The Auditors: California’s SB 813 and AB 1405 Create the First External Verification Architecture for AI Systems in the United States

1. What Newsom Signed

On 9 September 2026, Governor Gavin Newsom signed Senate Bill 813 and Assembly Bill 1405 into law. [Established — Governor of California official press release, “Governor Newsom signs first-in-the-nation AI safeguards to protect Californians, calls on the federal government to do its part,” 9 September 2026.]

Senate Bill 813 establishes a legal framework for Independent Verification Organisations — accredited third-party entities with authority to assess AI systems and models for compliance with California state law. Assembly Bill 1405 creates complementary standards for what those assessments must cover, including the documentation AI developers must produce for IVO review. [Established — Transparency Coalition, “California Gov. Newsom signs two bills to create nation’s first AI auditing framework,” 9 September 2026; ABC7, “California Gov. Gavin Newsom signs landmark bill creating AI safety measures.”]

The IVO framework applies to AI systems meeting defined thresholds of computational scale and deployment reach — the same class of systems addressed by the 2025 SB 53 transparency law. Developers whose systems meet those thresholds must make their systems available for IVO assessment on a defined schedule; IVOs publish their findings; non-compliance is reportable to the California Attorney General. [Established — Governor press release; Transparency Coalition analysis, September 2026. Note: the full text of the enacted bill has not been fully reviewed at time of publication; these characterisations reflect the signing press release and contemporaneous reporting from multiple outlets.]

Newsom described the signing as establishing “first-in-the-nation standards for independent assessments of AI systems” and called on the federal government to “do its part.” [Established — Governor of California press release, 9 September 2026.]

2. What Independent Verification Organisations Are

The IVO concept is adapted from regulatory frameworks in finance, aviation, and nuclear power, where independent external assessors evaluate compliance with technical and legal standards. The structural feature that distinguishes external auditing from self-certification and disclosure is authority: an IVO does not merely receive reports from AI developers; it conducts its own assessment of the AI system and reports its findings independently of the developer. [Established — Transparency Coalition analysis; standard description of third-party auditing frameworks as applied to regulated sectors.]

In financial regulation, the closest analogue is the Public Company Accounting Oversight Board, created by the Sarbanes-Oxley Act of 2002 following the Enron and WorldCom accounting frauds. Before Sarbanes-Oxley, public companies’ financial statements were audited by firms engaged and paid by the companies themselves — a structural conflict of interest. SOX created the PCAOB to oversee and standardise those audits, and gave the SEC authority to enforce the framework. [Established — SEC and PCAOB historical documentation; general legal knowledge.]

SB 813 imports this logic into AI: the developer builds the system; an accredited independent organisation assesses it; the assessment is published. The practical question the analogy raises is whether AI systems can be assessed with the standardisation that financial statements can. Accounting standards (GAAP, IFRS) provide a common language for auditors. AI systems operate across vastly different architectures, training data, and deployment contexts; there is no equivalent of GAAP for large language models or multimodal systems. [Assessed — this is a genuine open question in AI governance, consistent with the published commentary of AI safety researchers and the Wilson Sonsini analysis of SB 53.]

SB 813 sidesteps this by establishing IVO accreditation criteria and requiring California to define assessment standards by a specified date — creating the auditing authority first and deferring technical standards-setting to a subsequent regulatory process. That sequencing is the right order of operations. The alternative — waiting until assessment standards are fully agreed before creating the authority to assess — produces indefinite delay. [Assessed — analytical characterisation; consistent with historical regulatory sequencing in comparable sectors, including financial auditing and aviation safety.]

3. Why External Auditing Changes the Compliance Architecture

Prior AI regulation in California operated through two mechanisms: self-certification (developers describe their own risk frameworks and safety protocols) and disclosure (developers report to state agencies on critical safety incidents). [Established — SB 53, signed 29 September 2025, established these requirements; described by the Transparency Coalition and Future of Privacy Forum in their analysis of California’s AI framework.]

