance officers: The LRIA is not a future regulatory burden but a present opportunity. Run before deployment and documented well, it is evidence of diligence—in litigation, before regulators, and in the procurement processes that increasingly demand it. Absent or perfunctory, it is becoming evidence of the opposite. The companies that build the assessment into their AI governance now will set the standard the rest are eventually measured against.

REFERENCES

1. Stanford Institute for Human-Centered Artificial Intelligence, The AI Index 2026 Annual Report, ch. 4 (Economy) (2026).
2. International Covenant on Economic, Social and Cultural Rights, arts. 6–7, Dec. 16, 1966, 993 U.N.T.S. 3.
3. U.N. Comm. on Econ., Soc. & Cultural Rights, General Comment No. 18: The Right to Work, U.N. Doc. E/C.12/GC/18 (2006); U.N. Comm. on Econ., Soc. & Cultural Rights, General Comment No. 23: The Right to Just and Favourable Conditions of Work, U.N. Doc. E/C.12/GC/23 (2016).
4. Regulation (EU) 2024/1689 of the European Parliament and of the Council (Artificial Intelligence Act), 2024 O.J. (L 1689).
5. Directive (EU) 2024/2831 of the European Parliament and of the Council on Improving Working Conditions in Platform Work, 2024 O.J. (L 2831).
6. U.N. Guiding Principles on Business and Human Rights, U.N. Doc. HR/PUB/11/04 (2011).
7. Ziad Obermeyer, Brian Powers, Christine Vogeli & Sendhil Mullainathan, Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations, 366 Science 447 (2019).
8. Int’l Labour Org., Standard-Setting Committee on Decent Work in the Platform Economy, Int’l Labour Conf., 113th Sess. (2025).
9. Int’l Labour Org., Recommendation No. 202: Social Protection Floors (2012).

Michael A. Santoro is a professor at Santa Clara University’s Leavey School of Business, and Brewer D. Stone is a graduate of the University of St. Andrews (Scotland) and a fellow of the AI, Ethics, and Human Rights Lab at Santa Clara University’s Leavey School of Business. This post is based on their recent article, “Artificial Intelligence, Human Rights, and the Future of Work: Global Governance and International Organizations in the Age of AI,” published in the Journal of Human Rights and available here. An open-access version is available on ResearchGate.