When an officer uses AI to draft a report, three things need to survive past the moment of filing for anyone to check the report later: what the tool was given, the AI’s own first draft before anyone touched it, and a record of who used the tool, so that what changed before an officer signed the final version can be shown. California already requires agencies to keep the first draft and an audit trail naming the person who used the AI and any footage it drew on. Utah requires a disclaimer on the report and the author’s certification that it was reviewed for accuracy, but not that the draft be kept. As far as we can tell, no other state has a statute written for AI-drafted reports as of this writing, and the CJIS Security Policy has no provision written for them. Whether that surviving record is later discoverable to the defense is a question for state discovery and Brady rules, not for this page, but a record that was never kept cannot be produced either way.
What changed between the AI draft and the filed report, and how do you show it?
California’s SB 524, chaptered as Penal Code section 13663, sets the clearer floor here. It defines a report’s first draft as “the initial document or narrative produced solely by artificial intelligence” and requires the agency to retain that first draft for as long as the official report itself is retained. Except for the signed official report, no draft created with AI may be treated as the officer’s statement. The first draft and the filed report are two different documents by design, and keeping both is what makes the difference between them visible later, to a supervisor, an auditor, or defense counsel.
Utah’s SB 180 does not go that far. It requires a disclaimer on any report created wholly or partly with generative AI, and a certification from the author that they reviewed it for accuracy, but it does not require the first draft itself to be kept. Under Utah’s statute alone, an agency can meet the letter of the law and still have nothing left to compare the filed report against once the AI’s original output is gone.
Whether either document, the AI’s first draft or the audit record around it, is discoverable to the defense is decided by the jurisdiction’s own discovery rules and, in a criminal case, by Brady v. Maryland’s disclosure obligations for material favorable to the defense. That is a state-by-state and often a case-by-case question, and this page will not generalize an answer for it. What the retention requirement settles is narrower and comes first: whether the record exists at all for a court to ever rule on.
Who is accountable for what the AI was given?
California’s statute puts a name on two different things, and it helps to keep them separate. It requires the signature of the officer or agency member who prepared the report, “verifying that they reviewed the contents of that report and that the facts contained in the official report are true and correct.” That signature is what makes the report the officer’s statement at all. Separately, California requires the agency’s audit trail to identify the person who used artificial intelligence to create the report and, if any, the video or audio footage used to create it. The signer is accountable for the content. The audit trail is what makes the act of using the tool itself attributable to a specific person, which is a different fact and a different failure mode if it goes missing.
Utah folds both into one certification: the author of the report certifies they read and reviewed it for accuracy. Utah also places a second layer of accountability on the agency itself, which must maintain a policy naming which generative AI technologies staff may use, for which tasks, with an acknowledgment that violating the policy can bring administrative discipline. Neither state makes the vendor accountable for what a report says. California does reach the vendor on a different point: a contracted vendor may not share, sell, or otherwise use the information an agency gives its tool except for the agency’s own purposes or under a court order, though it may still access the processed data for troubleshooting, bias mitigation, accuracy improvement, or system refinement.
Preservation and discoverability are two different questions. A department controls the first one: keep the input, the first draft, and a record of who touched it. Whether a court later orders that record produced is a separate question that state discovery rules and Brady decide, case by case.
What does state law currently require?
As far as we can tell, as of this writing California and Utah are the only two states with a statute specifically addressing AI-drafted police reports, and what each one asks for is different enough that treating them as interchangeable would be a mistake.
| Requirement | California (Penal Code sec. 13663, SB 524) | Utah (Code sec. 53-25-901, -902, SB 180) |
|---|---|---|
| Effective | January 1, 2026 | May 7, 2025 |
| Disclosure on the report | Required, on each page or in the body, with a stated sentence | Required, a disclaimer that the record contains AI-generated content |
| Officer sign-off | Signature verifying review and accuracy | Certification that the author reviewed for accuracy |
| First-draft retention | Required, as long as the official report is retained | Not required |
| Audit trail of who used AI | Required, plus footage used, if any | Not required |
| Agency AI use policy | Only a policy requiring the disclosure and signature above | Required, naming permitted tools and tasks |
| Vendor data restrictions | No sharing, sale, or other use beyond agency purposes or a court order; access for troubleshooting, bias mitigation, accuracy, and refinement allowed | Not addressed |
Neither statute reaches an agency outside its own state, and as far as we can tell, no other state has enacted an equivalent law as of this writing. Bills addressing AI-written police reports have been introduced elsewhere, and at least one prosecutor’s office, not a legislature, has separately told agencies it will not accept reports drafted with AI, which is a charging-office decision rather than a statute. A number of states have enacted broader AI governance laws that could touch a law enforcement agency in some way. Whether any of them reaches the retention or disclosure of an AI-drafted report the way California’s and Utah’s statutes do is unsettled here: it depends on each law’s own text and scope, which this page has not checked state by state. Confirm the current rule in your own state before relying on this page: this is a fast-moving area, and a bill introduced this year can change the answer by the time it is read again.
