AI-written police reports and administrative records: where the data they summarize comes from

Updated on September 12, 2026

AI in government has moved past pilots. In Texas, deputies file AI-generated police reports drafted from body-worn camera audio; in Germany, a ministry runs permit applications through a language model; in Italy, a draft decree would extract biometric data from faces at stadiums. Between the fact and the record, a new link has appeared: the model that writes the text.

Most of the debate about AI police reports looks at that link: does the model hallucinate, does anyone read the draft, should the record say that software wrote it. Lawmakers in California and Utah have started to answer, and their answers share a blind spot: they assume the material the model summarizes is sound.

That assumption decides the value of the record. If that material was not certified when captured, the record inherits a gap that no review of the prose can close: review checks whether the story hangs together, not whether the source is genuine. The quality of AI-generated police reports and government records is settled at the source, before generation.

An AI-written government record is a document that a public body issues on the basis of a text generated by a language model from documentary material (audio, photographs, readings, screenshots, uploaded files) and then reviewed and signed by an official. Take Draft One, the Axon tool used by the Johnson County Sheriff’s Office in Texas. According to the COPS Office of the US Department of Justice, Draft One processes body-worn camera audio with a variant of ChatGPT and returns a draft narrative; officers must sign a statement that the report was generated with the tool and that they reviewed and corrected it. 404 Media reports that the draft is ready in about five minutes, after which the deputy edits and exports it. The model decides how the material is told; the material decides what the record is worth.

Where the language model enters AI police reports and administrative records

In AI police reports and administrative records alike, the model enters at the drafting stage, after the facts are collected and before an official signs.

The cases on the table: police reports in Texas, workflows in German public administration, and Italy’s implementation of the AI Act

In Texas, the Johnson County Sheriff’s Office used Draft One for the report on a Flock camera search for a woman who had self-managed an abortion (404 Media). In Germany, the federal digital ministry built Spark Workflow, which converts applicants’ files into Markdown and, with predefined prompts, checks them and drafts legal assessments; Minister Wildberger aims to cut infrastructure procedures by 50 to 80 percent (ministry press release, heise). In Italy, the draft decree adapting national law to the AI Act would extract biometric data from faces at stadiums, stations and demonstrations, searchable for seven days; the data protection authority (July 14, 2026) found it inconsistent with Article 26(10) of the AI Act and asked that it happen only afterwards, on recorded footage, for a specific operational need (Matteo Flora, Agenda Digitale).

What human review checks and what it cannot check

Human review of AI-generated police reports and government records checks the text, not the source: a reviewer can catch an invented sentence or a wrong name, but cannot establish when a photograph was taken, whether a recording was cut, or whether an uploaded file is the original.

The human in the loop is asked to do a job the loop does not equip them for. On netzpolitik.org (August 12, 2026), Stefan Kaufmann observed that Spark Workflow has undergone no scientific validation and has no defined error threshold, that it returns different results for identical input, and that the person in the loop cannot reliably judge whether the output is correct. The Texas case adds another limit. Draft One does not keep the first draft: “we don’t store the original draft and that’s by design”, an Axon product manager told 404 Media. Once the deputy has exported the report, nobody can separate what the model wrote from what the officer wrote. A reviewer who cannot tell where the text came from, and holds no certified record of where the data came from, is verifying coherence, not truth.

The question that decides the value of the record: where does the data come from

The value of an AI-written government record is the value of its source. Attention belongs where the digital provenance of the material is established: at capture.

Photographs, instrument readings, recordings and screenshots: the documentary base of the record

Every government record in the public sector, from the police report to the permit file, rests on photographs, instrument readings, recordings and screenshots. The model reads a transcript or description of this material, never the event, so its data provenance is the provenance of that base: certified time, place and integrity pass to the record, and so does their absence.

The inherited gap: when the source is not certified, the prose does not fill it

A record cannot state more about its source than the source can prove about itself. A challenge rarely concerns the model’s grammar. It asks when the photograph was taken, by whom, and whether it has changed since: chain of custody questions.

