The Police Report Wrote Itself.
Did Anyone Check It Against the Footage?
July 2026 · VdVLaw
When AI drafts the first version, the absurdity is not the technology. It is treating the summary like the source.
Summaries are not facts. The report is the document. The footage is the source.
The legal profession got there first
Let us start with the legal profession, because apparently lawyers volunteered to be the crash-test dummies for this particular technology.
A lawyer asks AI for case law. The AI responds with beautiful citations, confident quotations, and the warm glow of fake competence. The lawyer files it. Then a judge does the one thing the software was desperately hoping nobody would do. .......The judge checks.
Suddenly the perfect case is not a case at all. Not bad law. Not distinguishable. Not from the wrong jurisdiction. Just imaginary. The legal equivalent of citing People v. Trust Me Bro and hoping the court is too busy to notice.
This has happened often enough that it is no longer a cautionary tale. It is a genre. In June 2026, the Ninth Circuit sanctioned two attorneys after briefs contained nonexistent cases, misattributed quotations, and serious misrepresentations of real authority. A California attorney was fined $10,000 after an appellate court found that 21 of 23 case quotations in his opening brief were fabricated. A Colorado attorney accepted discipline that included a 90-day actual suspension after relying on AI-generated fake authority.
AI did not invent laziness. It just gave laziness a very nice font.
So let us be fair. This is not only a police problem. It is not only a defense problem. It is not only a prosecutor problem. This is a human problem. AI did not invent laziness. It just gave laziness a very nice font.
The lesson is simple: AI can produce something that looks finished, sounds confident, and may be partly nonsense. The machine does not know it is nonsense. That is part of the charm, if by charm we mean a small professional liability grenade wearing a bowtie. The checking belongs to the human.
Now hold that thought, because the same problem has walked across the courtroom and picked up a badge.
The first story in the case
Police reports have always mattered more than they probably should. They shape charging decisions, plea negotiations, bail arguments, sentencing conversations, investigation strategy, witness preparation, and the first impression of nearly everyone who touches a criminal case. A police report is not evidence the way body camera footage is evidence, but in practice it often becomes the first working version of the case.
The report becomes the story. Not because everyone is careless, but because criminal cases move fast, calendars are crowded, and the first clean narrative has gravity. The prosecutor reads it. The defense reads it. The client reacts to it. The court may hear it summarized. The report may not be the evidence, but it often gets treated like the tour guide wearing a little hat and confidently pointing at the exhibits.
Now that first version may not be written entirely by the officer.
Axon Draft One uses body camera audio to generate draft police report narratives that an officer can review, revise, and copy into a reporting system. The U.S. Department of Justice COPS Office describes these tools as producing a first draft from body camera audio after the footage is uploaded and transcribed.
The report wants to become the story. The footage may have other plans.
That may save time. It may also create a polished summary of something the tool did not actually see. And that is where the defense work begins.
The machine heard the movie playing in the next room
Here is the part that sounds like a joke and is instead just true: the tool works from audio. It does not watch the video the way a person watches the video.
The COPS Office explains that the transcription is based entirely on audio, that officers are encouraged to narrate the situation in real time, and that these AI tools are not able to parse or summarize the video's visual content.
Sit with that for a second. A document whose job is to describe a physical encounter may now begin with a tool that experiences the encounter the way you experience a movie playing in the next room. It hears the dialogue. It misses the picture. Then everyone acts impressed because the summary arrived quickly. That is not analysis. That is fast paperwork wearing a little crown.
Because the tool works from audio, officers are encouraged to narrate what they are seeing in real time. In theory, that helps the report. In practice, it can also make body camera audio sound like a nature documentary hosted by the arresting officer.
'Here we observe the subject moving his left hand.' Lovely. David Attenborough would like a word.
But did the video show that? Did the hand move before or after the command? Was the object actually reachable? Was the person confused, injured, intoxicated, frightened, off balance, or trying to comply? Was someone else speaking? Did the officer's narration describe what happened, or did it describe what the officer believed was happening?
Summaries are not facts
A police report is a summary. An AI draft is a summary. A transcript is a summary. Attorney notes are summaries. Even a good discovery memo is a summary.
Some summaries are useful. Some are necessary. Some are accurate. Some are dangerously clean. But none of them are the event.
The source matters: the footage, the audio, the timestamp, the medical record, the witness statement, the jail call, the thing said before the quoted line, and the thing visible on the edge of the frame that somehow never made it into the report, probably because inconvenient facts are famously bad at networking.
This is the real absurdity of relying on AI without human review. The human asks the machine for a shortcut. The machine gives the human a confident summary. The human treats the summary like the source. Then everyone stands around surprised when the shortcut becomes the scenic route to sanctions, suppression issues, impeachment, or professional humiliation.
