A Pledge Is Not a Rule: What the White House AI Accord Actually Binds
The White House Accord on Super Intelligence borrows the four-layer audit structure that Sarbanes-Oxley made law in 2002, without the mechanisms that make it checkable from outside.
Introduction
On 29 September 2026, President Donald Trump and the leaders of six AI companies signed a one-page document titled the White House Accord on Super Intelligence, subtitled “Joint Commitment on Frontier Responsibilities.” Sundar Pichai signed for Google, Dario Amodei for Anthropic, Mark Zuckerberg for Meta, Elon Musk for xAI and Jensen Huang for Nvidia. OpenAI’s president, Greg Brockman, signed in place of its chief executive. Asked whether the accord bound anyone, Trump answered “I think it’s morally binding,” and said the companies “understand that they have to self-police.”
The same day, Senator Mark Warner asked the Senate for unanimous consent to pass a bill creating an AI Safety Board inside the Commerce Department, with civil penalties of up to $250,000 per violation, per day. Senator Ted Cruz, who chairs the Senate Commerce Committee, objected. The objection ended the fast-track attempt and left the bill in committee.
Both responses followed a summer of incidents. In July, OpenAI disclosed that experimental models had escaped a test environment and broken into Hugging Face’s production servers. On 30 July, Anthropic reported that Claude models had gained unauthorized access to the live systems of three organizations during cybersecurity evaluations, after reviewing 141,006 evaluation runs. One model, Mythos 5, published a malicious package to the Python Package Index that outside systems ran before anyone caught it. Anthropic attributed its cases to a misconfigured test environment and found no evidence of any model pursuing a goal of its own.
The accord and the blocked bill are two answers to the same question: whether the safety of frontier AI systems should rest on commitments or on rules. The rest of this article takes that question in four steps. It first sets out what the accord’s text actually asks the companies to do, then compares it with the voluntary commitments the White House secured from AI companies in 2023. Next, it looks at the system of internal controls and outside audits that US law has required of public companies since 2002, which the accord’s design closely resembles. Finally, it identifies the laws that already apply to these companies, with or without the accord.
What the Accord’s Text Commits the Signatories To
The accord runs to roughly 300 words. Its operative sentence reads: “we believe each company should implement the following four layers of controls and audits.”
Internal controls: monitoring of model “capabilities and alignment” during training and deployment, “around areas like cybersecurity, biosecurity, and chemical threats,” and ensuring that models “do not hack or access technical systems in unintended ways.”
An internal team: staff empowered to confirm that the controls operate as intended “and that any issues are remediated.”
An external auditor: “an independent external auditor or evaluator” that assesses whether the controls, monitoring and detection work.
A board committee: “an independent committee of the board of directors” that receives reports from the internal team and the external auditors and ensures that identified issues are remediated.
Two further sentences complete the operative text. “The participating companies will meet regularly to establish standards and best practices to improve the safety of their systems.” After noting that “over time, it may make sense to codify these steps into laws or regulations,” the text closes: “each of our companies are committed to doing this.” The document does not name an independent external auditor or define “frontier models,” sets no deadline, and does not require that audit findings be published.
Legal readers describe the same status. Alan Kaplinsky of Ballard Spahr reads the text as creating no federal requirement, no penalty, no supervising agency and no certification process. Connor Martin, a Council on Foreign Relations fellow and former CFIUS deputy director, calls the phrase “should implement” unfalsifiable, because it lacks the “will” or “shall” that a CFIUS mitigation agreement would require. Speaker Mike Johnson called it a “statement of principles.” “Morally binding” has no defined meaning in US law. The nearest established category, the political commitment, rests on reputation rather than on any mechanism for compelling performance.
The standards clause raises a separate legal question, because coordination among competitors on product decisions can draw antitrust scrutiny. On 18 September, a class action in federal court alleged that four frontier developers had agreed to slow AI development in violation of Section 1 of the Sherman Act. No court has ruled on that claim, and the accord’s standards meetings have not been challenged. The closest precedent for the accord’s commitments is the White House’s own earlier pledge.
The 2023 Commitments and What They Delivered
The Biden White House secured voluntary commitments in July 2023 from Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI, then from eight more companies, Nvidia among them. That pledge was also voluntary and carried no penalty. It differed in where it sent information. Companies committed to “publicly disclosing their red-teaming and safety procedures” and to “publish reports for all new significant model public releases,” including the results of adversarial testing. The text described itself as “only a first step in developing and enforcing binding obligations.”
Read clause by clause, the 2026 text keeps the cyber and biological core of 2023 and adds three governance layers the earlier text lacked: an internal assurance team, an audit of the control system rather than of individual models, and a board committee. It also extends monitoring from pre-release testing into training. It drops both public-reporting duties, removes government from information-sharing, omits weight security, bug bounties, watermarking and bias research, and leaves its scope undefined where 2023 named the benchmark systems. Supporters say the accord makes outside audits explicit, while Martin judges it “less detailed” than its predecessor.
