California AI Governance Guide for Local Governments

The California Public Records Act defines a public record as a writing containing information relating to the conduct of the public’s business that is prepared, owned, used, or retained by a state or local agency, regardless of physical form.^[1]

The definition of “writing” includes words, pictures, sounds, symbols, email transmissions, and every other means of recording a communication or representation, regardless of storage method.^[2]

When an identifiable, nonexempt public record is held electronically, an agency generally must make it available electronically upon request and, when applicable, in an electronic format in which it holds the information.^[3]

What this may mean for AI

An AI prompt, output, transcript, recommendation, score, log, review note, or system record may qualify as a public record when it satisfies the statutory public-business and agency-use tests. That conclusion depends on the facts and is an application of general public-records law, not an AI-specific rule. Do not state that every prompt, output, or vendor log is automatically a public record.

Questions local officials should ask

• Does the material relate to the conduct of the public's business?
• Was it prepared, owned, used, or retained by the agency?
• Did it inform a recommendation, communication, employment action, enforcement action, permit, benefit, healthcare communication, school action, or other government decision?
• Is the material stored only in a vendor platform?
• Can the agency search and export it without vendor assistance?
• Is it an identifiable record in a format the agency holds?
• Does a statutory exemption or confidentiality rule apply?
• Can exempt information be segregated from disclosable information?
• Is the human reviewer's role documented?
• Can the agency preserve relevant material during litigation, an audit, an investigation, or an active records request?

Recommended governance action

Create interim internal guidance explaining:

1. 01when AI-assisted work should be preserved;
2. 02which office classifies the record and applies exemptions;
3. 03what prompts, outputs, metadata, versions, logs, and human-review notes are needed for consequential uses;
4. 04how records are searched and exported from vendor systems;
5. 05how confidential, privileged, security-sensitive, and personal information is protected.

California does not establish one retention period for every AI-related record.

Government Code section 34090 generally prevents a city from destroying a city record less than two years old and establishes an approval process involving the legislative body, department head, and city attorney, subject to listed exceptions and other laws.^[4]

Government Code section 26202 provides a separate framework for county records and permits destruction of certain records more than two years old when its conditions are met.^[5]

Government Code section 60201 provides a separate destruction and retention-schedule framework for special districts. It preserves specified categories of records, including records subject to a pending Public Records Act request and records relating to pending claims or litigation.^[25]

Other statutes, local charters, records schedules, litigation holds, audits, grants, employment rules, law-enforcement requirements, education requirements, healthcare requirements, and function-specific laws may require longer retention.

Practical implication

Do not create an arbitrary “AI records retention period” for every system.

Classify AI-related records according to:

• the public function they document;
• whether the entity is a city, county, school district, special district, law-enforcement agency, healthcare entity, or another body;
• the underlying transaction, communication, or decision;
• any use-specific statute;
• the agency's adopted retention schedule;
• litigation, investigation, audit, and public-records constraints.

Minimum inventory fields

For every AI-enabled system, record:

• records created and received;
• prompts and outputs retained;
• model and system versions;
• audit and activity logs available;
• human-review records;
• storage location;
• export formats;
• applicable schedule or retention authority;
• records owner;
• vendor responsibilities;
• legal-hold capability;
• termination, migration, and deletion process.

The California Constitution identifies privacy as an inalienable right and separately protects public access to information concerning the conduct of the public’s business. Local agencies should assess privacy, access, due process, and equal-protection questions together rather than treating transparency or privacy as an absolute answer in every case.^[6]

Civil Code section 1798.29 applies its breach-notification provisions to local agencies. It requires notice when qualifying personal information in agency-owned or licensed computerized data is acquired, or reasonably believed acquired, by an unauthorized person under the statutory conditions. It also addresses agencies that maintain data they do not own and identifies covered data that can include medical information, health-insurance information, specified biometric data, and automated license-plate-recognition information.^[7]

AI procurement implications

Before placing resident, employee, student, patient, public-safety, biometric, location, or other sensitive data in an AI system, determine:

• whether the data may lawfully be used for the proposed purpose;
• whether the vendor may retain or reuse it;
• whether prompts, attachments, outputs, and derived data enter model training or evaluation;
• which subcontractors and model providers receive it;
• where it is stored and processed;
• how the agency learns of unauthorized access;
• who conducts investigation, notice, remediation, and preservation;
• whether deletion requests can be executed without destroying records the agency must retain.