A summary can point you somewhere. It cannot sit in the witness chair wearing a little bowtie.
AI can be useful. Nobody serious should pretend otherwise. But the usefulness ends the moment the human stops checking. A summary can point you toward the source. It cannot replace the source. A report can tell you what someone says happened. It cannot prove that the footage agrees. A transcript can tell you what the software thinks it heard. It cannot tell you tone, hesitation, confusion, body language, timing, or what was visible but never spoken.
That is where discovery review matters. Not because the attorney needs someone to make folders pretty. Pretty folders are nice. We are all pro-folder. A labeled folder is the little mint on the pillow of litigation.
But a mint is not a defense.
The clean version is not always the true version
The danger of an AI-assisted report is not that it sounds sloppy. The danger is that it sounds good.
AI does not get flustered. It does not write, 'This happened fast and I may have missed something.' It does not pause and say, 'I am not sure whether that was resistance or confusion.' It produces organized prose with the serene confidence of something that was definitely there, despite the small inconvenience that it was not. It heard a recording. That is not the same as understanding the encounter.
The report does not have to be completely false to be dangerous. A completely false report waves its arms and announces itself like a drunk uncle at a wedding. The harder problem is the mostly accurate report: mostly accurate but incomplete, technically correct but misleading, organized but missing the one fact that changes the meaning of everything around it.
The audio may capture 'stop resisting.' The footage may show a person trying to understand a command, shifting weight on uneven ground, reacting to pain, or being held in a position where compliance is not as simple as the report later makes it sound. The report may describe intoxication while the footage shows poor lighting, exhaustion, a medical issue, a language barrier, fear, or a stumble that has less to do with alcohol and more to do with being human under pressure. The report may describe a clean sequence. The video may show a messy one.
A clean report is not the same thing as an accurate case.
None of that appears magically because the file is organized. Someone has to watch. Someone has to listen. Someone has to compare.
What attorney-directed discovery review brings to the table
The attorney keeps the strategy. That part matters. Discovery review is not about replacing the attorney's judgment. It is about giving the attorney a better board to play on.
Working under the attorney's direction, I review the discovery through the issues the attorney identifies: suppression, force, identification, intoxication, mental health, medical explanation, timeline, witness credibility, officer credibility, mitigation, or whatever else actually matters in that case. Sometimes the theory is clear. Sometimes the theory is still forming. Sometimes the file is doing that charming thing where it points in three directions and makes everyone pretend that is normal.
I am not just asking, 'What is in the file?' I am asking, 'What does the file do to the attorney's theory?'
That means reading the report against the source. Watching the video instead of trusting the description of the video. Listening to the audio instead of trusting the transcript. Checking the sequence against the timestamps. Looking for the place where the report sounds more certain than the source allows. Looking for the fact that appears in the footage but never appears in the report. Looking for the officer's interpretation dressed up as observation. Looking for the sentence that seems harmless until the ten seconds before it make it very much not harmless.
Organized discovery tells you where the file is. Source-tested review tells you what the file does.
That is not file organization. That is source-tested analysis.
The result should not be a pile of notes that makes the attorney do the work twice. That would be rude, and worse, inefficient. The result should be a usable map: what the report says, where the source is, what the source actually shows, why the difference matters, and what question comes next. Sometimes that question is for the officer. Sometimes it is for the client. Sometimes it is for a witness, investigator, toxicologist, medical expert, or mitigation specialist. Sometimes it is simply the question that helps the attorney stop a bad assumption from becoming the theory of the case.
California changed the rules, but disclosure is not review
California responded to AI-assisted police reports with Senate Bill 524, now codified at Penal Code section 13663. The law requires agencies using AI for official reports to disclose AI use, identify the AI program used, retain the first AI-created draft for as long as the official report is retained, and maintain an audit trail identifying the person who used AI and the video or audio footage used to create the report, if any. The statute also makes clear that, except for the final signed official report, an AI-created draft does not constitute an officer's statement.
That is important. It gives the defense something concrete to ask for. But disclosure is not review. A label saying 'AI was used' does not tell you whether the report is accurate. A retained first draft may show where the AI started, but it does not necessarily show every later edit, omission, emphasis, or wording choice.
Axon's current product guide says Draft One can retain original AI-generated drafts, but it also explains that only the original, unedited version is saved. Once the user inserts the draft into Report Writer and begins editing, those edited versions are not saved. Axon also describes the retention setting as off by default so agencies can update policies or retention schedules first.
So even with California's statute, the defense cannot assume everything arrives in one tidy bundle with a bow on it. The attorney has to ask. Then someone has to compare. The robot is not going to raise the issue for you. That would be convenient, but apparently the robot is busy inventing appellate authority and writing confident summaries of video it did not see.
Disclosure tells you AI was involved. Review tells you whether the report matches the source.