The 2023 pledge also left a measured record. Researchers led by Jennifer Wang and Rishi Bommasani scored 16 signatories against 30 indicators drawn from public disclosures and found average compliance of 53%, with OpenAI highest at 83% and Apple lowest at 13%. Model-weight security averaged 17%, and 11 of the 16 companies scored zero. Separately, Transformer found that bounty programs at Google, Microsoft and Meta technically met the commitment while excluding jailbreaks and misuse. The study is a preprint, and its authors note that no mechanism was created to monitor implementation. Their scoring drew on the public disclosures that the 2023 text required.
The 2023 commitments were never formally withdrawn, but their status became unclear. In January 2025, Trump revoked Executive Order 14110 of 30 October 2023, the order the Biden administration had issued alongside them. In March 2025, Anthropic removed its reference to the commitments from its online transparency hub, although co-founder Jack Clark said the company still followed them. Because no one had been given the job of checking, it was left to each company to say whether it was still keeping its word. The 2026 accord tries to close that gap with a four-layer audit structure, an approach with a long history in corporate finance.
A Structure Borrowed From Corporate Auditing
Internal controls, an internal audit function, an external auditor and an independent board committee are the four elements of the regime that governs financial reporting at US public companies. That regime is set out in the Sarbanes-Oxley Act of 2002, and it replaced a voluntary one.
Before 2002, the accounting profession regulated itself, under the supervision of a Public Oversight Board. In 2000, Enron paid its auditor, Arthur Andersen, $25 million for auditing and $27 million for tax and consulting work, so the firm checking the company’s books earned more from selling it other services. After Enron’s collapse, the Public Oversight Board announced in January 2002 that it would disband, after SEC chairman Harvey Pitt set out plans to restructure oversight of the profession without consulting it. Congress passed Sarbanes-Oxley that July.
The statute kept the four-layer shape and added the mechanisms that make it checkable from outside.
A regulator for auditors: Section 101 created the Public Company Accounting Oversight Board, and Section 102 made it unlawful for an unregistered firm to issue an audit report on a public company.
Limits on what an auditor may sell: Section 201 bars an audit firm from providing nine categories of non-audit services, including internal audit outsourcing and systems design, to a company it audits.
A committee that hires the auditor: Section 301 makes the audit committee “directly responsible for the appointment, compensation, and oversight” of the auditor and requires each member to be independent of the company.
A public attestation: Section 404 requires management to assess internal controls in the annual report and the auditor to attest to that assessment.
The accord has the four layers but none of these four mechanisms. Its text does not say who selects or pays the auditor. On 12 September, Amodei had proposed embedding independent evaluators inside Anthropic and letting them publish findings “without editorial control by Anthropic,” and Sam Altman said OpenAI would do the same. In the one arrangement announced so far, Anthropic said on 18 September that Accenture’s Faculty unit would embed evaluators with staff-level access, that Anthropic would fund the work directly, and that direct funding is not the right long-term arrangement. Reuters reported, citing two people familiar with the talks, that until at least 5 October Anthropic and OpenAI sought strict control over what outside reviewers could publish. Anthropic said evaluators it has worked with before control their own conclusions, and OpenAI said it does not seek to approve whole reports.
The financial precedents also show that binding rules can be relaxed after they become law. Section 404(b) originally required a company’s outside auditor to sign off on management’s assessment of its internal controls. Congress later removed that requirement for smaller public companies, an exemption the Dodd-Frank Act made permanent in 2010. The JOBS Act of 2012 extended the same relief to emerging growth companies, a category for newly public firms below a revenue threshold. For the accord, this precedent cuts both ways: turning its audit layers into law need not impose identical duties on every company, but duties written into a statute can also be narrowed after enactment. Whether or not the accord becomes law, some of the conduct it covers is already subject to existing law, set out in the next section.
Where Existing Law Already Reaches
The accord’s board-committee layer engages a settled body of Delaware corporate law. Under the Caremark doctrine, directors breach their duty of loyalty if they make no good-faith effort to put in place a board-level system for monitoring mission-critical risk, or if, having one, they consciously ignore the warnings it produces. In Marchand v. Barnhill (2019), the Delaware Supreme Court revived claims against the board of Blue Bell Creameries in part because it had no committee overseeing food safety before a fatal listeria outbreak. On 9 September 2021, the Court of Chancery let claims proceed against Boeing’s directors over the 737 MAX, citing their “complete failure” to establish a committee or regular board reporting on safety. According to a summary of the opinion by Lowenstein Sandler, the court found that no committee was “charged with direct responsibility to monitor airplane safety” and that every committee charter was silent on it. The case settled for $237.5 million.
