This Report Was Written Either Fully or in Part

Since January, every California police report drafted with help from artificial intelligence has had to carry the same sentence, word for word: “This report was written either fully or in part using artificial intelligence.” It is twelve words long. Outside California, the departments using these tools often can't say how many such reports they have filed. The missing sentence is a large part of why.

That sounds like an exaggeration. It isn't, and the reason sits underneath every crime number we publish. Before an incident becomes a row with an offense code, a time and an address, it is a paragraph somebody wrote. More of those paragraphs now begin as a machine draft. In most places, the draft disappears the moment it is used.

How a draft gets made

The most widely discussed tool is Axon's Draft One. According to Axon, it “generates report narratives based on body-worn camera audio.” The officer finishes a call, the audio goes through a language model, and a narrative comes back. Axon says “creativity is turned off.”

The draft arrives with gaps on purpose. The Electronic Frontier Foundation, which investigated the product in July 2025, describes drafts “sprinkled with bracketed placeholders” the officer has to fill in. A peer-reviewed study we cover below quotes Axon describing “intentionally included errors for correction.” Then the officer signs an acknowledgement and the narrative goes into the department's records system, where it looks like any other report.

That last part is the whole story for anyone who works with the data. Once the text lands, nothing in the record says where it came from unless the department makes the officer say so.

“We don't store the original draft”

EFF's investigation turned on one admission. At a public roundtable, Axon's senior principal product manager for generative AI said the company does not keep the first draft, and that this is by design. His stated reason was to avoid “more disclosure headaches” for customers. Axon's own product page agrees on the mechanics: each use is logged “without storing the draft itself.”

So a finished report can't show you which sentences the officer wrote and which the model wrote. Nobody can diff the two, because one of them no longer exists.

The logs don't help much either. EFF found that agencies can export an activity log for one report, or for one officer. It found no way for an agency to export a list of every report that used the tool. Here is how that played out in records requests:

• Lafayette, Indiana.The police department told EFF, “we do not have the ability to create a list of reports created through Draft One.”
• Palm Beach County, Florida.The sheriff's office requires officers to include a disclosure line. It found more than three thousand Draft One reports by searching for that keyword within a timeframe.
• Frederick, Colorado. Axon first told the department there was no filter for Draft One reports. Then it said it tracks which reports use the tool internally, and exported the data itself.

Put those three next to each other. The vendor can count. The agency that bought the tool, and is answerable for its reports, often can't. The one agency in the set that could count did it with a text search on a sentence it had required its own officers to type.

Two states, two very different laws

California's SB 524, signed on October 10, 2025, and now Penal Code section 13663, is the more serious of the two. (The bill carries no effective-date clause, so it took effect under the state's default rule on January 1.) It does four things a data person should care about.

1. The fixed sentence.The report must prominently state the twelve words in our headline, and name “every specific artificial intelligence program used,” either on each page or in the body of the text.
2. The first draft survives.The agency must keep “the first draft created” for as long as it keeps the report. The law defines that draft as the narrative “produced solely by artificial intelligence,” and it says the draft is not an officer's statement.
3. An audit trail. The agency must record who used the tool, and which video and audio footage went into the report.
4. The officer signs. The officer has to verify that the facts in the official report are true and correct.

Read item two against the product manager's quote and you can see what the legislature was aiming at.

Utah's SB 180, in effect since May 7, 2025, is lighter. A report made “wholly or partially” with generative AI needs “a disclaimer that the report or record contains content generated by artificial intelligence” and a certification that the author reviewed it for accuracy. Utah doesn't say what the disclaimer must say or where it goes. It says nothing about keeping drafts.

That difference in wording sounds like a lawyer's detail. For counting, it decides everything. In California, a search for one exact string should find every compliant report in the state. In Utah, each agency can phrase its disclaimer its own way, and a researcher has to discover every variant before the count means anything.

