Redistributed

General

AI Executive Orders: What Changed and What Didn't

Published March 11, 2026

policyregulationusexecutive-ordersgovernance

In October 2023, President Biden signed the most sweeping federal action on AI the United States had ever produced. Fifteen months later, President Trump rescinded it. This sequence tells you almost everything you need to know about the fragility of American AI governance — and about who benefits when regulation arrives by executive order rather than legislation.

The real story isn't about the policies themselves, though they matter. It's about the structural gap: the United States still has no durable AI law. Every protection the federal government has extended to workers, consumers, and communities affected by AI systems can be withdrawn by a single signature. That is not a bug in the system. For the largest AI companies, it is a feature.

What Biden's Executive Order Actually Did

Executive Order 14110, signed on October 30, 2023, was a 111-page document that attempted to do what Congress had not: create a federal framework for AI oversight. It was sprawling, ambitious, and built on a creative legal maneuver — invoking the Defense Production Act to compel reporting from AI companies (White House, 2023).

The key provisions fell into several categories:

Safety and security. Companies developing "dual-use foundation models" — systems with potential military or intelligence applications — were required to notify the federal government and share the results of safety testing before releasing them. The threshold was set at models trained using more than 10^26 floating-point operations, or models trained on primarily biological sequence data using more than 10^23 operations. The National Institute of Standards and Technology was tasked with developing standards for red-teaming and safety evaluations.

Transparency and watermarking. Federal agencies were directed to develop standards for authenticating AI-generated content and detecting synthetic media. The Commerce Department was tasked with developing watermarking guidelines — a technically difficult problem that remains largely unsolved.

Worker protections. The order directed agencies to study AI's impact on the labor market and develop principles for employers deploying AI in ways that affect workers. The Department of Labor was tasked with guidance on best practices for employers using AI, including protections against AI-enabled surveillance and ensuring that AI complements rather than displaces workers.

Equity and civil rights. Federal agencies were required to assess how their AI use might affect equity and civil rights. The Department of Justice received direction to address algorithmic discrimination in the criminal justice system. The Department of Health and Human Services was tasked with developing safeguards against AI-driven discrimination in healthcare.

Federal talent pipeline. The order launched a government-wide AI talent surge, streamlining hiring for AI professionals and creating new fellowship programs. This was one of its more pragmatic provisions — the government's ability to oversee AI depends on whether it can hire people who understand it.

What the Order Got Right

The Biden EO made a genuine structural contribution by naming the problem clearly: AI systems affect real people, and someone needs to be checking the work. Its use of the Defense Production Act was a creative workaround for Congressional inaction — the DPA had been designed for wartime industrial mobilization, not technology regulation, but it gave the executive branch a legal hook to compel reporting from private companies.

The NIST standards work was particularly important. NIST's AI Risk Management Framework, already in development, provided a common vocabulary for evaluating AI systems. As Meredith Whittaker, president of Signal and former co-lead of the AI Now Institute, has argued, the fundamental question in AI governance is whether we have the institutional capacity to evaluate these systems at all (Whittaker, 2023). NIST's work began building that capacity.

The equity and civil rights provisions acknowledged something that the technology industry's preferred framing — innovation, competitiveness, national security — systematically obscures: AI systems redistribute power, and that redistribution has winners and losers. When a hiring algorithm screens candidates, when a healthcare algorithm allocates resources, when a criminal justice algorithm assesses risk, someone's life changes. The Biden order, whatever its implementation gaps, put that reality on paper.

What Trump's Revocation Changed

On January 20, 2025, President Trump signed an executive order revoking Biden's AI EO in its entirety. The replacement framework, articulated through a subsequent executive order titled "Removing Barriers to American Leadership in Artificial Intelligence," took a sharply different approach (White House, 2025).

Where Biden emphasized guardrails, Trump emphasized acceleration. The new framework prioritized:

  • Deregulation: Removing what the administration characterized as burdensome reporting requirements
  • Competitiveness: Framing AI governance primarily through the lens of international competition with China
  • Industry self-governance: Relying on companies to police themselves rather than mandating federal oversight
  • Energy infrastructure: Directing attention to the power and data center capacity needed for AI development

The rhetorical shift was total. Safety, equity, and worker protection essentially disappeared from the federal AI vocabulary.

What Actually Changed on the Ground

Less than you might think — and more than the policy documents suggest.

Many of Biden's directives were still being implemented when they were rescinded. Federal agencies were mid-process on rulemaking, guidance development, and hiring initiatives. Some of this work simply stopped. Other parts continued under existing statutory authority that doesn't depend on executive orders.

NIST's AI safety standards work, for instance, continued. The AI Risk Management Framework had its own institutional momentum, and NIST's standards development process involves extensive industry participation — the same companies that lobbied against mandatory reporting were happy to participate in voluntary standards-setting, where they could influence the outcome.

Federal agencies continued adopting AI systems, but the equity assessments that would have accompanied those deployments were no longer required. The practical effect: the government kept buying and deploying AI, but stopped asking who might be harmed by it.

The talent pipeline initiatives largely stalled. Government AI hiring, already difficult given private sector compensation, lost its executive-level champion.

The Structural Problem: Executive Orders as Policy

Here is the core issue: an executive order is not a law. It is a directive from the president to the executive branch. It requires no Congressional vote, no committee markup, no floor debate. It can be signed in an afternoon and undone the same way.

