Investigation
Bend Privacy Alliance / Axon patents series, Part 10 of 11 · September 2026
Axon’s redaction tools and patents show how police video can be transformed by AI detection, transcripts, tracking and policy choices—and why the transformation itself needs an audit trail.
Ask for police body-camera footage and the video that comes back may contain blurred faces, hidden license plates, obscured screens and stretches of muted audio. It is easy to think of that process as digital black ink: someone identifies private information, covers it up and releases what remains.
That description is increasingly incomplete.
Axon now sells AI-assisted redaction tools that can detect heads, plates, screens, identity documents and personally identifiable information in audio. Its software can track objects through video, use transcripts to locate portions of speech, generate masks and produce a new redacted evidence file while leaving the original unchanged.
Axon also holds patents describing more ambitious systems. One addresses redaction rules tied to the intended use of evidence; another describes using redaction criteria across related recordings. Those patent capabilities should not be confused with current commercial deployment. A patent shows that Axon designed and sought protection for a system. It does not prove that Axon currently sells every described capability, that a particular customer can use it, or that Bend Police has activated it.
Still, the patents and current products together reveal something important about what redaction has become. A redacted police video may be the end product of a longer process involving machine detection, timestamps, object tracks, configuration choices, legal reasons and human corrections.
The question is no longer simply what was hidden. It is whether anyone can reconstruct how and why it disappeared.
The file you receive may not be the file that was recorded
Axon’s current Evidence documentation provides a reassuring starting point. When media is extracted from Redaction Studio, Axon says the system creates a new evidence file while leaving the original unchanged. The extracted item remains linked to its parent and has its own audit trail.
In other words:
original recording → redaction process → separate derivative
That distinction matters for chain of custody because an authoritative source can remain intact even when a heavily transformed version is produced for public release or another purpose.
Preserving the original, however, answers only part of the accountability question. A reviewer may also need to know who or what identified the sensitive material, whether a mask originated with software or a person, what configuration was active, which exemption was assigned, and whether anyone later changed the result. The transformation itself can become an important record.
Redaction is no longer just drawing a blur box
Axon’s current Redaction Assistant can inspect video for supported object classes including heads, license plates, screens, mobile-device screens, notebooks or paper items, and identification documents. It can also use transcripts to identify personally identifiable information in audio.
For supported categories, users can choose between Detect only and Auto-redact. Axon’s March 2026 release notes also document transcript-based Audio PII redaction, mute/bleep options and adjustable timing margins before and after detected speech.
Those settings affect the file that another person eventually sees or hears. Suppose someone states a home address while a body camera is recording. The system can use the transcript to identify the address and locate the corresponding portion of audio. If the transcript timing is slightly early or late, the agency can apply additional time before and after the detected words.
A larger margin may better protect the speaker, but it can also remove more surrounding conversation. Neither outcome necessarily reflects misuse. Redaction serves important purposes: victims, children, witnesses, medical information, addresses and unrelated private material often need protection before government video is released. Axon’s Redaction Studio quick-start guide instructs users to review processed video to confirm that sensitive information has been properly masked.
The important point is that the released artifact can reflect both the original event and a later set of technical and policy choices about what should remain perceptible.
One decision can follow an object through time
Axon’s patents provide a more detailed look at what automated redaction can do beneath the interface.
One Axon patent, US11122237B2, describes detecting light-emitting screens in recorded video, creating masks over the relevant pixels and assigning the detected screen a persistent track ID. That track lets one redaction instruction apply across later frames rather than being reset at each one.
The patent also describes associating later detections with a previous track ID, including after the object disappears and returns. That is useful because an officer’s laptop, mobile device or other screen may move constantly within body-camera footage.
It also changes the provenance question. The relevant history is no longer only that a mask appeared at a particular moment; it may include which detection created the track, which frames were associated with it, what redaction instruction was attached, and whether a human later modified the result.
Axon’s current commercial Redaction Assistant also advertises object tracking. Its product materials say unique heads can be followed throughout a video, including when they leave and later re-enter the frame. That should not be confused with facial recognition. Tracking a head through one recording is not the same as identifying a person by name or matching a face against an identity database.