Recommended contract safeguards

Require clear provisions addressing:

1. 01permitted purposes;
2. 02data minimization;
3. 03security standards;
4. 04subcontractors and model providers;
5. 05model training and secondary use;
6. 06incident notification and cooperation;
7. 07preservation of evidence and logs;
8. 08agency access and export;
9. 09deletion and certified destruction;
10. 10conflicts between deletion duties and government retention requirements.

Effective January 1, 2026, Civil Code section 1714.46 provides that, in an action against a defendant that developed, modified, or used AI alleged to have caused harm to the plaintiff, the defendant may not assert as a defense that the AI autonomously caused the harm.^[24]

The section preserves other affirmative defenses, including evidence relevant to causation or foreseeability, and evidence relevant to another person’s or entity’s comparative fault.^[24]

Procurement and governance implications

• Do not assume that labeling a system “autonomous” transfers responsibility away from the agency or vendor.
• Define responsibility for design, configuration, deployment, human review, monitoring, correction, suspension, and incident response.
• Preserve model, version, input, output, human-action, and audit evidence needed to evaluate causation and responsibility.
• Require vendor cooperation, insurance information, indemnity analysis, and access to relevant records, subject to counsel's review.

California’s Fair Employment and Housing Act applies to employers that include the state, political and civil subdivisions, and cities.^[9]

Effective October 1, 2025, California’s Civil Rights Council regulations clarify how existing employment antidiscrimination law applies when employers or their agents use automated-decision systems. The Civil Rights Department explains that such use may violate the law when it harms applicants or employees based on protected characteristics, that disability-related assessments may constitute unlawful medical inquiries, and that relevant automated-decision data falls within employment recordkeeping obligations.^[8]

Government Code section 12946 requires covered employers to preserve applications, personnel, and other listed employment records for at least four years, with longer preservation after notice of a verified complaint.^[9]

Local systems requiring review

• résumé screening and applicant ranking;
• recorded or chatbot interviews;
• personality, aptitude, or game-based assessments;
• background or integrity scoring;
• promotion and succession tools;
• performance evaluation;
• productivity and activity monitoring;
• scheduling and assignment;
• discipline or termination recommendations;
• tools using voice, face, emotion, movement, location, or behavioral data.

Questions for HR, counsel, and procurement

• What employment decision does the system make or facilitate?
• Is the vendor acting as the agency's agent?
• Which inputs, proxies, scores, and classifications are used?
• How was the system tested for the agency's jobs and applicant population?
• Could the tool screen out a person because of a disability?
• What accommodation or alternative process is available?
• Can a qualified human meaningfully review and change the outcome?
• Which automated-decision data, assessments, and employment records must be retained?
• Can the agency reconstruct why a person advanced, was rejected, or received an adverse action?
• Can the vendor change the model or scoring process without agency approval?

Recommended governance action

Require legal and civil-rights review before deployment, document job-related purpose and validation evidence, provide an accommodation route, preserve required employment and automated-decision records, test outcomes, and prohibit a vendor from making material model or scoring changes without notice and review.

Do not state that the regulations create a universal notice, impact-assessment, or appeal procedure for every California local-government AI system.

Penal Code section 13663 applies to defined AI systems that automatically draft police-report narratives from audio or video recorded by in-car or dash-mounted cameras or body-worn cameras, or that analyze an officer’s dictated report to produce an automatically enhanced narrative.^[10]

For covered uses, each law-enforcement agency must maintain a policy requiring:

• identification of each AI program used;
• the prescribed statement that the report was written fully or partly using AI;
• the signature of the officer or agency member verifying review and that the facts are true and correct;
• retention of the first AI-generated draft for as long as the official report;
• an audit trail retained for as long as the official report;
• identification in that audit trail of the user and the video or audio footage used, if any.

