State AI-in-Policing Disclosure Laws: What California and Utah Now Require for AI-Written Police Reports

Two states now have statutes on the books that say, in effect: if a machine helped write it, the report has to say so. California’s SB 524 took effect January 1, 2026, and Utah’s SB 180 has been in force since May 2025 — making Utah first to the issue and California the more prescriptive of the two. Meanwhile, one of Washington State’s largest prosecuting offices has taken a different position entirely: it won’t accept AI-assisted police narratives at all, regardless of what disclosure accompanies them.

For agencies watching this unfold from outside California and Utah, the practical question isn’t academic. Report-writing tools like Axon’s Draft One are already in wide use, more vendors are entering the space, and the gap between “what the law requires,” “what a specific prosecutor’s office will accept,” and “what a defensible practice actually looks like” is where departments get exposed. This is a plain-English comparison of the two enacted laws, the King County policy that goes further than either, and what the difference means for agencies building their own AI-in-report-writing policy in 2026 — a policy that, per our governance framework for efficiency AI versus decision AI, needs to sit alongside oversight for every other AI tool the agency runs, not just report drafting.

What California’s SB 524 Actually Requires

SB 524 amends the California Penal Code (adding Section 13663) and applies to any official law enforcement report generated using generative AI, in whole or in part. The requirements are specific rather than aspirational:

• Page-level disclosure. Every page of a report that used AI assistance must carry a conspicuous statement — the law specifies language along the lines of “This report was written either fully or in part using artificial intelligence” — along with identification of the specific AI program used.
• Officer sign-off. The officer or agency member who used the tool must sign to verify they personally reviewed the content and that the facts stated are true and correct to the best of their knowledge. The AI draft cannot substitute for the officer’s own certified statement.
• Audit trail and retention. Agencies must retain the first AI-generated draft — before any human edits — for as long as the final report is retained. That preserved draft, plus a record of the underlying inputs (such as body-worn camera audio or video that fed the AI tool), must be available to prosecutors, defense counsel, and auditors.
• Vendor data restrictions. Companies that supply the AI tool are barred from selling or otherwise using the department’s input data beyond what’s needed to provide the contracted service.

The retention requirement is the one creating the most operational friction. Some widely deployed report-writing tools were not built to preserve a clean, unedited first draft as a separate artifact — meaning agencies adopting or continuing to use them may need vendor-side changes, not just a new disclosure line in the report template, to actually comply.

What Utah’s SB 180 Requires — And How It Differs

Utah moved first. Governor Spencer Cox signed SB 180 in March 2025 and it took effect May 7, 2025, roughly eight months before California’s law came online. The core requirements are lighter-touch than California’s:

• Agency-level AI policy. Every Utah law enforcement agency must adopt a written policy specifying which AI technologies it permits, and where and how they may be used.
• Disclosure in the report. Any report generated wholly or partly by generative AI must disclose that fact.
• Accuracy certification. The officer who authored or approved the report must certify, in a form with legal weight, that they reviewed it for accuracy.

What Utah’s statute does not impose, compared to California’s: there is no statutory page-by-page disclosure format, no mandated preservation of the AI’s unedited first draft as a discrete audit artifact, and no vendor data-use restriction written into the law itself. Utah leaves more of the specifics — what the disclosure looks like, how review is documented day to day — to each agency’s own policy, rather than dictating exact report language and retention rules at the state level.

Practically: an agency that is SB 180-compliant in Utah is not automatically SB 524-compliant if it operates in California, and vice versa. Multi-state vendors and any agency drafting a policy that might need to satisfy either state’s court system should build to California’s more specific bar rather than Utah’s floor.

The King County Prosecutor’s Office Goes Further Than Either Law

Neither California’s nor Utah’s statute is the strictest position in the country. That belongs to the King County Prosecuting Attorney’s Office in Washington State, which has taken the position that no amount of disclosure fixes the underlying reliability problem — and has said so since 2024, well before either state law existed.

In a memo to area police chiefs, Chief Deputy Prosecutor Daniel J. Clark stated the office would not accept any police narrative “produced with the assistance of AI,” citing a specific incident in which an AI-assisted draft referenced an officer who was not present at the scene — an error Clark noted could easily survive review given the volume of material officers process on deadline, and one that could prove “devastating” for a case, a community, and the officer’s own credibility on the stand. As of reporting through early 2026, King County’s position has not changed, even as the office has fielded vendor pitches (including from Axon) for newer AI tools; the office has publicly said it does not want the county to be “the guinea pig for the rest of the nation.”

