In the United States, AI regulation at work is advancing without waiting for Washington

According to data from Jobscan, 97% of Fortune 500 companies now use an ATS — an applicant tracking system — to filter applications. This figure took ten years to become commonplace. The law took only eighteen months to begin responding — but not from the capitals one might have expected.

While Europe debates the AI Act and the American federal administration still searches for its direction, it is American states that have moved first. Illinois classified discriminatory use of AI in employment among civil rights violations as of January 1, 2026. Connecticut is preparing an obligation as of October to report mass layoffs linked to automation. A dozen other states have filed similar proposals since 2024. This is not a coordinated strategy. It is an upward, fragmented pressure that nonetheless produces concrete law.

The essentials

  • According to SHRM, 25% of organizations used AI in HR in 2024, a rate reaching 43% according to 2025 data; 97% of Fortune 500 companies use an ATS (conventional applicant tracking system), according to Jobscan
  • Illinois integrated discriminatory use of AI in employment into its civil rights law, effective January 1, 2026
  • Connecticut SB 5 requires employers to report mass layoffs caused by AI beginning October 2026
  • A federal Executive Order from December 2025 prepares a preemption doctrine that could erase part of this structure
  • The European AI Act, more coherent in its design, will take effect for high-risk systems no earlier than late 2026, leaving an operational void that American states are beginning to fill

When AI decides who gets hired, the question becomes political

The recruitment algorithm is no longer a decision-support tool. In most large American companies, it has become the first filter — sometimes the only one. Platforms like HireVue or Pymetrics analyze facial micro-expressions, tone of voice, responses to cognitive games to produce a compatibility score. Candidates no longer speak to a recruiter during the first interview. They speak to a camera.

The problem is not theoretical. Independent audits, notably those conducted by the Algorithm Watch organization and the work of professor Ifeoma Ajunwa at the University of Georgia, have documented systematic biases in these systems: Black and Hispanic candidates receive lower scores for equivalent résumés, women are disadvantaged in male-dominated positions, candidates over 50 are filtered before even reaching a human manager. This is not a technical malfunction. It is the automated reproduction of historical biases in the labor market, at unprecedented speed and scale.

It is precisely this mechanism that Illinois chose to address head-on. The law amending the Illinois Human Rights Act — in effect since January 1, 2026 — qualifies as unlawful discrimination the use of an AI system that produces disparate effects on groups protected by law, even in the absence of discriminatory intent. The notion of disparate impact is borrowed from American civil rights jurisprudence, which has recognized since the 1971 Griggs v. Duke Power decision that discrimination can be measured by its effects, not by its intentions. Illinois simply extended this principle to the algorithmic domain.

The burden of proof now falls on the employer. They must demonstrate that the tool they use is validated for the position in question, that it has been tested for disparate impact, and that less discriminatory alternatives did not exist. This reversal is significant: it is no longer sufficient to say “the algorithm is neutral.” You must prove that it is.

Connecticut invents a right of alert for algorithmic layoffs

Where Illinois works on employment access, Connecticut tackles the exit. Senate Bill 5, adopted in May 2026 and signed by Governor Lamont on May 27, 2026, is applicable beginning October 2026. It requires companies to notify the state when they carry out mass layoffs whose cause is linked to the introduction of an automated system. This notification obligation already existed for factory closures or massive restructuring, under the 1988 federal Worker Adjustment and Retraining Notification Act. Connecticut extends this scope to automation.

The stated objective is twofold. First, to give state employment agencies sufficient notice to prepare retraining programs. Second, to build a public database on the real impact of AI on workforce numbers, in the absence of any reliable federal statistics on the subject. The American Bureau of Labor Statistics does not have a specific category for layoffs linked to automation. Connecticut’s decision amounts to creating this data where it did not exist.

This type of measure is less spectacular than banning discrimination. Its effect is however potentially more structuring. Data transparency is the prerequisite for any serious public policy on AI’s impact on employment. You cannot correct what you do not measure. The question of invisible precarity in labor market transitions arises in the same terms in France: the absence of granular data on the real causes of contract terminations prevents any serious diagnosis.

Several other states have followed analogous trajectories. California adopted rules in 2023 requiring employers to notify employees of the use of automated systems in decisions affecting them. New York City has had an obligation since January 2023 for annual audits of recruitment AI tools, with publication of results. Maryland’s AI Act of 2024 prohibits the use of facial analysis in job interviews without explicit candidate consent. The American map of AI regulation at work now resembles an archipelago: discontinuous, unequal, but inhabited.

Why states moved before Congress

The short answer: because Congress has been incapable for ten years on technology matters. The longer answer is more interesting.

American states have a long tradition of labor legislation ahead of the federal level. California invented discrimination laws before the 1964 federal Civil Rights Act generalized them. Seattle’s minimum wage preceded any serious congressional discussion by ten years. This pattern repeats. When the federal government is blocked, states experiment. When experiments succeed, they migrate.

There is also direct pressure from workers and their organizations. American unions, weakened since the 1980s, have found in AI regulation a terrain for partial reconquest. The AFL-CIO made algorithmic transparency a lobbying priority from 2022 onward. The Service Employees International Union negotiated contractual clauses on AI in several sectors. This organized pressure is not sufficient to produce federal law in a divided Congress, but it is sufficient to trigger legislative initiatives in Democratic-leaning states.

The result is regulation with variable geometry. A candidate for a position at a company based in Chicago benefits from protections that their counterpart in Houston does not have. This asymmetry is not unlike that observed in other sectors where regulation follows technology with a lag: law is built where political pressure is sufficient, not where the problem is most intense.

