Redistributed

General

Doc Review: The White House AI Executive Order

Published March 11, 2026

doc-reviewpolicyusexecutive-orderregulation

Document: Executive Order 14110 — "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence" (October 30, 2023) Source: Federal Register

On October 30, 2023, President Biden signed the most comprehensive federal action on artificial intelligence in US history. Eighty-seven days into the next administration, President Trump revoked it.

The lifecycle of Executive Order 14110 — from ambitious policy framework to rescinded document — tells you nearly everything you need to know about how the United States governs AI. It also demonstrates why executive action, no matter how well-intentioned, is structurally incapable of addressing AI's redistribution questions on its own.

What This Document Is

Executive Order 14110 is a presidential directive spanning over 20,000 words across eight major sections. It touches safety, security, privacy, equity, civil rights, consumer protection, worker support, innovation, competition, and international leadership. In scope, it attempted to be the US equivalent of the EU AI Act — a comprehensive governance framework for artificial intelligence.

There is a critical difference. The EU AI Act is legislation, passed by elected bodies, enforceable by law, requiring formal repeal. An executive order is a directive from one person's desk. It binds federal agencies, not private companies. It lasts exactly as long as the sitting president wants it to.

Biden's team knew this. The EO was designed as a bridge — a way to begin the machinery of AI governance while Congress failed to pass AI legislation. It used existing statutory authority, primarily the Defense Production Act, to create reporting requirements for the most powerful AI systems. The idea was to establish norms, build institutional capacity, and create facts on the ground that a future administration would find difficult to reverse.

That bet did not pay off.

What It Claims

The EO opens with eight guiding principles. AI should be safe and secure. It should promote responsible innovation and competition. It should support workers. It should protect civil rights, consumer privacy, and equity. The US should lead internationally on AI governance. Government should use AI responsibly.

These principles are unobjectionable to the point of near-meaninglessness. No one publicly opposes "safe and secure AI" or "responsible innovation." The question is always what comes after the principles — what mechanisms exist to enforce them, what happens when they conflict with each other, and who decides.

The EO's answer: federal agencies decide, through a cascade of reports, standards, guidance documents, and pilot programs. Over 50 distinct actions were assigned across more than a dozen agencies, each with specific deadlines. The Commerce Department, the Department of Energy, the National Science Foundation, the Department of Homeland Security, the Department of Health and Human Services — all received mandates.

What's Actually In It

The EO's provisions fall into several categories, each with different levels of ambition and enforceability.

Safety Testing and Reporting

The headline provision: companies developing foundation models trained using more than 10^26 floating-point operations (FLOPS) must notify the federal government and share the results of safety tests. This threshold was designed to capture only the most powerful models — at the time of signing, roughly GPT-4 class and above.

The legal authority for this requirement came from the Defense Production Act of 1950, a Cold War-era law that gives the president broad powers to direct industrial activity in the interest of national defense. Using the DPA for AI regulation was legally creative and politically controversial. It framed frontier AI development as a national security matter — a characterization that some applauded and others saw as a dangerous precedent for government involvement in technology development.

The EO also directed the National Institute of Standards and Technology (NIST) to develop standards for red-teaming — adversarial testing designed to find dangerous capabilities before public release. NIST was tasked with creating frameworks for AI safety evaluation, developing guidelines for content authentication and watermarking, and establishing best practices for managing AI risk. This institutional investment is arguably the EO's most durable legacy — NIST's frameworks continue to influence how labs document safety (as visible in Claude's Model Card and Gemini's Safety Documentation).

Privacy

The EO called for advancing privacy-preserving technologies, including research into cryptographic tools that allow AI training on sensitive data without exposing individual records. It directed agencies to evaluate how their data collection practices interact with AI systems, and to strengthen guidelines on federal use of commercially available personal data.

What it did not do: create any new privacy rights, establish any data protection standards with the force of law, or restrict how private companies collect and use data for AI training. The US still has no comprehensive federal privacy legislation. The EO acknowledged this gap without closing it — a gap that OpenAI's Terms of Service and similar corporate policies fill with terms favorable to platforms. State-Level AI Regulation has partially stepped in where federal action has not.

Equity and Civil Rights

Section 7 directed federal agencies to prevent AI from exacerbating discrimination. Specific mandates included using AI to address housing discrimination and unfair lending practices, developing best practices for AI use in the criminal justice system, and examining algorithmic discrimination in healthcare.