The probative value of an AI-written police report or government record does not exceed the probative value of the material it summarizes. Under Federal Rule of Evidence 901, the party offering an item must produce evidence sufficient to support a finding that the item is what it claims to be. Rules 902(13) and 902(14) go further for electronic material: records generated by an electronic process, and data copied from a device or file, are self-authenticating when a qualified person certifies the process, for copied data through digital identification, typically a hash value. None of this applies to the prose the model produced; it applies to the photograph, recording or screenshot underneath it. If those were captured with a hash and a certified timestamp, authentication becomes a matter of checking. If not, the official’s signature is asked to cover a gap it cannot see.

EU law points the same way: under Article 41 of eIDAS, a qualified electronic timestamp carries a presumption of accurate date and time and of data integrity, so a hash sealed with a timestamp at capture gives the base what the model cannot.

What the law asks: human oversight, transparency and reasons grounded in facts

The law on AI in government regulates the system, the drafting layer and the moment of proof, and each layer requires that whoever signs a record can reach the facts behind it.

AI Act: human oversight (Articles 14 and 26) and transparency obligations (Article 50) for public use

Article 14 of the EU AI Act (Regulation (EU) 2024/1689) requires that high-risk AI systems be designed so that the people assigned to oversee them can understand the system’s capacities and limits, interpret its output correctly, decide not to use it or to disregard it, and intervene or stop it (artificialintelligenceact.eu). Article 26 places duties on deployers, public authorities included: assign oversight to competent people, use the system according to its instructions, keep the logs it generates; paragraph 10 restricts post-remote biometric identification in criminal investigations. Annex III lists public uses among high-risk systems: law enforcement, migration, essential public services, justice. These provisions set a standard the source has to meet: an overseer cannot interpret an output correctly without access to the material it describes, and cannot decide to disregard it without knowing what it rests on.

Article 12 requires automatic logging, which underpins the record-keeping duties under Article 26; Article 50 sets the transparency obligations for AI-generated content. The Regulation does not certify what enters the system: a limit of the AI Act public bodies must cover themselves.

Human oversight under Article 14 of the AI Act presupposes that the reviewer can trace the output back to its source. TrueScreen makes every capture independently verifiable, so the decision to accept or set aside a draft can rest on documentation instead of on the draft itself.

Public records and the duty to state the facts: California SB 524, Utah SB 180 and the EU AI Act as the international benchmark

California SB 524 requires a police report drafted in whole or in part by AI to carry a disclosure and the first draft to be retained; Utah SB 180 requires a notice on AI-generated content and the officer’s certification that accuracy was verified (EFF, December 2025).

Neither law reaches the source of AI police reports: a disclosure that AI drafted a use-of-force report says nothing about whether the body-worn camera footage behind it is complete. That is what authentication under Rule 901 asks in court, and where the EU AI Act, with overseers able to understand and disregard the output, is the benchmark for government records.

What must a public authority do when it uses AI in a procedure?

A public authority that uses AI in a procedure must be able to show how each record’s text was produced and where its facts come from:

1. Disclose the use of AI in the record (California SB 524, Utah SB 180, Article 50 of the AI Act).
2. Assign human oversight to competent people who can understand, correct or disregard the output (Articles 14 and 26 of the AI Act).
3. Certify the documentary base before generation: photographs, readings, recordings, screenshots and uploads receive a hash, a digital seal and an official timestamp at capture.
4. Keep the logs and the first draft, so that the system’s output and the official’s changes can be reconstructed (Articles 12 and 26; SB 524).
5. State the facts with a verifiable source: every factual statement traceable to a certified item.

Practical cases: the field inspection and the case report

The report drafted from photographs taken with a service phone that were never certified

A code enforcement officer photographs an unauthorized occupation of public land with a service phone; a language model drafts the report from the photographs, and the officer signs. Weeks later, the party contests the when and the where, and the office holds only gallery files and a recollection.

