Independent verification organizations.

The entire California Artificial Intelligence Standards and Safety Commission is eliminated. This included a 7-member body with specific appointment criteria (large business rep, small business rep, academic AI safety expert, civil society rep, worker rep, public accountant, and Attorney General), four-year terms, and a provision making it inoperative if AB 1709's e-Safety Advisory Commission was established instead.

The legislative findings and declarations section (old Section 1, subsections a through j) is entirely removed. This section stated California's global AI leadership, the need for consensus-based standards, and the intent to support safe development while preserving innovation.

The voluntary pilot program provision is removed. The old version allowed the commission to establish voluntary pilot programs to evaluate the effectiveness of developed standards.

The two-tier standards system is removed. The old version defined 'minimum compliance standard' (criteria for meeting basic safety, efficacy, reliability, and security requirements) and 'advanced safety standard' (criteria for substantially exceeding minimums and incorporating emerging best practices). These are replaced with a simpler framework of agency-developed 'requirements' and 'criteria.'

The definition of 'covered AI audit' is narrowed. The old version covered assessments under minimum compliance or advanced safety standards. The new version limits it to audits assessing whether an AI system meets minimum safety, efficacy, reliability, or security requirements necessary to comply with applicable state law.

The responsible agency is consistently changed to the Government Operations Agency throughout. The old version inconsistently referenced both the Business and Consumer Services Agency (for establishing the commission) and the Government Operations Agency (for other functions).

The IVO process changes from 'registration' to 'designation.' The new version adds more detailed application requirements including qualifications demonstrating competence, information addressing designation criteria, proposed benchmarks/technologies/metrics/methodologies, and documentation for the agency to verify accuracy.

The detailed criteria for what IVOs must be capable of assessing is significantly simplified. The old version listed five specific capabilities (assessing safety/efficacy/reliability/security/robustness, identifying foreseeable risks, verifying risk levels are appropriate, determining need for ongoing monitoring, and employing sufficient technical expertise). The new version consolidates this to three criteria: assessing risks and identifying metrics/methodologies, employing sufficient technical expertise, and managing conflicts of interest.

New provision allowing IVOs to accept payment from assessed parties at reasonable market rates, but prohibiting terms that condition payment or its amount on the results of the assessment. This clarifies the financial independence expectations for IVOs.

New requirement (8898.1(f)) that the agency regularly review and revise requirements and criteria to reflect changes in state law, technological developments, industry standards, and emerging best practices.

New requirement (8898.1(g)) to align procedures and criteria with existing professional and regulatory audit and assurance standards to promote consistency and comparability across frameworks.

New requirement (8898.1(h)) to minimize duplicative compliance obligations by allowing reports, assessments, audits, or assurance engagements prepared for substantially similar requirements to be used for purposes of this chapter.

The 'agency' responsible for establishing the commission and overseeing IVOs changed from the Business and Consumer Services Agency to the Government Operations Agency, shifting oversight to a different branch of state government.

Two new limitations were added: the chapter does not require any entity to engage an IVO or undergo a covered AI audit as a condition of developing, deploying, or operating an AI system in California, and it does not require an IVO to conduct compliance audits in order to register.

The consultation list for the commission/agency was expanded to explicitly include startups and other emerging technology companies, and private entities were split into those that develop and those that deploy or operate AI systems.

A new section (8898.1) requires the Government Operations Agency to develop requirements for IVOs by January 1, 2028, including what information an applicant IVO must submit about its benchmarks, technologies, metrics, and methodologies.

A new section (8898.3) requires registered IVOs to submit annual reports to the agency and Legislature, including summaries of standards and methodologies and descriptions of changes to governance policies or funding relevant to conflicts of interest. Includes redaction provisions for trade secrets, cybersecurity, public safety, and national security.

Working groups convened by the agency must now include engineers from competing AI companies and AI safety experts, and the agency must provide a report to the Legislature on working group findings under Section 9795.

The IVO definition was expanded to require that an IVO have no financial, operational, or management dependence on the party being assessed or its affiliates, and be free from the assessed party's control in reaching conclusions, including through contractual safeguards and conflict of interest policies.

The term 'covered audit' was renamed to 'covered AI audit' and its definition was revised to focus on assessing whether an AI system meets minimum safety, efficacy, reliability, or security requirements necessary to comply with applicable state law.

Added 'lapses in cybersecurity' as a new ground for suspending or terminating an IVO's registration, alongside existing grounds like failures to adhere to standards and conflicts of interest.

The operative condition requiring AB 1405 of the 2025-26 Regular Session to be enacted before this chapter takes effect (former Section 8898.6) appears to have been removed, making the bill no longer contingent on that companion measure.

Commission appointment authority changed: the civil society representative is now appointed solely by the Speaker of the Assembly (previously listed as 'appointed by the Governor, Speaker of the Assembly'), and the worker interests representative is now appointed solely by the Senate Rules Committee (previously listed as 'appointed by the Governor, Senate Rules Committee').

The President of the California Board of Accountancy (or designee) was removed as a standing commission member. The accounting profession seat is now filled by an individual appointed by the Governor who is authorized to practice public accountancy and has AI-related assurance experience.