The point
There is another wrinkle, because optimism is not a discovery strategy. California has a broad Public Records Act exemption for law enforcement investigatory records. That does not mean a criminal defense attorney cannot seek discoverable material in a criminal case. Public records access and criminal discovery are different worlds. But it does mean the public may not be able to evaluate this technology case by case, and it means the defense cannot assume that someone else has already checked whether the report matches the recording. Professional-looking documents are exactly how we ended up with lawyers citing cases from the enchanted forest.
The report is not the event.
It is the written version of the event, and now sometimes it is the machine-polished written version of the audio version of the officer's version of the event. Which is a lot of versions before anyone has actually watched the thing we all agreed to call body camera footage.
That is the absurdity. We built a system where the camera records the encounter, the software listens to only part of it, the report turns that partial version into clean paragraphs, and then everyone acts surprised when the clean paragraphs need to be checked against reality.
The absurdity is not that AI made a summary. The absurdity is pretending the summary checked itself.
Because the footage is the closest thing the file has to the event itself. Not perfect. Not complete. Not magic. But closer than a summary wearing a badge and a spell-check subscription.
Summaries are not facts. The report is the document. The footage is the source.
The work is finding the daylight between the document and the source. That is where the easy story starts to wobble. That is where better questions come from. That is where impeachment can begin. And sometimes, after everyone has admired the very professional-looking report, that is where the case quietly raises its hand and says: Actually, that is not what happened.
Sources
1. United States Court of Appeals for the Ninth Circuit, Lnu v. Blanche, No. 24-4790, Order filed June 3, 2026. Sanctions order involving nonexistent cases, misattributed quotations, gross misrepresentations of real cases, and AI-related candor issues. https://cdn.ca9.uscourts.gov/datastore/opinions/2026/06/03/24-4790.pdf
2. CalMatters, California lawyer's ChatGPT use is why courts want AI regulation. Reports the $10,000 fine after a California appellate court found that 21 of 23 case quotations in an opening brief were fabricated. https://calmatters.org/economy/technology/2025/09/chatgpt-lawyer-fine-ai-regulation/
3. People v. Zachariah C. Crabill, 23PDJ067, Colorado Presiding Disciplinary Judge, Stipulation to Discipline. Colorado attorney discipline involving AI-generated fictitious cases, with a 90-day actual suspension and probationary conditions. https://www.coloradolegalregulation.com/wp-content/uploads/PDJ/Decisions/Crabill%2C%20Stipulation%20to%20Discipline%2C%2023PDJ067%2C%2011-22-23.pdf
4. Axon, Draft One Product Guide: Draft One. Describes Draft One as powered by body camera audio and used to generate draft narratives for review and revision. https://www.axon.com/help/draft-one/software/draft-one/draft-one.htm
5. U.S. Department of Justice, Office of Community Oriented Policing Services, Using AI to Write Police Reports. Explains that AI police-report tools work from body camera audio, that transcription is based entirely on audio, and that the tools do not parse or summarize visual video content. https://cops.usdoj.gov/html/dispatch/01-2025/ai_reports.html
6. California Penal Code section 13663. Requires AI-assisted official reports to disclose AI use, identify the AI program used, include a verifying signature, retain the first AI-created draft, and maintain an audit trail. https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=PEN§ionNum=13663
7. California SB 524, Law enforcement agencies: artificial intelligence. Bill history and statutory background for California's AI-assisted police report requirements. https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB524
8. Axon, Draft One Product Guide: Auditing and reporting. Explains original AI-generated draft retention, that only the original unedited version is saved, edited versions are not saved, and the retention setting is off by default. https://www.axon.com/help/draft-one/software/draft-one/auditing-reporting.htm
9. California Government Code section 7923.600. Provides the Public Records Act exemption for certain law enforcement investigatory records, subject to statutory exceptions. https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=GOV§ionNum=7923.600
10. First Amendment Coalition, California Public Records Act: Investigatory Records. Explains that law enforcement investigatory reports and files are generally not required to be made public under the investigatory-records exemption. https://firstamendmentcoalition.org/resources/explainers/explainer-cpra-investigatory-records/
VdVLaw provides attorney-directed legal support services and is not a law firm. Robert van der Vijver is not an attorney. Nothing in this article is legal advice, and nothing here creates an attorney-client relationship. This article is general information written for attorneys and does not address the facts of any particular case. Criminal discovery obligations, public records access, evidentiary issues, retention issues, ethical duties, and law enforcement technology policies can vary by case, agency, county, court, and date. Statutes, rules, policies, and technology features change over time. Attorneys remain solely responsible for their own legal analysis, discovery decisions, ethical duties, litigation strategy, and verification of all statutes, rules, agency policies, discovery responses, and source materials.