The mirage of AI deregulation
One of the most interventionist approaches to technology governance in the United States in a generation has cloaked itself in the language of deregulation. In early December 2025, President Donald Trump took to Truth Social to announce a forthcoming “One Rule” executive order on artificial intelligence (AI). Trump warned that US leadership in AI would be “DESTROYED IN ITS INFANCY” by the meddling of “50 States, many of them bad actors, involved in RULES and the APPROVAL PROCESS.” But beneath the bluster was a consequential policy move: the federal preemption of state authority to govern AI, framed as the removal of bureaucratic obstacles from the path for American technological dominance.
This social media post and the executive order that soon followed crystallized the prevailing narrative about the administration’s approach to AI governance: that it is fundamentally deregulatory. This narrative is misleading. Not in its particulars—the Trump administration has indeed rescinded Biden-era AI safety and innovation policies and pressured allies to reduce what it calls “innovation-killing regulations”—but in its core framing. To describe this AI policy as “deregulatory” is to confuse the removal of certain kinds of rules with the absence of governance itself.
To speak of regulation is to conjure a particular image: the federal agency, the notice-and-comment process, the rule published in the Federal Register. This is regulation as visible, formal, legible—and therefore open to dispute and debate. But governance does not vanish when it stops looking like this. Decades of social science research make clear that the steering of policy occurs through mechanisms far more varied than formal rulemaking. Much of the classic work on regulation focused on how interests contest rules once they are on the books. Economist and Nobel laureate George Stigler showed that regulation serves concentrated interests, but his analysis still presumed the formal rule as the site of struggle. Other strands of scholarship have shown that regulatory power often operates earlier and through the shaping—and breaking—of shared expectations about what kinds of governance are legitimate or even imaginable. Political scientists Martha Finnemore and Kathryn Sikkink identified one such mechanism: norm entrepreneurship, the work of establishing collective understandings about appropriate behavior that shape what is possible before any rule is written.
All administrations engage in norm-setting; this is an ordinary feature of governance. What distinguishes the current approach is not norm entrepreneurship itself but its character: a greater reliance on executive discretion than on deliberative, public processes.
The Trump administration is engaged in norm destruction—breaking expectations about transparent governance and public oversight while installing new assumptions about how technological development should be directed. What it has advanced is not the absence of AI regulation but its rearrangement, often by caprice: intensive state intervention operating through industrial policy, trade restrictions, immigration controls, equity stakes in private firms (selected by the state), the redirection of research funding, and the strategic preemption of state authority. Many of these actions face legal challenge, and some may not survive judicial review. But the pattern itself—the systematic preference for executive discretion over deliberative process—reveals an approach to governance that will shape AI policy regardless of how individual cases are decided. This is not deregulation. Not in the least. It is hyper-regulation by other means.
When Michael Kratsios, the White House science and technology advisor, traveled to Montreal in December 2025 to address G7 technology ministers, he was not merely lobbying against specific rules. He was working to establish as common sense the proposition that government oversight represents an obstacle to innovation rather than a precondition for it. The “trusted AI ecosystem” he advocated—built on “smart, sector-specific regulations tailored to each nation’s priorities”—is itself a normative framework, one that privileges certain values (a narrow idea of innovation, national competitiveness) while marginalizing others (innovation broadly understood to include social and economic benefits, precaution, public accountability, the rights of those affected by AI systems). The Trump administration is not withdrawing from AI regulation; it is attempting to set the terms on which AI governance will be understood globally.
The material dimensions of this governance project are no less significant. In August 2025, the US Department of Commerce became the largest shareholder of Intel, acquiring a 9.9 percent stake in exchange for $8.9 billion in funding. The administration has since moved to take equity positions across semiconductors, critical minerals, and nuclear energy—committing more than $10 billion to ownership stakes in at least nine firms in six months.
Investment in semiconductors and critical minerals was a priority of the Biden administration as well. What distinguishes the current approach is not industrial policy itself but how it operates—circumventing ordinary accountability mechanisms in favor of discretionary power.
This is industrial policy of a scope and ambition that would have seemed inconceivable even a decade ago–—particularly from a party historically committed to free enterprise and opposed to government involvement in the private sector. It represents not the absence of state intervention in markets but its intensification, conducted through regulation by ownership. When the government acquires equity in private firms, it gains leverage: influence over where factories are built, which technologies are prioritized, and which partnerships are deemed acceptable.
The influence these ownership stakes confer exceeds what formal regulation could achieve. As a condition of approving U.S. Steel’s acquisition by Japan’s Nippon Steel, the administration secured a “golden share,” granting it veto power over major corporate decisions. When the company later planned to close an Illinois facility, the government intervened. This is regulation exercised through discretion rather than administrative procedures that allow for public notice, comment, and judicial review.