A photograph taken with a service phone and never certified is a file whose date, place and integrity rest on the word of whoever holds the phone. The metadata written by the camera can be edited with free tools; the file can be re-saved, cropped or replaced without leaving a trace. Certified capture changes the object. At the instant the picture is taken, a SHA-256 hash of the file is computed, following the NIST FIPS 180-4 standard: a 256-bit digest that changes completely if a single bit of the image changes. The hash, the verified metadata (time, position, device identifiers) and the file are then sealed with a digital seal and an official timestamp. Any later modification can be detected by recomputing the hash. What the photographs show, and when and where, no longer depends on the narrative.

The same applies to AI-written police reports drafted from body-worn camera footage, such as a use-of-force report, or to any evidence a law enforcement agency captures in the field. A file whose timestamp can be altered cannot carry the record; photo verification after the fact only raises doubts.

The case report that summarizes documents uploaded by third parties with no proof of when and by whom

A permit office receives drawings and a technical report through a portal; a Spark-type workflow drafts the assessment. Months later, the applicant claims a different version was submitted in time, and the office holds only a portal log and an email thread, both editable by their administrator. A certified screenshot of the portal page, or a certified copy of the file taken at receipt, fixes what arrived, and when. Public bodies use TrueScreen, the Data Authenticity Platform, to give a certain date and demonstrable integrity to documents uploaded by third parties, before a case report summarizes them.

How is the source of an AI-written record certified?

TrueScreen, the Data Authenticity Platform, certifies at the source the material that feeds the police report or government record: photograph, recording, instrument reading or screenshot receive a digital seal and an official timestamp at the instant of capture, with verified provenance metadata and a technical report open to independent, repeatable verification.

Certifying the source of an AI-written record means fixing, at the moment of capture, what the material is, when and where it was acquired, on which device, and that it has not changed since, in a form a third party can check independently. TrueScreen does this with a patented forensic methodology aligned with ISO/IEC 27037, the international standard for handling digital evidence. Capture happens in a controlled environment, so that no person, software or system can alter the content while it is being acquired. The platform then verifies device integrity and the authenticity of metadata (time, position, IP address, identifiers), and produces a technical report in PDF and JSON with the full chain of custody. Certification is completed with an official digital seal and an official timestamp, recognized internationally: the standard Article 41 of eIDAS sets for a qualified timestamp.

Real-time capture inside the platform certifies the material from the moment it exists: the highest probative value. Importing an existing file certifies it from the import onward: TrueScreen attests that it has not changed since, not how it was born. For the documentary base of a government record, capture at the source closes the gap, and the legal value of the timestamp runs from then.

The package is self-contained: an expert witness, the opposing party or a court can recompute the hash and check seal and timestamp with standard tools, without TrueScreen. The seal shows that the file has not changed; the forensic capture shows that it is what it claims to be. During capture, alteration is prevented; afterwards, it is detectable.

TrueScreen guarantees the how, not the what: it attests that the capture is genuine and the content unaltered since, not that what stood in front of the camera is what the operator says, and it does not judge the record’s merits. It produces certified documentation with probative value, which becomes evidence when a court admits it and, like any evidence, can be challenged; source code and logs can then be examined by an expert under NDA.

The result is a clean separation between the certified material (what happened, when and where) and the text the model drafted. A challenge to the text is answered by review; a challenge to the facts, by the certified package. The debate about how the model writes AI police reports and government records becomes a debate about prose.

FAQ: AI-written police reports, government records and data provenance

Can AI be used to write police reports?
+

Yes, and it already happens. Draft One, a tool by Axon, transcribes an officer’s body-worn camera audio and generates a draft narrative that the officer edits, signs and exports. According to the COPS Office of the US Department of Justice (January 2025), it runs on a variant of ChatGPT, and officers must sign a statement that the report was generated with the tool and that they reviewed and corrected it. The Johnson County Sheriff’s Office in Texas used it, as 404 Media reported. Whether such a report holds up depends less on the drafting than on the material it summarizes: if audio, footage and photographs were certified at capture, the report inherits their value; if not, their gaps.