The commission's duty to review and revise published standards no longer requires action 'at least annually.' The requirement is now simply to 'regularly review' standards, procedures, and criteria and revise them as appropriate.

The commission's duty to consult with stakeholders (CPAs, AI auditors, IVOs, academic institutions, private entities, consumer/labor/civil society organizations, government agencies, standards-setting bodies, and others) changed from permissive ('may consult and coordinate') to mandatory ('shall consult'), though the word 'coordinate' was dropped.

The definition of 'covered audit' was simplified by removing redundant sub-items that restated the same concept. It now simply means an assessment conducted in accordance with a minimum compliance standard or an advanced safety standard.

Section 8898.2(b) was broadened from requiring objective criteria only when developing 'advanced safety standards' to requiring them when developing 'standards' generally, meaning minimum compliance standards must also meet the same objective criteria requirements.

The bill's focus narrowed significantly. Entire sections on IVO designation process (8898.3), IVO application plan requirements (8898.4), and IVO obligations (8898.5) were removed. The commission no longer designates IVOs through an application process, maintains a registry of IVOs, submits biennial reports to the Legislature, or maintains formal liaison relationships with state agencies.

A new commission member was added: an individual authorized to practice public accountancy in California, appointed by the Governor, who represents the accounting and auditing profession and has experience providing assurance services related to AI and technology systems. Standards development now explicitly references national and international auditing and assurance standard-setting organizations and professional accountancy bodies as sources to evaluate.

The definition of 'covered audit' was expanded to include assessments conducted in accordance with a minimum compliance standard (previously only state law compliance audits and advanced safety standard assessments qualified). The 'minimum compliance standard' definition now specifically requires criteria for both operational effectiveness and design/implementation of internal controls. The 'IVO' definition was redefined from a broad category of private entities to specifically an AI auditor designated by the agency.

Commission membership was restructured from a complex system with 5 Governor-appointed categories, 3 ex officio members, detailed qualification criteria, conflict of interest rules, per diem provisions, and term limits to a simpler 7-member structure. The civil society member now requires appointment by both the Governor and Speaker of the Assembly, and the workers' interests member now requires appointment by both the Governor and Senate Rules Committee.

Two new criteria were added for determining whether an AI auditor qualifies as an IVO: (E) employing or engaging personnel with sufficient technical expertise to conduct covered audits, and (F) identifying and managing potential conflicts of interest that may undermine the integrity, quality, or independence of covered audits.

The commission's authority to establish a fee structure for charging fees to applicants and designated IVOs was removed, along with its general regulatory authority under former Section 8898.6.

The commission name was changed from 'California AI Standards and Safety Commission' to 'California Artificial Intelligence Standards and Safety Commission.' References to 'AB 1709' were changed to the formal 'Assembly Bill 1709.' The standards review requirement was changed from 'Review' to 'Regularly review.'

The entire MRO (Multistakeholder Regulatory Organization) framework is removed. The Senate version would have had the Attorney General designate private entities as MROs for a renewable 3-year period to certify AI model safety, with detailed plan requirements covering CBRN threats, whistleblower protections, and certification scope/duration. None of this survives in the Assembly version.

The bill's operation is now contingent upon enactment of AB 1405 of the 2025-26 Regular Session. Additionally, if AB 1709 is enacted and establishes an e-Safety Advisory Commission, this bill becomes inoperative and the 'commission' would instead refer to that body.

The responsible agency shifts from the Attorney General (who designated MROs in the Senate version) to the Business and Consumer Services Agency (which must establish the commission in the Assembly version). The commission's fee-setting authority also transfers from the Attorney General to the commission itself.

The Senate version's rebuttable presumption that an AI developer exercised reasonable care if their product was MRO-certified (which would have reduced litigation risk for certified developers) is removed entirely.

New procedures for suspending or terminating AI auditor registration based on: failure to adhere to standards, material misrepresentations in audit reports, conflicts of interest impairing independence, failure to maintain adequate documentation, or conduct calling into question the auditor's integrity or competence.

New definitions added: 'advanced safety standard' (criteria for AI systems that substantially exceed minimum compliance), 'AI auditor' (entity conducting a covered audit for a third party), 'covered audit' (compliance verification or advanced safety assessment), and 'minimum compliance standard' (baseline safety requirements to comply with state law).

The definition of 'independent verification organization' is modified. The Senate version defined an IVO as a private entity designated by the commission. The Assembly version adds that an IVO must be an AI auditor that has demonstrated expertise in assessing whether AI systems substantially exceed minimum compliance requirements and achieve demonstrable safety outcomes.

A two-tier standards system is introduced: the commission must develop both minimum compliance standards and advanced safety standards, publish them in publicly accessible format, review them at least annually, and include a prominent disclaimer that publication does not constitute state endorsement of any AI system or model.

A specific deadline is added: the Business and Consumer Services Agency must establish the commission on or before July 1, 2027. The Senate version had no such deadline.