Immigration policy provides another, less visible lever of AI governance. Restrictions on high-skilled immigration—implemented through visa screening, fees, and administrative guidance—shape who participates in US AI research and development. Concerns about the potential displacement of US workers are not unfounded, but the governance implications extend beyond labor markets. Restricting this pipeline does not just limit who works in laboratories, firms, and universities; it determines which perspectives and expertise circulate within them, shaping both what gets built and what questions can be asked about its consequences.
At the same time, research funding has been redirected through executive action toward narrowly defined AI development priorities, while work examining the social, labor, and health implications of the widespread deployment of AI models, tools, and systems has been defunded or deprioritized. The administration has framed research on algorithmic discrimination, disparate impact, and the environmental costs of AI systems as political bias rather than empirical inquiry. This rhetorical move attempts to sideline the substantive premise that AI governance should attend to collective goods. Democratic accountability concerns not just how decisions are made but whose interests they protect and whose harms they make visible.
These decisions restructure the intellectual resources of the scientific enterprise itself, influencing not only what technologies are built but which questions can be asked about their consequences. This is AI governance operating upstream of regulation, through control over the people and knowledge that constitute the field.
Perhaps the clearest illustration of this paradox—and of how far this approach departs from the states’ rights principles historically central to conservative governance—is the preemption of state AI laws. Trump’s “One Rule” rhetoric frames the problem as regulatory burden: companies cannot navigate fifty different approval processes. This is not an invented problem. Regulatory fragmentation can impose real coordination costs, and poorly designed state laws might create obstacles without corresponding benefits.
But the solution matters as much as the problem it addresses. Federal preemption by executive action, rather than through congressional action, replaces diffuse but locally accountable policy with concentrated but unaccountable governance. Framed as relief from regulatory burden, preemption represents an aggressive assertion of federal authority that forecloses democratic experimentation at the state level. Whatever one’s view of state-level AI policy, federal preemption is itself a form of regulation—one that concentrates power while insulating it from local accountability.
Here the regulation-deregulation binary collapses entirely. The Trump administration is not removing government from AI regulation; it is concentrating governmental power at the federal level while deploying it through mechanisms—investment, ownership, research funding, immigration controls, and preemption—not typically classified as “regulation.” The effect is a regulatory regime simultaneously more intensive and less transparent, less democratically accountable, than the rulemaking it displaces. Political economist Dani Rodrik has observed that governments always engage in industrial policy; the question is whether they do so deliberately and transparently or “surreptitiously and without an overall strategic frame.”
The current administration’s AI policy may be a case study in the latter. The decisions being made now may determine US technology research, development, and deployment infrastructure for decades; early choices at critical junctures lock in trajectories that become increasingly difficult to reverse. An administration proclaiming deregulation is making precisely such choices through mechanisms that escape the ordinary character of regulatory debate. These choices are hyper-regulatory, interventionist, yet outside the traditional vocabulary of regulation while exercising power every bit as consequential.
Procedures such as public notice, comment periods, reasoned decision-making, judicial review, and congressional approval are hard-won mechanisms of public oversight. These principles must be extended to the new modalities through which governance now operates: the equity stake chosen in private, the research priority set by executive action, the visa rule implemented through guidance.
Regulation is not just a matter of abstract procedures and frameworks. Governance is always governance of someone. When a graduate student loses funding for scholarship on algorithmic bias, that is not an abstraction—it is AI policy in practice. When a machine learning researcher cannot obtain a visa to return to her lab, that is AI regulation at work. And when a rare earth miner’s labor unfolds within constraints set by trade policy, ownership arrangements, and government-defined priorities, governance is operating through material conditions rather than formal rules.
When decisions about the technologies that will shape our work, our health, our capacity to know and to communicate are made without public justification, outside ordinary channels of accountability, we are not governing ourselves. We too are being governed. The mirage of AI deregulation obscures this truth. Power has not retreated. It has moved—and democratic accountability requires following it there. Seeing clearly is where that work begins.
(1)eLetters
eLetters is a forum for ongoing peer review. eLetters are not edited, proofread, or indexed, but they are screened. eLetters should provide substantive and scholarly commentary on the article. Neither embedded figures nor equations with special characters can be submitted, and we discourage the use of figures and equations within eLetters in general. If a figure or equation is essential, please include within the text of the eLetter a link to the figure, equation, or full text with special characters at a public repository with versioning, such as Zenodo. Please read our Terms of Service before submitting an eLetter.