What should a department’s AI policy require for record keeping?
Even where no statute requires it yet, the two enacted laws describe a floor worth meeting voluntarily. A department writing or updating its own AI use policy should require, at minimum:
- Name the tool and the task. Utah’s statute requires an agency’s policy to say which generative AI technologies staff may use and for which tasks. A policy that only says “AI use must be reviewed” does not meet even that floor.
- Keep the input and the first draft, not only the signed report. Whatever the tool was given, footage, dictation, or notes, and whatever it produced before anyone edited it, retained for as long as the filed report itself is retained. California requires that for the first draft, along with an audit trail naming any footage used, and it is a reasonable standard to adopt regardless of where the department sits.
- Log identity and timing separately from content. Who ran the tool, on which report, and when, is a fact about the process, not about the narrative, and it needs its own record so it survives even if the report is later amended.
- Require a named person to certify the final version. Both enacted statutes require this in substance: a signature or certification that a specific person reviewed the content and stands behind its accuracy.
- Match retention to the schedule the report itself is held under, not a separate schedule invented for AI. A generation record that expires before the case file it supports is not there when discovery finally asks for it.
- Treat the vendor agreement and the audit record as two different documents. A services contract can restrict what a vendor does with agency data, and California’s statute already restricts the vendor directly, but a contract is not itself a log of who used the tool on which report. One does not substitute for the other.
What does the CJIS Security Policy say about AI-drafted reports?
The CJIS Security Policy, at version 6.1 (June 25, 2026), has no provision written for AI-drafted reports or for generative AI. Its one mention of artificial intelligence is in the discussion of malicious code protection (SI-3), as a technique malware detection can use. That is not really a gap. The policy applies to every entity with access to criminal justice information or that operates systems used to process, store, or transmit it, so an AI report-writing tool that receives CJI is such a system, and the policy’s audit and accountability controls apply to it as to any other, with no AI-specific exception and no AI-specific extra requirement. What that means in practice for an agency’s own CJIS audit, and what a state auditor actually asks about AI use today, is covered in what the CJIS Security Policy requires before AI touches CJI, and this page will not repeat that ground. The short version worth carrying over: CJIS sets the floor for controlling and auditing access to criminal justice information. California’s and Utah’s report-specific statutes sit on top of that floor. Neither replaces the other.
Sources
- California Penal Code section 13663, added by SB 524 (2025 to 2026 Regular Session), chapter 587, Statutes of 2025, approved by the Governor and filed with the Secretary of State October 10, 2025. Chaptered bill text, read September 25, 2026.
- Utah Code sections 53-25-901 and 53-25-902, enacted by SB 180 (2025 General Session, chapter 330) as sections 53-25-601 and 53-25-602 and numbered 901 and 902 in the Utah Code, effective May 7, 2025. Enrolled bill text and current code section, read September 25, 2026.
- CJIS Security Policy, version 6.1, June 25, 2026, FBI Criminal Justice Information Services Division. Full policy text, read September 25, 2026.
Where Verillian fits
Verillian governs AI use on the devices your organization manages. A checkpoint on each device sits between your people’s AI tools and agents and the AI providers they reach. For the Anthropic API format it enforces your policy before a request leaves the device; for the other providers your policy names, it records the usage. Each record is signed on the device it came from and hash-chained to the one before it, and your own admin server flags any entry that does not link or verify when it arrives, so the record is tamper-evident and stays on your own infrastructure. Redaction is best-effort, not a guarantee that every value is caught. The admin server runs where you choose: on-prem, private cloud, or air-gapped. Mac is supported today, with Windows and Linux in early access. For a department, that record can help evidence one element of California’s audit trail, the person who used AI, whenever that use starts on a managed device, though it is not a substitute for the first draft the statute says to keep. The architecture is aligned with CJIS Security Policy v6.1, not certified, because compliance with that policy is verified by audits of the agencies that use it, not by certifying products.
Verillian does not see inside a vendor’s own cloud. When a vendor’s service calls a model on the vendor’s servers, as an ambient scribe or a hosted report-writing tool does, the record of what that model received is created on the vendor’s side, and the contract is your lever for it. What Verillian gives you is the record of AI use that starts on your own devices.
Our compliance mappings show the controls the platform is designed to support, the audit trail page shows what a device-side record contains, and the public safety section of the site covers what this looks like for an agency already running AI tools day to day.