"The SANDBOX Act Uses AI Policy as a Trojan Horse for Unfettered Deregulation" is a report by Nicole Alvarez and Adam Conner of the Center for American Progress (CAP), published in September 2026 on or before September 25. It analyzes the SANDBOX Act (S. 2750), the bill Sen. Ted Cruz (R-TX) introduced on September 10, 2025 as the centerpiece of his AI policy framework, and argues that its waiver program would operate as a general deregulatory mechanism rather than a regulatory sandbox in the conventional sense. It concludes that "Congress should reject the SANDBOX Act and any similar proposal that uses AI as a vehicle for broad regulatory waivers without significant safeguards or attention to the material conditions under which such waivers would operate." The acknowledgments credit Aidan Mostashari for contributions.
The report is an advocacy analysis. Its descriptions of the bill's mechanics can be checked against the bill text; its characterizations of intent and effect are the authors' positions.
Context the report sets out
The report places the bill within what it describes as the second Trump administration's agenda of "American artificial intelligence (AI) dominance." It characterizes America's AI Action Plan (July 2025) as framing rapid AI adoption as a national priority and treating regulation as the principal obstacle to it. It states that in September 2026 the United States "secured G20 support from the Chinese delegation" for the Carolina Principles for Emerging Technologies, which it describes as encouraging governments to use regulatory sandboxes and experimental exemptions to accelerate emerging technologies; its citation for that description is a September 3, 2026 article on the Winzheng site.
On federal preemption, the report states that the administration and congressional Republicans have focused on preempting state AI laws, which President Trump described as a "chaotic patchwork," arguing that Congress needs time to develop a national framework, "despite state laws currently being the main source of frontier AI oversight." It argues that Cruz's September 2025 "Legislative Framework for American Leadership in Artificial Intelligence" and the December 2025 executive order (Executive Order 14365) show that this approach "is not truly deferential to existing federal law." See AI Federalism.
The authors state that their analysis rests on the discussion draft the Senate Commerce Committee released on September 10, 2025, and that the proposal "has not been revised or advanced since." They argue that the bill should be examined even if it never reaches the Senate floor, because "discussion drafts often serve as the starting point for future legislation."
Summary of the argument
The report identifies five structural features it considers the source of the bill's main problems:
- No meaningful limit tying eligibility for a waiver to actual AI use.
- A definition of eligible "covered provisions" that extends well beyond AI-related regulation.
- Concentration of authority in the director of the White House Office of Science and Technology Policy (OSTP), who can oversee the program, override agency objections and initiate waivers.
- A mechanism by which temporary relief can become the basis for permanent amendment or repeal of federal requirements.
- Separation-of-powers concerns from letting the executive branch suspend or modify rules Congress mandated.
Departure from the conventional sandbox model
Citing a Congressional Research Service overview of Consumer Financial Protection Bureau sandboxes and the Utah legal-services sandbox, the report describes conventional regulatory sandboxes as narrowly tailored programs that let a limited set of firms test products under temporarily waived requirements, "subject to close supervisory oversight, specific conditions, and the continued preservation of baseline protections." It argues that the SANDBOX Act "inverts the premise of a traditional sandbox": it permits complete waiver or modification of federal rules to allow real-world deployment "directly to consumers, including at scale," and places "greater emphasis on scrutinizing existing federal regulations than on evaluating the technology itself." It groups the bill with two House proposals it says contemplate similar waiver authority for financial firms, the July 2025 Unleashing AI Innovation in Financial Services Act and the May 2025 draft of the Digital Asset Market Clarity Act (H.R. 3633). See AI Regulatory Sandbox.