What matters here is simpler: persistent object tracking allows one redaction decision to affect many later frames.
A transcript can become a control layer over the recording
Another Axon patent family makes the transformation even more explicit.
US10943600B2 describes systems that align transcript words with precise locations in audiovisual evidence. Once the transcript contains timestamps, it can do more than help someone read along. It can tell the system which part of the recording to alter.
The relationship is straightforward:
word → timestamp → portion of evidence → transformation
The patent describes identifying words or transcript portions, using their timestamps to locate the associated audiovisual material and removing or redacting the corresponding segment. Some claims contemplate that happening without further human intervention after the triggering words have been detected.
Current Axon Audio PII functionality follows the same broad logic. The system uses the transcript to identify categories of personally identifiable information and associates those detections with locations in the audio so that a reviewer can apply or confirm redaction.
That creates several distinct ways the process can fail. The transcript can mishear a word, or the text can be correct while the timestamp is slightly wrong. The PII classifier can mistake ordinary speech for sensitive information, or the classification can be right while the configured mask removes too much or too little surrounding audio.
Those are different kinds of errors. Evaluating automated audio redaction therefore requires looking not only at whether the final mute “worked,” but at transcription accuracy, timing accuracy, classification accuracy and the resulting redaction window.
Axon’s patents describe something broader: redaction by meaning and purpose
Object detection and transcript redaction still begin with relatively concrete things: a screen, a plate, a head, a spoken address.
Another Axon patent describes a more abstract architecture.
US12380235B2, granted to Axon on August 5, 2025, describes analyzing captured video to identify visual and semantic properties, generating alignment information that connects those findings back to the source, and applying rules to determine whether the identified material should be altered.
The claims state that those rules can be applied according to the intended use of the captured video.
The architecture therefore does not have to stop at identifying an object. It can also assign meaning to information and use rules to decide whether that information should remain visible in a particular version of the evidence.
The specification discusses categories including names, race, gender, age and other descriptive or identifying information. That does not establish that Axon’s current Redaction Assistant commercially detects all of those categories, and the present commercial documentation reviewed for this article describes a narrower set of visual objects and Audio PII classes.
That distinction is essential:
Axon patented it is not the same as Axon sells it.
Neither establishes that Bend has access to it, and none of those propositions establishes that Bend used it.
The patent is still significant because it shows an architecture in which meaning can become machine-identifiable data, aligned back to the source and subjected to redaction rules.
“The redacted video” may not be one thing
If redaction can depend on purpose, another assumption begins to break down.
We often speak of “the redacted video” as though one correct transformed version exists. Different recipients, however, can legitimately have different access rights. A public-records release may hide information that prosecutors or defense counsel can receive. A court order may require a specific field to be obscured, while a training or media version may require additional privacy protection.
Axon’s patent architecture contemplates different policies generating different redaction criteria from the same original recording. That flexibility can be useful and legally necessary, but it means “the redacted video” is an incomplete description.
The same original evidence can produce more than one redacted derivative. The accountability question is which version you received—and what rules produced it.
One camera’s redaction can become another camera’s instruction
Axon’s cross-redaction patents extend that idea beyond a single recording.
US20190348076A1 describes generating redaction criteria that identify locations in evidence and the redaction tasks to perform there. Those criteria can be used with alignment information to locate corresponding portions of related recordings and support redaction across an incident evidence set.
The patent also contemplates preserving the original recordings and redaction criteria while generating redacted derivatives when needed. That creates a useful distinction between the source evidence, the instructions describing what should be redacted, and the rendered file eventually released.
The redaction criteria may therefore be as important for accountability as the MP4 itself because they explain what transformation was supposed to occur. Cross-recording propagation also creates an obvious risk: if the originating redaction is wrong because of an incorrect target, time interval, alignment or policy, the error can potentially affect related recordings.
The architecture appears clearly in the patent. Axon’s current public product materials, however, document batch redaction across multiple files rather than this specific time-aligned propagation workflow. Whether the patented cross-recording mechanism is commercially available today remains unresolved from the materials reviewed here.