The Department of Justice was tasked with addressing civil rights violations related to AI. The Department of Housing and Urban Development was directed to examine algorithmic tenant screening. The Department of Health and Human Services was to develop an AI safety program for healthcare.

These directives recognized what researchers like Virginia Eubanks, Safiya Umoja Noble, and Ruha Benjamin have documented for years: automated systems encode and amplify existing inequalities. As Eubanks argues in Automating Inequality (reviewed in Automating Inequality), digital tools for managing poverty tend to replicate the logic of the poorhouse — sorting, surveilling, and punishing those who can least afford it. The Gender Shades study and Why Removing Bias Doesn't Work explainer examine how these dynamics manifest in AI systems specifically. The EO's equity provisions acknowledged this dynamic. Whether agencies would have meaningfully addressed it remains untested.

Workers

The EO directed the development of principles and best practices for employers deploying AI in the workplace. It called for studies on AI's impact on labor markets and existing worker protections. The Council of Economic Advisers was tasked with reporting on AI's labor market effects.

The provisions were studies and principles, not protections. No new worker rights were created. No employer obligations were established. No enforcement mechanisms existed for workplaces that deployed AI in ways that harmed workers. The EO asked for information about the problem; it did not solve it.

Innovation and Competition

The EO included provisions to fast-track visa processing for AI researchers and expand access to computational resources for researchers and small companies. It directed agencies to promote competition in the AI marketplace and prevent dominant players from leveraging their position to disadvantage competitors.

These provisions revealed an internal tension in the EO. Sections on safety and equity imposed requirements on AI companies. Sections on innovation and competition aimed to reduce barriers for them. The document tried to regulate and promote simultaneously — a balancing act that satisfied neither side fully. This tension is not unique to the EO; it runs through nearly all AI policy, as Europe's AI Paradox explores in the EU context. The challenge is real: overly restrictive regulation risks concentrating AI capability in jurisdictions with fewer rules, while permissive approaches risk concentrating it in the hands of a few well-funded firms (see AI Market Concentration).

What It Accomplished

The EO was in effect for approximately fifteen months before its rescission. In that window, several things happened.

NIST began substantive work on AI safety standards, releasing draft frameworks for AI risk management and content provenance. This work built on NIST's existing AI Risk Management Framework (released January 2023) and represented genuine institutional capacity-building.

Some federal agencies initiated equity assessments of their AI systems. The pace was uneven — agencies with existing AI governance staff moved faster; those without lagged or ignored the mandates entirely.

Federal hiring for AI positions accelerated. The EO directed the Office of Personnel Management to streamline AI hiring across government, and agencies did begin building internal expertise.

Several companies voluntarily engaged with the safety testing framework, though the extent of their compliance was not publicly documented. The voluntary nature of much of the engagement was itself a signal: companies cooperated when cooperation was costless.

What the EO did not accomplish is more telling. No binding regulations were issued. No enforcement actions were taken. No penalties were imposed on any company for any AI-related activity. The EO generated reports, frameworks, guidance documents, and institutional activity — the administrative infrastructure of governance. It did not generate governance itself. Some institutional capacity — NIST frameworks, federal AI hiring, equity assessment processes — may outlast the order. But institutional capacity without enforcement authority is preparation without follow-through.

What's Missing

The EO's most significant omission is enforcement. Every obligation runs to federal agencies, not to private companies. The DPA-based reporting requirement was the closest thing to a binding mandate on industry, and even that relied on companies self-reporting their compute usage honestly.

There is no penalty structure. A company that trains a model above the compute threshold and fails to report faces no specified consequence in the EO itself. The DPA has its own enforcement provisions, but the EO did not operationalize them for AI. The document assumed compliance would come from legitimacy and institutional pressure, not from the threat of sanction.

The EO also avoided the hardest questions about AI redistribution. It did not address the concentration of AI capability in a small number of companies. It did not propose mechanisms for sharing the economic gains from AI with affected workers or communities. It did not establish data rights for individuals whose information was used to train AI systems. It gestured toward equity without building structures to achieve it.

As Langdon Winner observed in "Do Artifacts Have Politics?," the design of technical systems embeds political choices — and those choices tend to reinforce existing power arrangements unless actively contested. The EO contested those arrangements rhetorically while leaving them structurally intact. For approaches that attempt structural intervention, see Public AI, AI Cooperatives, and UBI and AI.