This means that every protection in Biden's AI EO was, structurally, temporary. Anyone paying attention — and the major AI companies were paying very close attention — knew this. The provisions existed on a four-year or eight-year clock. Investment decisions, product roadmaps, and corporate strategies could be planned accordingly.

Compare this to the EU's approach. The AI Act is legislation. It was debated for three years, amended extensively, and passed through the European Parliament, the Council, and trilogue negotiations. It is flawed and in places cumbersome. But it cannot be undone by a single person on their first day in office. That durability matters. Companies build compliance infrastructure around durable rules, not transient ones.

The United States Congress has introduced dozens of AI-related bills. None has passed. The bipartisan Senate AI working group, led by Majority Leader Schumer, produced a roadmap but no legislation. The structural incentives are clear: passing a law means taking a position, and taking a position means risking both industry opposition and constituent backlash.

States Filling the Vacuum

In the absence of federal action, states have moved. The patchwork is uneven, but it's real.

California's SB 1047 proposed safety requirements for large AI models, including a "kill switch" and mandatory safety testing. Governor Newsom vetoed it in September 2024, arguing it could drive AI companies out of California. But the bill's influence exceeded its fate — it forced a public debate about model-level safety obligations and demonstrated that state legislatures were willing to act where Congress would not.

Colorado passed the first comprehensive state AI discrimination law in 2024, requiring "deployers" of high-risk AI systems to conduct impact assessments, notify consumers, and provide avenues for appeal. It applies to AI systems used in consequential decisions about employment, education, financial services, healthcare, housing, and insurance.

Illinois enacted requirements around AI in hiring, mandating employer disclosure when AI is used to evaluate candidates and requiring consent for AI-analyzed video interviews.

These state-level actions create a fragmented regulatory landscape. And here the redistribution dynamics sharpen. Large companies — Google, Microsoft, Meta, Amazon — can afford teams of lawyers and compliance officers to navigate fifty different state regimes. A startup with twelve employees cannot. Regulatory fragmentation, in practice, functions as a barrier to entry. The companies that lobbied hardest against federal standards may find that the absence of those standards protects their market position more effectively than any patent portfolio.

This pattern is not new. As the history of regulatory capture demonstrates, incumbents often prefer complexity to clarity, because complexity rewards scale.

Who Benefits, Who Pays

The current arrangement — no federal AI law, a revoked executive order, a patchwork of state rules — produces clear winners and losers.

Who benefits: Large AI companies that prefer self-regulation. They have the resources to manage a fragmented regulatory landscape, the lobbying capacity to shape voluntary standards, and the market power to set de facto norms. The absence of federal equity requirements means they face no mandatory assessment of how their systems affect vulnerable populations. As Langdon Winner wrote in his foundational essay on technological politics, "the issues that divide or unite people in society are settled not only in the institutions and practices of politics proper, but also in the tangible arrangements of steel and concrete, wires and semiconductors" (Winner, 1980).

Who pays: Workers whose jobs are restructured or eliminated by AI systems without the transition support Biden's order contemplated. Consumers subjected to AI-driven decisions — in credit, insurance, healthcare, hiring — without transparency requirements. Communities, particularly low-income communities and communities of color, that bear disproportionate exposure to AI systems deployed by government agencies. And smaller AI companies that cannot afford the compliance complexity that regulatory fragmentation creates.

The question of who regulates foundation models remains essentially unanswered in the United States. The models exist. They are deployed in consequential settings. And the regulatory architecture is a whiteboard that gets erased every four years.

What Could Fill the Gap

Three paths forward exist, and they are not mutually exclusive.

Congressional legislation. The most durable option and the least likely in the near term. Bipartisan interest exists — Senators Blumenthal and Hawley, among others, have introduced bills — but the combination of industry lobbying, partisan gridlock, and genuine technical complexity makes passage uncertain. The most plausible legislative path may be narrow and sectoral: AI in healthcare, AI in hiring, AI in financial services, rather than a single comprehensive act.

Agency-level rulemaking. Federal agencies retain existing statutory authority that doesn't depend on executive orders. The FTC can pursue unfair or deceptive AI practices under Section 5 of the FTC Act. The EEOC can address AI-enabled discrimination under Title VII. The FDA regulates AI-based medical devices. This piecemeal approach lacks coherence, but it has the advantage of existing legal infrastructure and does not require new legislation. It also represents something that is genuinely working: agencies like the FTC have already brought enforcement actions related to AI, demonstrating that existing law is not powerless even without new legislation.

State convergence. If enough states pass similar AI laws, the practical effect approaches a national standard. Industry may eventually prefer federal preemption — a single federal law that overrides the patchwork — to managing fifty different regimes. This is, ironically, the dynamic that eventually produced federal privacy legislation discussions: state action creating enough compliance pain that companies asked Congress to intervene.

None of these paths is fast. None is clean. And in the meantime, AI systems are being deployed in hiring decisions, medical diagnoses, credit scoring, criminal sentencing, and a thousand other contexts where the stakes are high and the oversight is thin.

The executive order era of AI policy demonstrated two things simultaneously: that the federal government understands the stakes, and that it has chosen the flimsiest available tool to address them. Until Congress acts, every protection is provisional, every standard is voluntary, and the question of who AI serves will be answered by the companies building it — not by the public living with the consequences.

Sources

← Back to all articles
On this page