The statute also restricts a contracted vendor from sharing, selling, or otherwise using agency information except for the contracted agency’s purposes or under a court order. It permits vendor access for specified purposes including troubleshooting, bias mitigation, accuracy improvement, and system refinement.^[10]

For this section, the “first draft” is the initial document or narrative produced solely by the AI system. Except for the official report, a draft created using AI does not constitute an officer’s statement.^[10]

Procurement and implementation questions

• Does the product fall within the statute's report-drafting definition?
• Can it place the required disclosure in the required location?
• Does the workflow require officer review and signature?
• Is the first AI-generated draft automatically preserved?
• Can the agency preserve the audit trail for the full report-retention period?
• Can the system identify the user and source video or audio?
• Does the vendor contract comply with the statutory data-use restrictions?
• How will permitted vendor access for troubleshooting or refinement be controlled, logged, and reviewed?
• Can the agency retrieve drafts and logs without vendor assistance?
• How are inaccuracies, omissions, hallucinations, and conflicting evidence flagged?

Recommended enhanced review

Treat AI-generated police reports as a high-risk use. Review accuracy, evidence integrity, discoverability, disclosure, retention, security, bias, officer training, supervisory review, incident reporting, and suspension procedures before deployment.

Education Code section 49073.1 applies when a local educational agency contracts with a third party for specified digital storage, management, retrieval, or educational-software services involving pupil records.^[11]

The required contract provisions include:

• continued local educational agency ownership and control of pupil records;
• limits on third-party use;
• parent, guardian, or eligible-pupil review and correction procedures;
• security and confidentiality procedures;
• unauthorized-disclosure notification procedures;
• end-of-contract retention and deletion terms;
• FERPA compliance measures;
• a prohibition on targeted advertising using personally identifiable pupil-record information.

The K–12 Pupil Online Personal Information Protection Act separately restricts covered operators from targeted advertising, building profiles outside K–12 purposes, selling pupil information, and making unauthorized disclosures. It also requires reasonable security and specified deletion practices.^[12]

What this may mean for AI tools

These rules can matter when a school district uses AI tutoring, writing, translation, proctoring, plagiarism detection, assessment, behavioral monitoring, counseling, scheduling, special-education, or administrative tools that access pupil records or covered information.

Do not describe these provisions as one comprehensive school AI governance code.

School-district questions

• Does the tool access pupil records or collect information directly from students?
• Are student prompts, voice recordings, images, documents, search activity, and derived profiles covered?
• Who owns and controls pupil-generated content?
• Can parents, guardians, or eligible pupils review and correct information?
• Can the vendor use data to train or improve unrelated models?
• What deidentification process is used, and can the data reasonably be reidentified?
• What advertising, profiling, or product-development uses occur?
• How are security incidents reported?
• What happens to records and accounts at contract completion?
• What human review, accessibility, language access, and correction process applies?

Health and Safety Code section 1339.75 applies to a health facility, clinic, physician’s office, or office of a group practice using generative AI to create written or verbal patient communications about patient clinical information.^[13]

Covered communications must include:

• a format-appropriate disclaimer stating that the communication was generated by generative AI; and
• clear instructions explaining how the patient may contact a human healthcare provider, employee, or other appropriate person.

The requirement does not apply when the AI-generated communication is read and reviewed by a human licensed or certified healthcare provider before it is sent to the patient.^[13]

AB 489, effective January 1, 2026, prohibits AI or GenAI functionality and advertising from using protected professional terms in a way that indicates or implies that care, advice, reports, or assessments are being provided by a natural person with the relevant healthcare license.^[14]

Local-government meaning

A county hospital, public clinic, local public-health program, or other public entity should determine whether the specific facility, office, communication, and system fall within the statutory definitions. Do not state that every public-health chatbot or administrative message is covered.

Healthcare governance questions

• Does the communication concern patient clinical information or only scheduling, billing, or another administrative matter?
• Does the organization fall within a covered statutory category?
• Is the output read and reviewed by a licensed or certified provider?
• Does the disclaimer appear in the required format?
• Can the patient reach an appropriate human?
• Could the AI imply that it is a licensed professional?
• Are clinical judgment and medical-necessity decisions reserved for qualified humans where required?
• How are patient data, prompts, outputs, corrections, and incidents protected and retained?

This guide provides general policy and governance information. It is not legal advice, a comprehensive statement of California law, or a substitute for review by a city, county, school-district, special-district, healthcare, law-enforcement, or agency attorney. Laws, regulations, guidance, records schedules, bill statuses, and local ordinances may change. Officials should verify current requirements and the facts of a particular use before acting.