Detective Marcus Webb had twelve seconds of footage that changed a case. The AI-drafted report stated that the suspect "approached the officer in a threatening manner with both hands raised." Webb had adopted the draft on a long shift and submitted it. Three weeks later, defense counsel played the body-worn camera (BWC, the recording device on the officer's uniform) footage in a preliminary hearing and the judge watched those twelve seconds. Both hands were not raised. One hand was raised. The other was at the suspect's side. "Threatening manner" was, at a minimum, in dispute. The narrative that "both hands were raised" was nowhere in the footage. The AI had assembled the description from the audio, from a pattern of language typical in use-of-force incidents, and from a gap in what the microphone could detect. The phrase it produced was plausible, confident, and wrong. Webb spent four hours on the stand explaining a sentence he had not personally written, had not personally verified, and now could not personally defend.

What Footage-Grounded Means and Why It Matters

The verification pass described in this lesson has a specific name: the footage-grounded verification pass. Not the "review," not the "check," and not the "edit." Each word in the phrase matters.

"Footage-grounded" means the footage is the source of truth, not the AI draft, not your memory, not what typically happens on calls like this one. The BWC recording, supplemented by the CAD (computer-aided dispatch) entry, any surveillance footage, any dashcam recording, and your contemporaneous field notes, is the evidentiary record. Every factual claim in the narrative must be traceable to that record. If it is, the claim is grounded. If it is not, the claim is ungrounded, which is another name for what the AI literature calls a hallucination and what prosecutors call a problem.

"Verification" means active checking, not passive reading. Reading through a draft and thinking it sounds right is not verification. Verification means opening the footage, going to the relevant timestamp, and confirming that what the draft says happened is, in fact, what the camera recorded happening. That process takes longer than a read-through. It is supposed to. The time it takes is the time between a report that holds up and a report that is picked apart in a suppression motion.

"Pass" means you do it systematically, all the way through, for every factual claim. Spot-checking is not a pass. Checking the sections you remember being uncertain about is not a pass. A verification pass covers every sentence with a factual claim, from the first line of the narrative to the last. If a sentence describes something, attributes something, or asserts something about what happened, it gets verified.

The footage is not a supplement to the report. The report is a description of the footage. If they do not match, the report is wrong, not the footage.

The Gap-Fill Mechanism: How It Happens

To run a good verification pass, you need to understand exactly how gap-fills happen, because that understanding shapes what you look for. A gap-fill is not a random error. It follows a predictable pattern. Knowing the pattern makes it catchable.

A language model generating a police report narrative from BWC audio does not have access to what the camera saw. It has access to what the microphone captured: voices, ambient sound, partial statements, the sounds of movement. In a high-stress incident, the audio is often incomplete. There are moments where no one is speaking, where voices overlap, where the audio quality drops, or where a significant physical action happens without any spoken description. The model encounters these gaps and, because its job is to produce a coherent, complete narrative, it fills them. It fills them with what is statistically likely to be true given the call type, the preceding audio, and the patterns in its training data.

That is why Detective Webb's draft said "both hands raised." The model encountered audio that included sounds of a physical confrontation and language patterns consistent with use-of-force reports, and it produced a description that was coherent with those patterns. The phrase "hands raised" appears frequently in use-of-force narratives in contexts related to compliance or threat assessment. The model did not lie. It did not fabricate in the sense of intent. It completed the pattern. It was wrong.

The Three Gap-Fill Signatures

Gap-fills are recognizable when you know their signatures. There are three patterns that should send you immediately to the footage to verify.

The first signature is specificity without audio support. The draft says something specific: a specific hand position, a specific direction, a specific item in a specific location. You remember the moment in the call, but you do not specifically remember saying or narrating anything about that detail. Go to the footage. If the audio does not support the claim and the footage confirms it, fine: you can verify it from the footage. If neither the audio nor the footage supports it, you have a gap-fill. The model invented a detail to make the narrative complete.

The second signature is use-of-force boilerplate. Phrases like "threatening manner," "aggressive posture," "closing distance rapidly," and "preparing to strike" appear with high frequency in use-of-force narratives and therefore appear with high frequency in the training data of any model that has been trained on law enforcement reports. When the AI encounters an audio signal consistent with a physical confrontation, it reaches for this vocabulary. Some of it may be accurate. Some of it may be a pattern-completion gap-fill. Any phrase in a use-of-force section that you cannot specifically tie to something you observed, verbalized on camera, or captured on footage needs verification. That section of the report is the highest-stakes section, and it is also where the model is most likely to produce language it learned from other reports rather than language grounded in this specific recording.

The third signature is scene description without audio cues. The model cannot see the scene. It can infer from audio: if someone says "put that down," there may be an object present. If someone's voice echoes, the environment may be large or empty. But detailed scene descriptions, descriptions of lighting, of how objects were arranged, of environmental conditions, have no audio source. They come from the model's inferences and patterns. A scene description that is vivid and specific is not necessarily accurate. Every specific detail in a scene description needs a footage source.