The December 2025 Executive Order: the federal government takes back control, or tries

This is where the story becomes complicated. In December 2025, the Trump administration signed an Executive Order on AI that includes a preemption provision: the federal government reserves the right to neutralize state AI legislation deemed incompatible with the federal doctrine of promoting innovation. The wording is deliberately vague, but the intention is legible. Washington wants to take back control of a file it long ignored.

The preemption doctrine is a classic constitutional tool: by virtue of the Supremacy Clause of the American Constitution, a federal law takes precedence over a state law in its domain. Its application to AI would be unprecedented and legally contestable. The laws of Illinois and Connecticut rest on foundations of labor law and civil rights that historically fall under state jurisdiction. A federal preemption in these domains would require either a congressional law — difficult to obtain — or an executive doctrine whose legality would immediately be contested before courts.

Law firms that have parsed the Executive Order, notably K&L Gates and White & Case, estimate that the threat of preemption is real but its execution uncertain. Companies that have built their compliance on state laws find themselves in an uncomfortable position: they do not know whether the regulatory framework in which they invested will hold. This type of uncertainty is precisely what economic actors detest most — sometimes more than the constraint itself.

There is an irony in the situation. The administration that claims to defend American competitiveness against European regulation is itself producing regulatory uncertainty that the most serious companies regard as a risk. A legal director at a major technology company must today ask themselves whether their compliance with the Illinois IHRA will remain valid in eighteen months.

What Europe is watching from Brussels

The comparison with the European AI Act is inevitable, but it deserves to be nuanced.

The AI Act adopted in 2024 is a text of primary law, adopted at the Union level, with a legal coherence that the American patchwork will never be able to achieve. AI systems used in recruitment and human resource management are classified there as “high-risk,” which implies obligations of transparency, auditing, registration in a public database, and human supervision. This framework is more comprehensive than any American state law taken in isolation.

But the AI Act will not take effect for high-risk systems until late 2026 at the earliest — and some obligations will not apply until 2027. In the meantime, millions of job interviews have been filtered by unaudited algorithms in all member countries. The question is not whether European legal architecture is superior to the American mosaic. It probably is. The question is what happens in the operational void created by the slowness of implementation.

There is also a difference in mechanism worth noting. The AI Act is a product regulation: it applies to suppliers and deployers of AI systems. The laws of Illinois and Connecticut are outcome regulations: they apply to concrete effects on workers, regardless of the technical nature of the system used. These two logics are not mutually exclusive. They complement each other. An employer can use a system certified as compliant with the AI Act and nonetheless produce discriminatory effects. Outcome regulation captures what product regulation does not see.

This may be the most useful lesson for European policymakers. Not “let us do as the Americans do,” but “let us look at what their experimentation reveals about what our approach does not cover.”

What companies are doing while they wait

Facing this normative uncertainty on both sides of the Atlantic, the most exposed companies are not remaining idle. A trend is emerging: large international companies are converging toward the most demanding standard, not out of virtue, but from pragmatism. Applying the most constraining rule across all their operations is less costly than maintaining differentiated processes by jurisdiction.

IBM published in 2024 an internal policy on recruitment AI that incorporates stricter audit obligations than what current American regulation requires. SAP, which deploys HR tools in dozens of countries, has integrated bias detection mechanisms into its products to anticipate both the AI Act and American state requirements. These choices are not philanthropic. They reduce legal risk and facilitate commercialization in regulated markets.

There is something in this dynamic analogous to what occurred with GDPR: European regulation that American companies initially feared transformed into a de facto standard for their global operations, because the cost of dual compliance exceeded the cost of alignment upward.

AI in human resources may be following the same path. The laws of Illinois and Connecticut will never apply to a German company recruiting in Munich. But American companies operating in Europe will apply the most demanding standards to their entire HR workflow. And suppliers of HR solutions — Workday, Oracle, SAP, Greenhouse — will adapt their products for the most regulated market, which will mechanically become the standard product.

The question that remains open is one of speed. Illinois acted. Connecticut acted. Ten other states have projects underway. The federal Executive Order could block some of this momentum, or be overturned by courts. The European AI Act will provide within two years a more comprehensive framework, if its implementation follows the planned timeline. Between these two movements, millions of recruitment and layoff decisions will be made by algorithms whose biases no one has verified.

This is not a fatality. It is a choice of rhythm — and the actors who choose to move faster than their peers are already at work.

Sources

  1. K&L Gates / White & Case, “AI Employment Law: State-by-State Tracker”, Lexology, 2025-2026
  2. Illinois Human Rights Act, amendments effective January 1, 2026 (Illinois General Assembly)
  3. Connecticut SB 5, adopted May 2026, applicable October 2026 (Connecticut General Assembly)
  4. Society for Human Resource Management (SHRM), annual report on AI use in HR, 2024
  5. Federal Executive Order on AI, December 2025 (Federal Register, White House)
  6. Algorithm Watch, audits of algorithmic recruitment systems, 2022-2024
  7. New York City Local Law 144, rules on automated employment decision support tools, effective January 2023 (NYC Commission on Human Rights)
  8. SHRM 2024 Talent Trends: AI in HR (primary report)
  9. Illinois HB 3773 – National Law Review
  10. Connecticut SB 5 – Law and the Workplace
  11. Executive Order 14365 – White House (official primary source)
  12. EU AI Act – European Commission (primary source)
  13. White & Case – EO 14365 and preemption of state laws
  14. K&L Gates – AI Governance and preemption (May 2026)
  15. University of Washington – AI bias in resume screening (2024)
  16. Latham & Watkins – AI Act Omnibus deadline update