Enforcement shield and cure periods
The report describes the bill's protections for participants: while a waiver is in effect and the participant complies with its written agreement, the participant is shielded from civil or criminal enforcement of the waived provisions, and agencies may not pursue punitive actions such as civil penalties, fines, or license suspension or revocation. If the OSTP director finds noncompliance, the director must allow at least 30 days to cure and may grant further 30-day periods; only then may the director revoke the waiver. The authors argue that conduct during a compliant waiver period "remains permanently shielded from agency enforcement, even if it would have violated the waived provisions or caused substantial harm," and that because the director rather than the affected agency controls the compliance determination and cure process, agencies cannot act until the director revokes. The report notes that the bill preserves consumers' existing private rights of action and confers no immunity for criminal offenses not expressly identified in a waiver. It states that the bill's only consumer-protection provision is a public-disclosure requirement.
Application requirements and the two pathways
The report describes two routes to a waiver: a company application and an application initiated by the OSTP director. It argues that "regardless of the relevant agency's response, both pathways can lead to the same result": an agency that grants produces a waiver, an agency that misses its deadline produces one by default, and a denial can be overridden by the director on appeal. It characterizes the private application checklist as "largely procedural," since applicants must describe benefits, risks and mitigation measures but need not demonstrate that mitigation will work or that benefits will materialize. A director-submitted application requires only the director's opinion that the waiver "will advance the development, deployment, or use of artificial intelligence in the United States."
No threshold for AI nexus
The report argues that the bill does not define how central AI must be to the activity in question, so "a company need only assert that its activity involves AI," and that agencies are not directed to assess whether a product qualifies as an "artificial intelligence system" under the bill's own definition. It concludes that this "effectively transforms a program nominally aimed at AI experimentation into a general-purpose deregulatory tool." The authors add that even a well-defined AI nexus "would not make this bill acceptable." The report cites an October 16, 2025 AFL-CIO letter opposing legislation that would exempt companies from future AI regulations.
Breadth of "covered provision"
The report traces the bill's definition of "covered provision" through the Congressional Review Act's definition of "rule" at 5 U.S.C. § 804(3), which incorporates the Administrative Procedure Act definition at 5 U.S.C. § 551 minus three exclusions (rules of particular applicability, internal management and personnel rules, and procedural rules that do not substantially affect outside parties). It identifies two expansions in the bill: inclusion of "any associated guidance, frequently asked questions publications, bulletins, or associated, derivative material," and inclusion of "any rule the adoption of which is expressly required by statute." Statutes themselves are not covered, but the authors argue that waiving implementing rules can remove a statute's practical effect, using Clean Air Act air-quality standards as the example. They also argue that the term "provision" suggests a narrow unit, whereas the definitional chain makes the waived object the rule itself, so one burdensome provision could support waiver of the whole rule. Examples of informal materials the report says would fall in scope are Office of Federal Procurement Policy Policy Letter 11-01, the Department of Education's Title IX FAQs, and Office of the Comptroller of the Currency Bulletin 2023-17 on third-party risk management. See Congressional Review Act.
Federal procurement
The report argues that the Federal Acquisition Regulation (FAR), agency supplements such as the Defense Federal Acquisition Regulation Supplement (DFARS), and related guidance fall within the covered-provision definition, though individual procurement decisions do not. It notes that procurement has become a principal federal tool for setting AI safeguards and gives the example of an AI company seeking waiver of FAR 52.204-21 and DFARS 252.204-7012, which require contractor security controls, including those in NIST Special Publication 800-171, and cyber-incident reporting within 72 hours, in order to deliver AI systems to the Department of Defense faster.
Agency review weighted toward approval
The report states that agencies must evaluate applications on public benefits (efficiency, economic growth, technological advancement) weighed against three risk categories only: health and safety risk, risk of economic damage, and risk of unfair or deceptive trade practices. It argues this excludes environmental, labor and civil-rights considerations. It describes an asymmetry in the required record: a denial must explain how the waiver could cause the risks, their likelihood, why partial approval or reformulation would not suffice, and the information relied on, and must recommend mitigation where the agency would otherwise deny; an approval requires only a limited description of mitigation and consumer protection. It notes that "risk of economic damage" is defined as tangible, physical harm to a consumer's property or assets, and that the unfair-or-deceptive-practices category is tied to Section 5 of the FTC Act and two 1980s FTC policy statements, which agencies would have to apply "on a 90-day clock."
