Deployer Assurance Checklist (CA IVO Voluntary)

Use the LegInfo chaptered bill text (SB 813, Chapter 179, Statutes of 2026; approved/filed 2026-09-09). Definitions that matter: IVO means an AI auditor designated by GovOps as having demonstrated expertise assessing risks of an AI system or model and identifying the metrics and methodologies behind that assessment. Ignore unsigned or un-chaptered bills (including AB 1883 and similar) as if they were current California law.

§8898.4(a) is explicit: the chapter does not establish liability solely for failure to comply with a standard under the chapter; does not constitute state endorsement of any AI system; does not require IVO engagement or a covered AI audit to develop, deploy, or operate in this state; and does not require an IVO to conduct legal-compliance audits merely to register. Operate first; IVO is optional assurance, not a permit.

On or before 1 January 2028, GovOps must develop IVO application requirements (qualifications, designation criteria, proposed benchmarks/metrics/methodologies), procedures to suspend or terminate designation (including independence, misrepresentation, cybersecurity lapses), and published criteria for who qualifies. The agency must convene working groups (including engineers from competing AI companies and AI safety experts) and report findings to the Legislature. Until those criteria exist, there is no designated IVO list to “check the box” against.

If you want to be ready for a future voluntary IVO, collect what the statute says IVOs will be judged on: risk assessment of the system or model in operation, metrics and methodologies, technical expertise, conflict-of-interest and independence controls, and documentation. Keep that pack as operator evidence. Do not claim a legal safe harbor — §8898.4(b) says a standards-aligned audit is relevant to, but not conclusive of, a harm action.

The Governor’s 2026-09-09 signing post also covers AB 1405 (Bauer-Kahan): a state registry for AI auditors with independence, transparency, and integrity standards. Use it as companion context for who may later appear on a registry. It is not a substitute for SB 813’s voluntary IVO path and it is not a requirement to engage an IVO before operating.

Once GovOps designates IVOs, a deployer may voluntarily engage one. A designated IVO must report annually to the agency and Legislature no sooner than 12 months after initial designation (standards/methodologies, governance/funding relevant to independence, application-info changes), with allowed redactions for trade secrets, cybersecurity, public safety, or national security. Until designation exists, skip this step.

If you serve EU users, check two AI Act Article 50 duties separately. As a provider: if you offer a generative AI system under your own name, its outputs must be marked in a machine-readable format and detectable as AI-generated (Article 50(2)). This has applied since 2 August 2026; systems already on the market before then have until 2 December 2026. Do not assume your model vendor does it for you: OpenAI’s API text watermarking (5 Oct 2026) is opt-in for select models and off by default. As a deployer: text published to inform the public on matters of public interest must be disclosed as AI-generated, unless it has gone through human review or editorial control and a person or company holds editorial responsibility for it (Article 50(4)). Treat watermark checks as weak evidence either way. In OpenAI’s own tests at a 1% target false-positive rate, its detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages on psychology-type content, and far fewer on maths. Replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. OpenAI says a watermark does not identify the user, establish ownership or responsibility, or measure human contribution, and that finding no watermark does not prove a human wrote the text. Its detector is open only to approved researchers and expert organisations. The Commission’s Code of Practice on Transparency of AI-generated Content is a voluntary way to show compliance with these duties.