Log In to Submit a ResponseNo eLetters have been published for this article yet.
Recommended articles from TrendMD
- Museum With a HeartScience, 1999
- Refracting ElectronsScience, 2004
- A Colorectal CatalogScience, 2006
- Biofuel BugsScience, 2011
- Male lyrebirds mimic mobbing flocks when having sexTess Joosse, Science Now, 2021
- LISTEN: nerandomilast Phase III trial results in ILD discussed by a doctor and patient
Valenzuela & Kervitsky, ERJ
- LISTEN: 10-minute podcast on nerandomilast Phase III trial results
Valenzuela & Kervitsky, ERJ
- Effect of nerandomilast on outcomes in IPF
Oldham, AJRCCM
- Nerandomilast in idiopathic pulmonary fibrosis: data from the whole follow-up period of the FIBRONEER-IPF trial
Oldham, AJRCCM
- Effect of nerandomilast on outcomes in pulmonary fibrosis
Wisenbeek, ERJ
Information & Authors
Information
Published In
Science
Volume 391 | Issue 6782
15 January 2026
15 January 2026
Copyright
Copyright © 2026 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works.
Article versions
Submission history
Published in print: 15 January 2026
Authors
Metrics & Citations
Metrics
Article Usage
- 2 citation in Crossref
- 0 citation in Web of Science
Altmetrics
Citations
Cite as
Export citation
Select the format you want to export the citation of this publication.
Cited by
- Structured Absence, Social Media + Society, 12, 2, (2026).https://doi.org/10.1177/20563051261452539
- Is the English Hegemony in Health Publishing Coming to an End?, Depiction of Health, 17, 1, (1-4), (2026).https://doi.org/10.34172/doh.2026.01
Loading...
Embodied AI Calls for Embedded Ethics
The mirage of AI deregulation (15 Jan 2026, Vol. 391, Issue 6782) has highlighted hyper-regulation by other means (1). Forms of governance operating beyond formal regulatory regimes have attracted growing attention and debate (2). In a broader context, this line of reasoning is especially salient for embodied artificial intelligence, where early technical choices strongly condition how systems later operate in the real world.
Embodied AI differs in important respects from AI systems operating primarily in digital environments (3). Its decisions are enacted through physical motion and force, unfold in real time, and may directly affect human safety, dignity, and fundamental rights. Ethical considerations introduced only after deployment often prove insufficient. Embodied AI therefore calls for embedded ethics at the design stage (4).
Recent commentary has cautioned that technologies, and technologists, cannot be left to self-regulate (5). However, for embodied AI, the absence of ethical considerations during the development phase can render later governance ineffective or prohibitively costly. Early choices concerning data, sensing, and learning quickly stabilize into physical behaviors that are difficult to reverse once systems are deployed at scale; for example, if training datasets for service robots fail to account for the needs of visually or hearing-impaired individuals, biases may be amplified through physical action in real-world environments (6).
For embodied AI, embedded ethics can provide a first line of defense for responsible deployment in human society, Which can be promoted through governance mechanisms beyond formal regulation. Just as the flyball governor may not have made the steam engine absolutely safe, its adoption nonetheless marked a crucial milestone in enabling the Watt steam engine to be integrated into social life.
References and Notes
1. A. Nelson, The mirage of AI deregulation. Science 391, eaee4900 (2026). doi:10.1126/science.aee4900
2. W. Liang, K. Chen, S. Wang, Governance over ethics in science and technology: An important link in the implementation of the Healthy China initiative in the era of artificial intelligence. [in Chinese] (CASS, 2025); https://www.cssn.cn/skgz/bwyc/202512/t20251231_5964756.shtml.
3. N. Obayashi, T. Watanabe, K. Suzuki, Embodied intelligence paradigm for human-robot communication. Sci. Robot. 10, eads8528 (2025).
4. Y. K. Dwivedi, A. Schlosser, in The Cambridge Handbook of the Law, Policy and Regulation for Human-Robot Interaction, J. M. Bauer, M. Latzer, Eds. (Cambridge Univ. Press, 2024), pp. 245–263.
5. J. Stilgoe, Governance can’t be automated. Science 388, eadx3843 (2025).
6. H. Holden Thorp, Resisting AI slop. Science 391, 5-5 (2026).
*Corresponding author. Email: r.chai@bit.edu.cn (R.C.); liangwn@tsinghua.edu.cn(W.L.); shuo.wang@ia.ac.cn (S.W.).
Acknowledgments:
Funding:
Tsinghua Strategy for Heightening Arts, Humanities and Social Sciences “plateaus & Peaks” grant 2024TSG06402 (WL)