The limitation of self-certification and disclosure is information asymmetry: the developer knows more about their system than any regulator or external party, and the regulator can only assess what the developer chooses to report. This is precisely the structure that produced the accounting frauds SOX was designed to prevent. In AI, the information asymmetry is even more pronounced: AI systems are substantially more opaque than financial statements, and the harms they can produce are harder to detect from outside the system than accounting irregularities. [Assessed — structural characterisation consistent with the analysis published by the Future of Privacy Forum and Wilson Sonsini in their SB 53 analysis; the opacity of large AI systems relative to financial statements is well-established in the technical literature.]

External verification by an accredited IVO changes the power relationship between regulator and developer. The developer no longer controls the information flow; the IVO conducts independent testing and produces independent findings. It is harder to misrepresent to an external assessor who has direct access to the system and the legal authority to publish than to a disclosure regime that relies on self-report. [Assessed with high confidence — structural inference from the known limitations of self-certification versus external audit in analogous regulatory contexts.]

The caveat is enforcement. SOX’s power derived from the SEC’s authority to bring enforcement actions and from the personal criminal liability of executives who certified false financial statements. SB 813’s enforcement architecture — referral to the California Attorney General — is structurally less sharp. The first IVO assessments will test whether the disclosure-and-referral model is sufficient deterrent or whether stronger enforcement is needed. [Assessed — inference from the press release characterisation of enforcement; this assessment should be revisited against the full enacted text.]

4. The Federal Vacuum and September 30

Newsom’s call on the federal government to “do its part” describes a regulatory gap that is now structural. The Trump administration withdrew the executive order on AI safety issued by President Biden in October 2023. No federal AI safety legislation has passed Congress. The EU AI Act is in force. Russia’s AI law entered force on 1 September (Sounding No. 29). California has now enacted an IVO framework. The United States federal government has no equivalent. [Established — Governor press release; EU AI Act in force per Sounding No. 36 analysis; Russia AI law per Sounding No. 29; Trump administration executive order withdrawal per prior reporting.]

The practical consequence is that California’s framework becomes the operating compliance baseline for AI companies with California operations — which means essentially every major US AI developer. Google, Meta, Apple, OpenAI, Anthropic, and Microsoft all operate substantially in California. When California mandates IVO assessment, those companies face a de facto national audit requirement regardless of what Washington does. [Assessed — this is the standard mechanism by which California regulations achieve national effect; the historical precedent is California’s vehicle emissions standards, adopted by fifteen other states and effectively setting the US national standard for decades. LW.com, “California Assumes Role as Lead US Regulator of AI,” makes this same structural argument.]

Approximately 28 AI-related bills remain pending Newsom’s signature before 30 September. The Governor’s pattern since 2024 has been to sign narrowly scoped regulation and veto broader liability-imposing measures — consistent with the 2024 veto of SB 1047 and the 2025 signing of SB 53. [Assessed — pattern inference from Gizmodo, “Newsom Signs AI Industry-Approved AI Regulation Bills Into Law in California”; LW.com California AI regulatory analysis.] SB 813 and AB 1405 fit the pattern: they create institutional architecture without imposing direct liability on developers for what their systems produce. The pending bills likely include measures from both categories. Newsom’s September 30 choices will determine whether the IVO framework operates as a standalone auditing baseline or as part of a broader enforcement and liability regime.

Bottom line: SB 813 and AB 1405 do something no prior AI regulation in the United States has done: they create external auditing authority rather than a self-report regime. The analogy is Sarbanes-Oxley — the moment a sector goes from self-certification to external check. That moment is structurally significant because external auditing changes the power relationship between regulator and developer. California is doing what the federal government has declined to do. It will do it imperfectly; the technical standards are not yet written; the enforcement mechanism is weaker than the financial analogy. But the architecture exists. That is the change.