Run the count yourself

If you have narrative text from a California agency, through a records request or a feed that carries it, the count takes one line:

grep -ic "written either fully or in part using artificial intelligence" narratives.txt

Two notes on that. The law requires the sentence, but it can't make every officer type it correctly, so the result is a floor, not a census. And it only works on text. Most public incident feeds we ingest carry an offense label, a time and a location. They don't carry the narrative, for good privacy reasons. So the sentence is there, and almost nobody downstream can see it.

Why the narrative matters to a number

You might reasonably ask why a developer pulling burglary counts should care who wrote a paragraph. Here is the connection. The offense code in your feed is a judgment made from the facts in that paragraph.

We wrote about this in Stitches or No Stitches. Whether an assault is aggravated or simple can depend on how an injury is described. Whether a theft was a burglary depends on whether the narrative says someone entered a structure. A model summarizing camera audio decides what to mention and what to leave out, and the officer edits from there. If drafts lean in a direction, the codes can follow.

We don't know whether they do. Nobody does, and that is the point. Without the first draft you can't measure what the model contributed. Without a list of which reports used it, you can't even compare AI-drafted reports to the rest.

The errors that have surfaced so far are factual, not statistical. In September 2024 the King County Prosecuting Attorney's Office in Washington told police agencies it would not accept AI-assisted narratives. Its memo, published by GeekWire, cited a report that “included a reference to an officer who was not even at the scene.” Axon itself told Frederick that officers had gotten past its read-before-submit guardrails at least three times, according to EFF. A phantom officer is easy to catch in court. A slightly different description of an injury, repeated across thousands of reports, wouldn't be.

We have written before about what happens when speed outruns accuracy in public-safety software. In our post on AI crime alerts the line we kept coming back to was Old Man Trench's: “Systems like this are built to be fast. Safety requires being right.” A report is slower than an alert, and it lasts much longer.

The time it was supposed to save

Axon's pitch is time. Its product page says officers spend “up to 40% of their time” writing reports. That claim has been tested once in a randomized trial.

Ian Adams and colleagues ran it at the Manchester, New Hampshire, police department: 85 officers, 755 reports, randomly split between those who had the tool and those who didn't. The paper, published in the Journal of Experimental Criminology in October 2024, has a title that gives away the ending: “No man's hand: artificial intelligence does not improve police report writing speed.” A full-year check on 6,084 reports came out the same way.

Anchorage ran its own ninety-day trial at about the same time and declined to buy. “We were hoping that it would be providing significant time savings for our officers, but we did not find that to be the case,” Deputy Chief Gina Burington told Alaska Public Media.

One trial in one city doesn't settle it. Officers may get faster with practice, and the tools have changed since 2024. In our view, though, it shifts the burden. If a tool alters how the primary record of a crime gets written, the case for it should rest on more than a vendor's time estimate.

Connecticut seems to have reached a similar place. In April, the Chief State's Attorney imposed a moratorium on using AI to draft criminal reports, with the support of the State Police and the state's chiefs association. (New Haven had been testing Draft One on low-priority calls without arrests.)

If you build on police data

None of this shows up in a crime count yet, and we'd be surprised if it showed up soon. Violent crime is down 5.9 percent this year across the 547 agencies in the Crime Index, and that figure has no way to say how any of its incidents were written up. Still, a few habits are cheap now and expensive later.

1. Keep a provenance field, even if it's empty.If you store narratives, add a nullable flag for AI drafting and fill it from the California sentence where you can. Blank means unknown, never “no.”
2. Don't treat narratives as testimony in a model.If you train a classifier on report text, an unknown share of that text was written by another model. That's a contamination problem, and it's easier to plan for than to discover.
3. Watch adoption dates. When an agency starts using a drafting tool, log the date the way you'd log a records system migration. If offense mix shifts that quarter, you'll want to know what else changed.
4. Ask for the count. Under California law the first drafts and audit trail now exist as records. The public-records mechanics are much like the ones in our post on body camera footage.

Back in Indiana, the Lafayette department answered EFF's request plainly. It could not make a list. Somewhere in its records system are reports that began as a machine's paragraph, signed by officers, filed, coded and counted. They read exactly like all the others, which was the idea.