The Redistribution Question

The Biden EO acknowledged what many policy documents ignore: AI creates winners and losers. Its equity provisions, worker protections, and competition mandates reflected a genuine awareness that AI's benefits and costs are unevenly distributed — that the value created by these systems flows disproportionately to those who build and deploy them, not to those affected by them.

But the EO's form undermined its content. Three structural problems made it inadequate as a redistribution instrument.

First, executive orders are politically contingent. What one president grants, another revokes. On January 20, 2025, President Trump signed Executive Order 14148, "Removing Barriers to American Leadership in Artificial Intelligence," which rescinded Biden's EO in its entirety. The replacement emphasized deregulation, framing safety requirements as obstacles to American competitiveness. The worker protections, equity mandates, and civil rights directives disappeared overnight — not because they were evaluated and found wanting, but because political power changed hands.

This is not a partisan observation. It applies regardless of which party holds the White House. Any AI governance built on executive authority alone is governance with a four-year expiration date. Kate Crawford, in Atlas of AI, argues that AI systems are expressions of political and economic power. The executive order's lifecycle proves her point: the rules governing AI reflected the political interests of whoever held power at the time, not any durable public interest.

Second, voluntary compliance is not governance. The EO relied on agency action, industry cooperation, and institutional norm-building. These mechanisms work when political alignment exists between regulators and regulated. They evaporate when it doesn't. The companies that voluntarily shared safety test results in 2024 were under no obligation to continue in 2025.

Third, the EO avoided structural intervention. It did not propose taxation of AI-derived profits to fund worker transition programs (see Stock Buybacks for where those profits go instead). It did not establish data dividends or usage rights for individuals whose information trains AI systems. It did not create Public AI infrastructure that might offer alternatives to corporate platforms. The redistribution it envisioned was procedural — agencies reviewing their own systems — not structural.

Daron Acemoglu has argued — as explored in Power and Progress — that the economic benefits of AI are not automatic; they depend on institutional choices about how the technology is deployed and who captures its returns. The Biden EO gestured toward making those choices deliberately. But an executive order is not an institution. It is a starting point that needs legislative follow-through to become durable.

What This Tells You

The lifecycle of Executive Order 14110 reveals three things about US AI governance.

It is ad hoc. The United States has no AI law. It has had executive orders, agency guidance documents, voluntary commitments, and proposed legislation that has not passed. Each of these is weaker than statute, easier to reverse, and narrower in scope.

It is politically contingent. AI policy in the US oscillates with election cycles. The Biden administration prioritized safety, equity, and worker protection. The Trump administration prioritized deregulation and competitive advantage. Neither approach is durable. Both are policy choices, and both reflect whose interests are centered.

It is structurally unable to address long-term redistribution. The questions that matter most — who captures the economic value created by AI, who bears the costs, how gains should be shared — require stable institutions, enforceable rights, and democratic deliberation. Executive orders provide none of these. They are instruments of presidential will, not democratic governance.

If you want AI policy that survives a change of administration, you need legislation. If you want legislation, you need a Congress willing to pass it. If you want a Congress willing to pass it, you need a public that demands it. The executive order was always a bridge, not a destination. Its rescission proved why bridges need to lead somewhere.

The lesson is not that the EO was worthless — its institutional investments in NIST standards, federal AI expertise, and equity frameworks created real value. The lesson is that executive action without legislative follow-through is inherently fragile. For what durable AI governance might look like, see Democratic AI Governance, AI and Democracy, and the comparative approach in Europe's AI Paradox.

Sources

  • The White House. "Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence." October 30, 2023. Federal Register
  • The White House. "Executive Order on Removing Barriers to American Leadership in Artificial Intelligence." January 20, 2025. White House
  • Eubanks, Virginia. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin's Press, 2018. Publisher
  • Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
  • Acemoglu, Daron. "The Simple Macroeconomics of AI." NBER Working Paper 32487, 2024. NBER
  • Winner, Langdon. "Do Artifacts Have Politics?" Daedalus 109, no. 1 (1980): 121–136. JSTOR
  • NIST. "Artificial Intelligence Risk Management Framework (AI RMF 1.0)." January 2023. NIST
← Back to all articles
On this page