This is a prosecutorial charging-and-evidentiary policy, not a state law — it binds what King County’s own office will accept into a case file, not what a department elsewhere in Washington or in another state must do. But it matters for two reasons. First, it shows that disclosure-plus-certification, the model both California and Utah adopted, is not universally regarded as sufficient even by legal professionals working directly with these reports. Second, any agency operating in or near a jurisdiction with a similarly skeptical prosecutor’s office should confirm that office’s evidentiary posture directly — a report that is fully compliant with a state disclosure statute can still be rejected by the specific prosecutor deciding whether to charge the case.

Legally Required vs. Merely Best Practice

For agencies outside California and Utah trying to build a defensible policy now, it helps to separate what’s actually mandated somewhere from what’s simply good practice that no statute currently requires nationally.

Legally required (where applicable):

• Disclosure that AI was used, in some form (California, Utah)
• Officer certification of accuracy (California, Utah)
• Preservation of the original AI draft as a distinct record (California only)
• A written agency policy governing permitted AI tools and use cases (Utah; California requires agency policies by the same January 1, 2026 date, addressing similar ground)

Not required by either statute, but increasingly treated as best practice:

• Independent, tool-agnostic accuracy audits of a sample of AI-assisted reports on a recurring basis
• Documented training on what the AI tool does and does not reliably capture (nuance, sarcasm, overlapping speech in body-camera audio) before an officer is authorized to use it
• A defined escalation path when an officer or reviewer catches the AI tool inserting or omitting a material fact
• Contractual data-use limits with the vendor that go beyond the state’s minimum, particularly regarding whether body-camera audio processed by the AI is used to further train the vendor’s models

Agencies that only clear the statutory floor — a disclosure line and a signature — are meeting the law, not necessarily building a report-writing program that survives cross-examination.

What This Means for Agencies Building a Policy in 2026

An agency without a California- or Utah-style state mandate is not in a regulatory vacuum. Two enacted state laws provide a live template for what a disclosure regime looks like in practice, and King County’s office has documented, in public, exactly the kind of error that motivated the strictest response in the country. Three practical takeaways follow:

1. Build to the stricter template even without a state mandate. California’s audit-trail and unedited-first-draft retention requirement is more defensible in litigation than a bare disclosure line, and it costs little to require of a vendor at the procurement stage rather than retrofit later.
2. Confirm the receiving prosecutor’s office position before scaling adoption. A tool that is technically legal to use under a hypothetical state disclosure law can still result in a case a prosecutor’s office declines to accept, as King County has shown. That conversation belongs in the pilot phase, not after a caseload depends on the tool.
3. Treat vendor compliance claims skeptically. SB 524’s retention requirement exposed a real gap between what some popular report-writing products were built to preserve and what the law now demands. “Compliant” marketing language from a vendor is not the same as a demonstrated audit trail that a defense attorney can actually request and receive.

Frequently Asked Questions

Does federal law require disclosure of AI-written police reports?

No. There is currently no federal statute requiring disclosure or accuracy certification for AI-assisted police report writing. The requirements described here come from individual state legislatures — so far California and Utah — and federal agencies have not issued a binding equivalent as of this writing.

Which AI report-writing tools do these laws cover?

Both California’s SB 524 and Utah’s SB 180 are written to cover generative AI tools broadly, not any single named product, though commercial tools such as Axon’s Draft One are the most frequently cited examples in legislative debate and press coverage. If a tool uses generative AI to produce report language from inputs like body-camera audio, it falls within scope regardless of vendor.

Is King County’s policy a law, or can it change?

It is an internal prosecutorial policy set by the King County Prosecuting Attorney’s Office, not a statute, and it can be revised by that office at its own discretion. It has remained in place since it was first issued in 2024 and, per public reporting into 2026, has not been rescinded, though the office has continued evaluating newer vendor tools.

If my agency isn’t in California or Utah, do these laws apply to us?

No — state statutes only bind agencies operating within that state. However, agencies elsewhere should expect similar bills to be introduced as more legislatures take up AI-in-policing oversight, and building toward the California/Utah disclosure-and-certification model now reduces the compliance work if a similar law passes locally.

What happens if an agency violates SB 524 or SB 180?

Publicly available summaries of either bill do not appear to specify a standalone criminal or civil penalty tied to a documentation lapse. In practice, the more immediate exposure is evidentiary and reputational: a report lacking required disclosure or a preserved audit trail is more vulnerable to suppression motions and cross-examination on reliability, independent of any statutory penalty.

Agencies with specific compliance questions should direct them to their state’s peace officer standards board or legal counsel rather than relying on vendor guidance alone, since implementation details (exact disclosure wording, retention formats) are still being worked out by early-adopting departments.

The documentation discipline these laws demand for report drafting is the same discipline oversight bodies are now expecting for algorithmic decision-support tools more broadly — see our companion piece on predictive policing audit frameworks for what a defensible audit record looks like once an agency moves beyond report-writing AI into risk-scoring and deployment tools.