Running the Verification Pass: Step by Step

Here is how to run the footage-grounded verification pass in practice. This is a workflow, not a checklist of boxes. It requires your active judgment throughout.

Set Up the Comparison Environment

Open the draft and the footage at the same time. Most BWC platforms let you timestamp and annotate while viewing. If yours does, use that capability. You want to be able to move from a sentence in the draft to the corresponding moment in the footage without having to hold the timestamp in memory. Note the start time of the call and use the CAD entry to anchor your timestamps. If there are multiple clips, note which clip covers which time period.

Before you begin the comparison, read through the draft once without the footage to get the complete narrative in your head. Note any sections where you think "I'm not sure about that" or "that doesn't sound quite right." Those are the primary targets. Do not stop there, because the most dangerous gap-fills are the ones that do not raise any alarm. They are plausible. They fit. They are just wrong.

Work Paragraph by Paragraph

Take each paragraph in order. For each paragraph, identify every factual claim. A factual claim is any assertion about what happened, where, when, or how: any action, any object, any location, any statement, any description of a person's demeanor or physical state. Mark each one, whether that is a mental note, a highlight, or a physical mark on a printout.

For each marked claim, identify the timestamp in the footage where that claim can be verified. Open the footage to that timestamp. Confirm. If it matches, mark it verified. If it does not match, or if there is no footage segment that addresses the claim, mark it unverified and stop. An unverified claim does not go into the sworn report. It either gets corrected to what the footage shows, removed from the narrative, or flagged with a note that it is based on your personal observation (not the footage) with a specific explanation of what you observed.

This last point is important. There are things in any incident that your BWC footage does not capture: what you observed before the camera activated, what happened at the periphery of the frame, your own physical experience, things said while the audio was degraded. Personal observation is a valid source for a police report. The difference between personal observation and a gap-fill is that personal observation is something you actually observed. If you are writing a claim based on personal observation rather than footage, say so in the narrative: "I observed..." rather than "the subject was." That phrasing signals that the source is your direct observation, which is traceable to you as a witness, not to the camera as a recording device.

The Special Case of Quoted Statements

Quoted statements require a separate treatment within the verification pass. Any time the draft puts words in quotation marks or closely paraphrases a statement as something a specific person said, go to the audio. Find the moment. Listen. Do the words in the draft match the words in the recording?

This is a verbatim test for anything that appears in quotation marks. It is a substance test for close paraphrases. The distinction matters in court because a direct quote implies verbatim accuracy. If the suspect said "I didn't go near that car" and the draft quotes him as saying "I wasn't near the vehicle," those are not the same statement. The specific words matter: in cross-examination, an attorney can use the exact phrasing to argue the officer's report was not accurate and therefore cannot be trusted on any other claim.

For statements that are hard to hear clearly in the audio, two practices help. First, note in the narrative that the audio quality was poor at a specific point. That is not an admission of failure. It is an accurate description of the evidentiary record. Second, use paraphrase language rather than direct quotes for anything you cannot verify verbatim: "the subject indicated" or "the subject stated words to the effect that" rather than a direct quotation. The narrative remains accurate. The report does not falsely imply verbatim accuracy you cannot guarantee.

What to Do When the Footage Contradicts the Draft

The footage will sometimes contradict the draft. This is not a failure of the verification pass. It is the verification pass working. The question is what to do when it happens.

The first and non-negotiable rule: the footage wins. When the draft says one thing and the footage shows another, the draft is wrong. Correct it. Do not soften the contradiction with qualifying language. Do not use phrasing that implies the draft may have been right and the footage may be ambiguous. If the draft says "both hands raised" and the footage shows one hand raised, the correction is "one hand raised." That is the factual record.

Document the correction. Your field notes or your workflow log should record what the draft said, what the footage showed, and what you changed the report to reflect. That documentation is important for two reasons. First, it demonstrates the verification pass was real: you found errors and fixed them. Second, it provides a record if the original AI draft is ever subpoenaed or reviewed. The AI's error is documented; your correction is documented; the chain from draft to sworn account is traceable.

In a use-of-force incident or any incident where the factual record may be contested, consider flagging significant contradictions between the AI draft and the footage to your supervisor before submitting the report. Not because the correction is problematic, but because your supervisor should know the AI's initial draft contained a significant error and that the footage tells a different story. That knowledge is relevant to the agency's oversight of AI-assisted reporting and to any administrative review of the incident.

When the Footage Itself Is Incomplete

BWC footage is not a perfect record of any incident. Cameras can be obstructed. They can deactivate at critical moments. They can fail. The angle may not capture what was directly in front of the officer. Audio may be degraded by distance, noise, or equipment failure. When the footage is incomplete, the verification pass must acknowledge the incompleteness.