Establishes the California AI Standards and Safety Commission within the Government Operations Agency, with members appointed by the Governor (representing small AI companies, frontier AI, civil society, workers, and education/research) plus ex officio members including the Attorney General and Director of Emergency Services. Members serve 4-year terms, max 2 terms, receive $100 per diem, and face conflict-of-interest restrictions including a 1-year cooling-off period before working for an IVO.

Creates a new category of Independent Verification Organizations (IVOs) designated by the commission (not the Attorney General). IVOs are private entities, nonprofits, academic consortia, or multistakeholder partnerships that verify AI safety. They operate on 3-year renewable designations with revocation criteria similar to MROs.

The commission is required to maintain formal liaison relationships with state agencies deploying or procuring AI, provide AI technical expertise and risk assessment to those agencies, support agencies in understanding IVO standards, make recommendations for state procurement requirements for AI, and comment on agency-developed AI guidelines.

Adds a new definition of 'deployer' as a person or entity that implements, integrates, or makes operational an AI model or application within the state, whether directly or as part of a product or service. This expands the bill's coverage beyond just developers to those who put AI systems into use.

Throughout the bill, obligations and protections that previously applied only to 'developers' now also apply to 'deployers.' IVO plans must address both developers and deployers in auditing, risk mitigation, whistleblower protections, and certification reporting.

The legal protection for MRO-certified AI in civil actions was changed from an 'affirmative defense to liability' to a 'rebuttable presumption that the developer exercised reasonable care.' A rebuttable presumption is generally weaker than an affirmative defense because the burden shifts to the plaintiff to overcome it, but it does not fully bar the claim.

IVO plans must now include a proposed definition of acceptable levels of risk, measurable metrics to determine whether those levels produce beneficial outcomes, target levels with data sources and measurement methods, and an evaluation and reporting protocol for ongoing compliance. This is more prescriptive than the MRO plan requirements.

IVO plans must include procedures for ongoing supervision after certification, including monitoring for risks from fine-tuning or modifying AI after certification, evaluating adherence to mitigation measures, prescribing corrective action when mitigation fails, and revoking certification if mandated corrective action is not taken in a timely manner.

The commission (in addition to the Attorney General for MROs) may establish a fee structure for charging fees to IVO applicants and designated IVOs to offset reasonable costs of carrying out its duties.

The commission must submit a report to the Legislature every two years (not annually), covering summaries of IVOs and their standards, analyses, identified gaps or emerging risks, recommendations for legislation or regulation, and an assessment of the state's AI risk landscape.

The entire amendment to Business and Professions Code Section 2570.18.5, which would have extended occupational therapist client record retention from 7 years to 10 years, was removed from the bill. This unrelated provision no longer appears in SB 813.

Section 8898.3(a) was expanded so that MROs must now certify qualified AI models or applications that meet MRO-prescribed requirements, in addition to certifying developers' and security vendors' exercise of heightened care. This gives MROs direct authority over the AI products themselves, not just the people who build them.

New Section 8898.2.5 requires the Attorney General to adopt regulations establishing minimum requirements for MRO plans and conflict of interest rules for MROs, including reporting requirements on boards of directors and donors to ensure independence from the AI industry and transparency on revenues from certification services.

New Section 8898.2.5(b) authorizes the Attorney General to establish a fee structure for charging fees to applicants and designated MROs to offset reasonable costs incurred in carrying out duties under the chapter.

Section 8898.4(a) was changed from providing an affirmative defense (which the defendant must prove) to creating a rebuttable presumption that the developer exercised reasonable care if the AI model or application was MRO-certified at the time of injury. This is a substantially stronger legal shield for developers because the burden shifts to the plaintiff to disprove reasonable care.

Section 8898.4(b) was changed from limiting the defense's inapplicability to claims of intentional misconduct, to allowing the rebuttable presumption to be overcome by any admissible evidence the court finds contrary to it. This broadens the circumstances under which a plaintiff can defeat the presumption beyond just intentional misconduct.

Added an entirely new Chapter 14 (Sections 8898-8898.4) to Division 1 of Title 2 of the Government Code creating a Multistakeholder Regulatory Organization (MRO) framework for AI safety certification, expanding the bill from a single occupational therapy records provision to a major AI governance statute.

The original occupational therapy record retention change (increasing from 7 to 10 years) is retained but renumbered as a later section following the new AI provisions, making it a secondary component of the bill rather than its sole subject.

Requires MRO applicants to submit a plan addressing auditing of AI models, mitigation of high-impact risks including cybersecurity, chemical/biological/radiological/nuclear threats, malign persuasion, and AI model autonomy and exfiltration, plus whistleblower protections, data collection for public reporting, and interfacing with federal and other state authorities.

Establishes that in civil actions for personal injury or property damage caused by an AI model or application, MRO certification at the time of injury serves as an affirmative defense to liability, except for claims of intentional misconduct by the defendant.

MRO designations expire after 3 years and may be renewed. The Attorney General may revoke a designation if the plan is misleading, the MRO systematically fails to adhere to its plan, independence is compromised, technology renders methods obsolete, or a certified AI causes significant harm.

Defines key terms including 'artificial intelligence application,' 'artificial intelligence model,' 'developer,' 'multistakeholder regulatory organization (MRO),' 'plan,' and 'security vendor.'