The Report Remembers, Until It Doesn’t

The pitch for AI report writing lands like relief. The medic talks through the call, the system takes that spoken narration and the run data, drafts the report, and hands it back to clean up and sign. Twenty minutes of charting in the parking lot becomes two. For a crew at the end of a brutal shift, that is not a gimmick, it is a gift, and the vendor knows it.

Then it deletes the part that proves who wrote it.

What Axon’s Draft One Actually Does

Draft One is the body-camera-to-report tool a lot of public safety agencies are running right now. It listens to the audio, writes the first draft of the report, and lets the human edit before exporting the final version. That part works about as advertised. The problem is what happens at export. The Electronic Frontier Foundation dug into the design and found that the moment the officer exports the finished report, the system erases the original AI-generated draft, and with it any record of what the machine wrote versus what the human changed (Axon’s Draft One Is Designed to Defy Transparency, EFF, https://www.eff.org/deeplinks/2025/07/axons-draft-one-designed-defy-transparency).

It goes further than that. EFF reported there is no built-in way to export a list of all the reports Draft One generated, or even a list of everyone who used it, which makes an after-the-fact audit close to impossible (AI Police Reports: Year In Review, EFF, https://www.eff.org/deeplinks/2025/12/ai-police-reports-year-review). Outside reviewers reached the same place: the tool needs far more oversight and transparency than it currently offers, and the design choices run the other way (AI tool that writes police reports needs better oversight, StateScoop, https://statescoop.com/gen-ai-police-report-transparency-oversight-eff/).

This is not a thought experiment about some far-off rollout. One agency that requires an AI disclosure line on each report was able to isolate more than 3,000 Draft One reports generated in a single four-month window. The volume is already real, and the design that throws away the draft is the default, not the exception.

A Record You Cannot Reconstruct Is Not a Record

For as long as we have run calls, the report has been a human artifact you could question. Who wrote it. What they saw. What they decided and why. That is the whole point of a record in a high-stakes setting: it carries provenance, the chain of who is responsible for each line. You can put a name next to a sentence and ask that person what they meant.

An AI draft that vanishes on export quietly cuts that chain. The danger is not that the report is wrong. The danger is subtler and worse: the report reads clean, and it can no longer tell you who to believe. Was the clinical impression the medic’s own read, built from what they saw in the back of the medic unit, or was it the model’s inference from a noisy audio track that the medic was too gassed at end of shift to fully challenge? Once the draft is gone, nobody can answer that. The document survives. The authorship does not.

That is the move worth naming. We are not automating the writing of the record. We are automating away the evidence of who authored it.

This Is Heading Straight for the ePCR

Right now most of the noise is on the law enforcement side, because that is where Draft One lives. Do not let that fool you into thinking it is someone else’s problem. The same tooling, the same vendors, the same procurement gravity is already reaching toward the ePCR and the incident report. Picture an AI drafting a patient care report from the medic’s spoken handoff, the dispatch data, and the monitor feed, a tired medic giving it a light edit, and the report getting exported at the end of a 24. A year later it is evidence in a malpractice suit, or it is the central document in a CQI review.

The question on the stand is simple. Who wrote this clinical assessment, the medic or the model. If the honest answer is that the system deleted the part of the record that would tell you, that is not a documentation hiccup. That is a credibility problem for the entire department, and it lands on the crew that ran a good call and trusted the tool to remember it straight.

I will be honest about my own habit here. I use AI every day to write, and I am sure most of us do now, going back and forth with it as a thought partner until the final version is right. I never save those drafts. I do not even think about them. In everyday life that is probably fine; nobody is going to question how I worded an email. But if the document is a medic report that someone might have to defend later, the question stops being academic. Where do those drafts even live, and who decided to keep them, or not.

What “Provenance” Actually Means in Our Documentation

Provenance sounds like an IT word. In our world it is the after-action review, the QA process, the preplan, the lesson we swear we will not learn twice. The department learns from its records only if it can trust them, and it can only trust them if it can tell what the human actually observed apart from what the machine inferred. Strip that distinction out and your lessons-learned pipeline is running on sand. You are training the next crew on a record nobody can vouch for.

So the policy question is not whether to use AI to draft reports. That ship is moving and there are real reasons to be on it. The question is whether the department retains the AI draft, the human edits, and the difference between them as part of the permanent record, regardless of what the vendor’s default does. That is the line that protects you in court and the line that keeps your institutional memory honest. Several states are already moving to require AI-generated-report disclosures, and the agencies that define their own provenance and retention rules now will be the ones standing on solid ground later, instead of retrofitting under legal pressure after a report gets challenged.

Control your data before a vendor’s default controls your memory. That is not a slogan. It is a clause you write into the contract before you sign it.

The Department Stays the Author, or It Doesn’t

The tool can write the first draft. It cannot be allowed to erase the authorship, because the authorship is the asset. A department that demands preserved provenance, the draft, the edits, and the diff between them, stays the author of its own record even while it runs someone else’s software. A department that accepts the convenient default discovers, the first time a report is contested, that its documentation cannot be defended and its memory quietly belongs to a vendor it never meant to hand it to.

Knowledge gained once is supposed to deliver value forever. That only holds if the record holds. The report can be written by a machine. It still has to remember who decided, and right now, by design, the most popular version of this tool is built to forget.

Robert Grand is a Battalion Chief at Eugene Springfield Fire with 24 years of service. He writes Frontline Intelligence, a newsletter on operational doctrine, technology, and leadership in Fire & EMS.

If this landed, share it with someone in Fire & EMS who needs to hear it.