A claim based on footage that does not clearly show the relevant action is not a verified claim. It is an inference or an observation. Distinguish between what the footage clearly shows, what the footage is consistent with, and what the footage does not address. A report that accurately describes what the footage shows, and accurately notes what it does not show, is a stronger document than one that fills the footage gaps with confident assertions. The confident assertion that the footage cannot support is where the defense will probe.

Use the CAD entry and your field notes as secondary sources for claims the footage does not cover. The CAD entry is a timestamped record of the dispatch information, the call type, the assigned units, and the notes entered by the telecommunicator. If the CAD entry says "caller reported a male with a weapon," that is a sourced claim you can report accurately: "Per the CAD entry, the call type was reported as a male with a weapon." The claim is specific, sourced, and verifiable. It does not rely on the AI to have correctly inferred the call type from audio patterns.

The Use-of-Force Section: Highest Stakes, Highest Risk

If there was a use of force in the incident, the use-of-force section of the report requires the most intensive verification. This is not a suggestion. It is the section that will receive the most scrutiny in any civil action, criminal defense, administrative review, or oversight investigation. And it is, as noted above, the section where AI-generated language is most likely to draw on boilerplate patterns from other reports rather than the specific facts of this one.

Go through the use-of-force section sentence by sentence with the footage open. For every claim about what the subject did, confirm it in the footage. For every claim about what you did in response, confirm it in the footage. For every claim about the sequence, confirm it in the footage. Pay particular attention to:

• The precise moment when the subject's behavior changed from the AI draft's description and what the footage actually shows at that moment.
• Descriptions of physical posture, body position, or movement that the AI may have inferred from audio without visual confirmation.
• Any claim about what you "observed" that you did not personally narrate on camera and that is not clearly visible in the footage.
• The sequence of force escalation and de-escalation: does the draft accurately reflect what the footage shows about the timing and order of events?

If the use-of-force description in the AI draft contains language you know is not grounded in the footage of this specific incident, remove it or correct it. This is not optional and it is not a matter of degree. A use-of-force narrative that overstates the threat, softens the force, or invents details the footage does not support is not just an inaccurate document. It is a document that can result in suppression of evidence, a successful civil rights claim, disciplinary action, and in extreme cases criminal exposure for the officer. The verification standard for the use-of-force section is zero tolerance for ungrounded claims.

Building the Verification Habit

Officers who run the footage-grounded verification pass consistently, on every AI-drafted report, develop a competence that is hard to overstate. They become fluent in the gap between what the AI produces and what the footage shows. They develop a feel for the gap-fill signatures. They build a documentation practice that protects them over years of work, not just on the next report.

The officers who skip the verification pass on "routine" calls are the ones who discover, at the worst possible moment, that the gap-fill they let slide on a routine call turned out to be a significant fact in a case that did not stay routine. There is no such thing as a low-stakes call in retrospect. A call that looks routine at 11 PM can become a significant case by morning, and the report filed at 11 PM is the one that will be tested in court.

An agency that has deployed AI-assisted report writing without building the verification pass into the workflow has saved time on drafting and spent it elsewhere without accounting for the verification burden. That is a governance gap. The CJIS (Criminal Justice Information Services) Security Policy obligations sit with the agency, not the vendor. Accuracy of the report is the agency's responsibility. An officer who builds the verification habit is not just protecting themselves. They are filling a gap that the technology creates and that the agency's supervision structure may not yet have closed.

Key Takeaways

• The footage-grounded verification pass means actively checking every factual claim in an AI draft against the BWC recording, the CAD entry, and field notes, not reading through the draft and thinking it sounds right.
• Gap-fills follow predictable patterns: specificity without audio support, use-of-force boilerplate that the model draws from training data rather than the footage, and scene descriptions without audio cues. Knowing these signatures makes gap-fills catchable before they reach the sworn account.
• When the footage contradicts the draft, the footage wins. The correction is not "both accounts are consistent." The correction is the accurate description of what the footage shows, documented with a note of what was changed and why.
• The use-of-force section carries the highest stakes and the highest gap-fill risk. It receives zero-tolerance verification: every claim about what the subject did, every claim about what the officer did, every claim about sequence, confirmed against the footage before submission.
• Quoted statements require verbatim verification against the audio. If the audio quality is poor, use paraphrase language rather than direct quotes. An accurate paraphrase is a stronger document than an inaccurate verbatim quote.
• Personal observation is a valid source for claims the footage does not capture, but it must be explicitly labeled as such in the narrative. The distinction between "I observed" and a gap-fill is that personal observation is something you actually observed.
• CJIS obligations and report accuracy obligations stay with the agency and the officer, not with the AI vendor. The verification pass is how the officer meets those obligations in an AI-assisted workflow.
• The verification habit built on routine calls protects the officer on cases that do not stay routine. There is no low-stakes call in retrospect, and the AI-drafted report filed on a quiet night is the one that will be tested under the harshest conditions if the call escalates later.