Competitive effects and OSTP authority
The authors argue that the first waiver of a rule becomes evidence that the activity can occur "safely without" it, after which later applicants need only ask whether the agency can justify treating them differently. They argue that this pushes competitors to seek equivalent waivers "to avoid structural disadvantage in the marketplace" and that discretionary approvals would let an administration favor some firms over others. As an instance of what it describes as the administration being "accused of" using regulatory authority as leverage, the report cites the Federal Communications Commission's approval of the Paramount–Skydance merger following Paramount's $16 million settlement of a lawsuit President Trump had brought against CBS News.
The report describes OSTP as an office "with no rulemaking authority, no enforcement capacity, and no institutional tradition" in the regulatory domains the bill would reach. It lists the director's powers under the bill: submitting applications on the director's own initiative with no specific conflict-of-interest constraints; presuming no objection when an agency misses its deadline; deciding appeals of agency denials, including on the director's own applications; publishing granted director-initiated waivers in the Federal Register so that any person may apply to use them; and consulting with applicants on which provisions are likely eligible. It concludes that these powers would make the director "one of the most powerful regulatory actors in the executive branch."
Temporary relief as a path to repeal
Under Section 703, the report states, the director must send Congress an annual special message listing each covered provision waived or modified, how often, those for which applications were denied, and any the director recommends for repeal or amendment, including textual changes, based on participants being able to "operate safely without" them. The authors argue that this evidence depends on participants' own reporting of risks and incidents. If the congressional procedures track the Congressional Review Act, they state, only a simple majority is needed in each chamber; unlike the CRA, which reaches only newly finalized rules within a limited window, the SANDBOX Act mechanism applies to existing rules regardless of age.
Separation of powers
The report argues that the bill's constitutional problem lies in "empowering a lone executive branch official to suspend rules that Congress has required be issued." It discusses the nondelegation doctrine and its "intelligible principle" test, notes that the Supreme Court has not struck down a statute on nondelegation grounds in more than 90 years while some justices have suggested reviving the doctrine, and argues the bill "is aggressive enough to potentially do so." It states that the critique does not depend on a more aggressive doctrine: the bill delegates "the authority to unmake rules that Congress required to be in force," with the single standard of advancing AI supplying "no limiting principle." It draws on the English Bill of Rights of 1689, which barred the Crown from suspending laws without Parliament's consent, and on the Take Care Clause.
Corruption and abuse
The authors argue that the program lacks requirements to treat similarly situated applicants consistently, to explain differing outcomes, or to manage conflicts of interest, and that denials and delays could be used against disfavored firms. They state that the risk is greater in procurement because the federal government is the largest single purchaser of goods and services in the world. They add: "The point is not that any official will act corruptly. It is that the bill creates a structure in which corruption would be easy to carry out, difficult to detect, and largely immune to legal remedy."
Sectoral examples
The report gives three illustrations of rules it says could be targeted:
- Environment. The Clean Air Act Prevention of Significant Deterioration preconstruction permit, which requires best available control technology for a new major stationary source. A data-center operator installing on-site gas turbines could seek a waiver; the report notes that building without the permit ordinarily exposes a company to criminal penalties of up to five years' imprisonment, doubled for repeat violations. It cites the NAACP's April 2026 suit against xAI and MZX Tech over unpermitted turbines powering the Colossus 2 data center in Southaven, Mississippi, the NAACP's allegation that as of June 16, 2026 xAI operated 59 turbines without permits, and the Justice Department's motion to intervene and dismiss on national-security grounds, and argues that "a company in xAI's position would not need to wait for DOJ intervention." See NAACP v. xAI (Clean Air Act gas-turbine suit).