Are AI-written police reports admissible in court?
+

Admissibility is decided case by case, and the fact that AI drafted the text is rarely decisive. What a court examines is whether the report and the material behind it are what they claim to be. Under Federal Rule of Evidence 901, the proponent must offer evidence sufficient to support a finding that an item is authentic; Rules 902(13) and 902(14) allow electronic records and data copied from a device to be self-authenticating when a qualified person certifies the process, typically through a hash value. An AI-written report therefore stands or falls with its body-worn camera audio and photographs. California SB 524 adds a practical aid: the first AI draft must be retained.

What is human oversight under Article 14 of the AI Act?
+

Human oversight under Article 14 of Regulation (EU) 2024/1689 is the requirement that high-risk AI systems be designed so that the people assigned to them can understand the system’s capacities and limits, interpret its output correctly, decide not to use it or to disregard it, and intervene or stop it. Article 26 turns this into a duty for deployers, public authorities included: oversight must be entrusted to competent people, and the logs the system generates must be kept. Annex III lists law enforcement, migration, access to essential public services and justice among the high-risk uses. Oversight of a drafted record is only as good as the reviewer’s access to the material the draft summarizes.

What can a human reviewer not verify in an AI-drafted report?
+

A human reviewer can verify what the draft says and cannot verify where its data came from. Reading the text, an officer or official can catch invented sentences, wrong names, contradictions and missing passages. What they cannot establish from a desk is when a photograph was taken, whether a recording was cut, whether a screenshot was edited, or when and by whom a document was uploaded. Stefan Kaufmann made a related point about Germany’s Spark Workflow on netzpolitik.org in August 2026: the software has no scientific validation and no defined error threshold, and the person in the loop cannot reliably judge whether the output is correct. Those questions are answered by certified capture, not by rereading.

How do you prove where the data summarized in an AI-written report came from?
+

You prove it by certifying the material at the moment it is captured, before any model reads it, in a form a third party can check on their own. With TrueScreen, the Data Authenticity Platform, every photograph, recording, instrument reading or screenshot receives a hash, a digital seal and an official timestamp at capture, with metadata such as time, position and device identifiers verified, and a technical report in PDF and JSON with the full chain of custody. Anyone who receives the package can recompute the hash and check seal and timestamp with standard tools, without TrueScreen. An AI-written report is contestable to the extent that its source is: certifying the material at the origin closes that gap.

Does an AI-written government record have to say that AI was used?
+

In a growing number of jurisdictions, yes. California’s SB 524 (2025) requires a police report drafted in whole or in part by AI to carry a disclosure and requires the first draft to be retained; Utah’s SB 180 (2025) requires a notice on AI-generated content and the officer’s certification that the accuracy was verified, as the Electronic Frontier Foundation summarized in its December 2025 review. In the European Union, Article 50 of the AI Act sets transparency obligations for AI-generated content. A disclosure tells the reader how the text was produced. It says nothing about where the underlying photographs, recordings or files came from, which is a separate question that only certified capture answers.

What does a certified timestamp prove about a photograph or a recording?
+

A certified timestamp proves that a specific file existed, in exactly that form, at a specific moment. Under Article 41 of the eIDAS Regulation (EU) 910/2014, a qualified electronic timestamp enjoys a presumption of the accuracy of the date and time it indicates and of the integrity of the data bound to it. In practice the timestamp is applied to a hash of the file, such as a SHA-256 digest under NIST FIPS 180-4; if the file changes later, the hash no longer matches. A timestamp does not prove how the file was created, which is why timestamping at the instant of capture, with verified metadata, is worth more than timestamping an existing file later.