White House AI Accord Omits Rules for Independent Audits

Rebecca Stone

Six technology leaders signed the White House Accord on Super Intelligence on Sept. 29, accepting voluntary safety audits that Congress could later turn into law.

Google, OpenAI, Anthropic, Meta, xAI and Nvidia pledged to bring in independent outside auditors (third-party reviewers) to check whether their artificial intelligence safety controls work. The commitments remain voluntary. The accord leaves open the possibility of legally binding requirements.

What does the accord require of external auditors?

The accord does not define independence. It sets no access requirement. It names no measurement standard. Each signatory chooses its own auditor.

Those omissions permit materially different claims about assurance. One company might hire a firm with deep technical expertise and full system access. Another might hire a consultant that also sells testing services and reviews only a slide deck. Both could announce an independent audit. The labels would describe different work.

Security and compliance auditors have urged outside reviews of the most powerful artificial intelligence systems for years.

Patrick Sullivan, vice president of strategy and innovation at A-LIGN, has audited companies' security and compliance practices for more than 25 years. His firm also audits artificial intelligence controls. It would likely compete for this work. That stake strengthens his case for rules before selection begins.

Sullivan said: "A credible audit market protects the public, and it protects auditors who do the work properly." He also warns that rigorous auditors could lose work to providers offering weaker, conflicted services. His proposed rules address conflicts, competence (demonstrated expertise), access and oversight.

What must make an audit credible?

Congress created the Public Company Accounting Oversight Board after the accounting scandals of the early 2000s. The board inspects audit firms and enforces independence rules. Financial audits now earn public trust because someone audits the auditors.

That model offers a workable precedent for independence. Critics of rigid artificial intelligence audit rules raise a valid concern. The science of testing frontier AI models (the most powerful systems) remains young. Methods change with each new model. Fixed rules could freeze current techniques and slow necessary experimentation. Voluntary commitments give companies room to adapt.

Those tensions can be resolved without freezing technical practice, because independence, competence, access and oversight address the credibility of the tester rather than the specific test.

The appropriate response separates auditor eligibility from testing technique. Lawmakers need not dictate how to test an AI model. Those methods should keep evolving. Rules about who can perform the testing can endure.

Sullivan set out four conditions for credible external assurance:

• An auditor should not sell other services to the company it reviews.
• It should demonstrate proven expertise in the systems it checks.
• It should receive enough access to observe safeguards working in practice.
• Someone should check its work.

None of those requirements depends on the current testing technique. They govern independence rather than the choice of methods.

What groundwork already exists?

International standards bodies have published requirements for organizations that audit companies' artificial intelligence practices. Their guidance includes rules on auditor competence and impartiality (freedom from conflicting interests).

Connecticut passed a law this year creating a state-supervised pilot. The program will approve up to five outside organizations to verify AI systems against safety standards. Sullivan testified before the California Senate on the standards AI auditors should meet.

States are already working through these questions. Those state efforts show that governments can define auditor eligibility without freezing the technical methods needed to test AI systems and safeguards as they evolve. Washington can draw on that work.

Why must Washington set the rules first?

Sullivan warned: "Once companies start naming their auditors, the market will settle on its own definition of 'independent,' and it will probably be the most convenient one." If Congress later converts the accord into law, it will inherit the market's choice.

The accord establishes a four-layer structure: company safeguards, an internal check, an outside audit and board oversight. Finance and cybersecurity depend on clear expectations for each layer, but artificial intelligence currently lacks comparable standards that make external assurance credible and meaningful.

The third layer remains empty. Policymakers should determine who qualifies for the external-audit role before requiring companies to use it. That approach would give voluntary commitments substance and preserve room for testing methods to evolve.