- Health. Medicare Advantage prior-authorization rules at 42 C.F.R. §§ 422.101, 422.137, 422.138 and 422.566, including the § 422.101(b)(6) requirements to publish internal coverage criteria and show that added utilization-management criteria are clinically justified. An insurer using AI-driven utilization review could seek waiver of those transparency requirements. See Healthcare — AI Deployment.
- Labor. The Fair Labor Standards Act administrative-employee exemption at 29 C.F.R. Part 541. The report describes a hypothetical modification under which any employee who reviews or signs off on an AI system's output is deemed to exercise "discretion and independent judgment," removing overtime eligibility; it notes Department of Labor civil penalties of up to $2,515 per violation for repeated or willful overtime violations, which would be barred during the modification period. See AI Labor Disruption.
Conclusion of the report
The authors write that "Whatever its proponents claim, the SANDBOX Act is not an AI bill. It is a deregulatory proposal that uses AI as a pretext to place virtually the entire federal regulatory system under the authority of a single White House official." They recommend that outdated or burdensome rules be addressed through ordinary legislative and regulatory processes, and argue that weakening protections "risks reinforcing the perception that AI is a technology that benefits powerful actors at the public's expense."
Comparison with the bill text
The bill text as introduced on September 10, 2025 (congress.gov lists one version, referred to the Commerce Committee that day) matches the report's account of the covered-provision definition (§ 701(6)), the three risk definitions (§ 701(8), (10), (11)), the director-submitted application standard (§ 702(c)(3)), the 90-day agency decision deadline and presumption of no objection (§ 702(d)(5)(A), (C)), the asymmetric record requirements (§ 702(d)(5)(B)), the appeal procedure (§ 702(g)), the 30-day cure periods (§ 702(j)), the enforcement shield (§ 702(k)(2)–(3)) and the Section 703 special message (Source: congress.gov). Several provisions the report does not discuss appear in the text: an agency may request one 30-day extension of the 90-day deadline; waivers run for two years, renewable for up to four further two-year periods; the program sunsets 12 years after establishment; and participants must report consumer-harm incidents to the director and relevant agencies within 72 hours and file periodic reports, alongside the public disclosures the report describes. Section 703 requires Congress to complete action on a joint resolution within 60 legislative days of the special message, with committee discharge after 10 legislative days and a non-debatable motion to proceed in the Senate "notwithstanding Rule XXII" (Source: congress.gov). See SANDBOX Act (S. 2750).
Provenance
The report was captured from the CAP site on September 25, 2026; the page carries no machine-readable publication date, and its endnotes cite material "last accessed September 2026." Inside AI Policy reported the report on September 25, 2026 as opposition to "any legislation resembling" the SANDBOX Act and noted that the position runs against Rep. Josh Gottheimer (D-N.J.), who backs financial-services AI sandbox legislation (Source: insideaipolicy.com).
Two endnotes (13 and 108) cite the enforcement shield as "§ 701(k)(2)"; in the bill text the shield is § 702(k)(2), which endnote 37 cites correctly. The report calls the analyzed text a "discussion draft"; congress.gov records it as introduced and referred on the same date. Figure 1, a flowchart of the two application pathways, is not reproduced in the captured text beyond its caption.
Relationships
- contradicts: SANDBOX Act (S. 2750) — argues against enactment of the bill
- related: Ted Cruz — bill sponsor; Office of Science and Technology Policy (OSTP) — office the bill would empower
- instance-of: AI Regulatory Sandbox — critique of a general-purpose waiver program presented as a sandbox
- depends-on: Nondelegation Doctrine, Congressional Review Act — legal frameworks the argument draws on
- related: Unleashing AI Innovation in Financial Services Act (H.R. 4801) — financial-sector sandbox bill the report groups with the SANDBOX Act
- related: Carolina Principles for Emerging Technologies, America's AI Action Plan, Executive Order 14365 — administration policy the report places the bill within
- related: NAACP v. xAI (Clean Air Act gas-turbine suit) — case used as the environmental illustration
- related: AI Federalism — preemption context
- related: Center for American Progress, Nicole Alvarez, Adam Conner — publisher and authors