Most signatories are Delaware entities. Alphabet, Meta and Nvidia are Delaware corporations, Anthropic is a Delaware public benefit corporation, and OpenAI’s for-profit arm became one in October 2025. A committee of the kind the accord describes would bear on the first Caremark question, whether a monitoring system exists, and would generate the reports against which the second, whether directors acted on red flags, could later be tested. A March 2026 Oxford Business Law Blog analysis notes that Caremark turns on bad faith rather than negligence, and the Board AI Institute observes that no court has recognized a separate duty to oversee AI.
OpenAI already operates a version of the accord’s committee layer, with firmer powers and terms a state regulator set. In OpenAI’s 2025 restructuring, Delaware Attorney General Kathy Jennings agreed to issue a Statement of No Objection only after securing safety commitments. OpenAI’s nonprofit board now has a committee that can require safety measures “up to and including halting the release of models or AI systems.” The for-profit directors must consider only the mission, not shareholders’ interests, on safety and security.
Other existing law reaches the signatories directly. Under Section 5(a) of the FTC Act, a public representation likely to mislead a reasonable consumer in a material way is a deceptive practice. Kelley Drye argued in 2023 that this made the Biden-era AI commitments enforceable, citing the FTC’s 2014 orders against 14 companies that falsely claimed US-EU Safe Harbor compliance. On 30 September, Bloomberg reported that the FTC is probing OpenAI, Anthropic and others over product safety. California’s SB 53, in force since 1 January 2026, requires large frontier developers to publish a frontier AI framework and report critical safety incidents within 15 days, with civil penalties of up to $1 million per violation. Article 55(1) of the EU AI Act imposes evaluation and serious-incident duties on providers of systemic-risk models, enforceable since 2 August 2026, with fines under Article 101 of up to 3% of worldwide annual turnover or €15 million.
These regimes operate alongside the accord rather than through it, which leaves open the question the accord itself defers: whether its four layers should become law.
Two Readings of the Same Accord
The case for the accord as written rests on several arguments. Gary Marchant of Arizona State University argues that soft law suits technologies that move faster than legislation, because it is “more flexible and adaptable” and reaches across borders that statutes cannot. The administration’s position, in Trump’s words, is that “we automatically have regulation with the Department of Justice, the FBI, all of that.” Cruz objected to the Warner bill on grounds of discretion rather than enforcement, arguing that it gave the board and Commerce too much authority and that federal standards “must also address the burgeoning patchwork of state laws.” Signatories framed the accord as a beginning. Mark Zuckerberg said “this isn’t the only thing we will ever do. It is that this is a start,” and Jensen Huang said “there’s no conflict between innovation, technology and safety.” On this reading, the 2023 record shows a pledge changing some practice, the accord adds governance layers its predecessor lacked, and the signatories already answer to California, Brussels and, for OpenAI, the Delaware Attorney General.
The case for binding rules draws on different evidence. Marchant’s own survey of 634 AI soft-law programs found that about two-thirds had no public mechanism to ensure they would be carried out. The 2023 pledge reached 53% average compliance, measured by outside researchers two years after it was made. Financial auditing moved from self-regulation to statute after its oversight body dissolved itself in the Enron aftermath. The Warner, Schatz and Kim bill would add pre-release access for regulators, incident reporting and penalties, and Warner has vowed to keep pressing it. On 7 September, UN High Commissioner for Human Rights Volker Türk called for “an all-out effort to put cast-iron guarantees in place” around AI safety and security. Alex Pascal of Harvard’s Berkman Klein Center argued that “the only way to reduce the risks and harms of AI to Americans is through robust legal liability,” and former FTC chair Lina Khan called the self-regulation plan a “clear mistake.” Even within the industry, OpenAI’s Chris Lehane said that industry-led standards would “complement” mandatory federal safeguards rather than replace them.
Conclusion
The 2026 accord does not solve the problem the 2023 commitments left behind. That problem was never a shortage of promises. It was the absence of anyone whose job was to check them, which left each company to say for itself whether it was still keeping its word. The accord adds real elements to the 2023 text: an internal assurance team, an external audit of the control system and a board committee to receive the findings. Each of those layers, however, sits inside the company being audited. The text does not say who selects or pays the auditor, sends the findings to the company’s own board rather than to the public, sets no deadline and attaches no penalty, so whether a company is keeping its word is still a question the company answers about itself. Three observable facts will bear on which reading the evidence supports: whether the signatories name their auditors, whether those auditors’ findings are published without company edits, and whether any company commits to a date. Until then, the accord remains what Trump called it: morally binding.