Axon preserves part of the transformation history
It would be inaccurate to suggest that Axon produces redacted evidence without any record of what happened.
Current Axon documentation describes a downloadable Redaction Activity Report associated with the original evidence. The report can identify the type of redaction object and its start and end time, and administrators can enable customized “Redacted Object and Reason” values.
Axon Evidence audit trails are described as immutable. Those are meaningful safeguards because they allow an agency to preserve more than simply the untouched source and a finished blurred copy.
The public documentation reviewed for this article does not, however, establish that the exported Activity Report contains the complete computational history behind every transformation. The published field descriptions do not tell us whether each redaction can expose the detector or model version, confidence, a persistent track ID, exact spatial mask coordinates, the transcript token that triggered an audio mask, the timing-margin setting used, or whether a mask was generated by software, created manually, or later modified by a person.
Some of that information may exist elsewhere inside Axon, in logs or fields not covered by the public documentation. But the public materials reviewed for this article do not answer the question either way, and that gap matters.
There is a difference between knowing what was ultimately hidden and being able to reconstruct why the system hid it.
Privacy can fail in both directions
Redaction is unusual because the civil-liberties problem runs both ways.
Under-redaction can expose a victim, witness, child, medical detail, home address or uninvolved person. Over-redaction can hide information that the public, press, defense or another legally entitled recipient should have received.
Automation can improve both problems or make either one worse. A better object detector may prevent a face from being accidentally disclosed, while a false positive may obscure someone who had no legal basis for being hidden. A longer audio margin may protect a spoken address, but it may also erase part of the sentence that explains what happened next.
That is why automated redaction should be evaluated as more than a privacy feature. It is also a mechanism for controlling access to information.
The broader semantic architecture raises another issue. If a redaction system identifies sensitive descriptive information in order to protect it, the system first has to derive or classify that information. Is the classification retained, and if so, who can see it, reuse it or search it — and how accurate is it to begin with?
A tool designed to protect sensitive information can itself create sensitive derived data.
This matters beyond public-records requests
The issue is easiest to see in a public-records release because the redactions are visible, but transformed evidence can matter elsewhere.
A defense attorney may need to know whether the file received is an untouched original or a derivative, what was removed and whether the unredacted source still exists. A prosecutor may need confidence that a derivative accurately reflects the underlying evidence. A court may need to determine whether missing audio or visual content was absent from the original recording or introduced later.
Different bodies of law determine who is entitled to which records, and nothing here assumes that every internal detector output must be disclosed to every recipient. The more basic accountability principle is that the agency itself should be able to reconstruct the process.
If software proposed the transformation, that should be knowable. If a human changed it, that should be knowable. If a particular policy or configuration produced the result, that should be knowable. If several derivatives were created from the same original, their lineage should be knowable.
Otherwise, an authoritative source can remain perfectly preserved while the process that determined what everyone else was allowed to see becomes difficult to audit.
Bend’s unanswered question is configuration
Bend already operates within a broad Axon environment.
Bend Privacy Alliance’s earlier review of City records identified Axon Evidence, Respond Plus, Standards, Capture and other Axon systems in the City’s production software environment. The same local record shows how Bend’s Axon architecture accumulated across multiple procurements rather than through one comprehensive decision.
That does not establish that Bend has licensed or enabled every Redaction Assistant feature discussed here. The patent deep dive reached the same conclusion: local licensing, configuration and actual redaction provenance remain deployment questions rather than patent questions.
The next local inquiry is therefore straightforward. Does Bend license Redaction Assistant? Is Audio PII available? Which detectors are enabled? Are they configured for Detect only or Auto-redact? What audio margins are used? What exemption or reason fields are required? Does Bend retain Redaction Activity Reports? Can it distinguish AI-generated masks from manual masks? Are track IDs, confidence values or model information available? Can any redaction criteria operate across synchronized recordings?
Those questions do not require speculation about what Axon might someday build. They ask what the City actually has, and they are answerable through records.
If evidence is transformed, the transformation should leave evidence behind
Redaction performs an essential function. Government should not expose a child’s face merely because a camera recorded it. A victim’s address should not become public because it was spoken within earshot of a body camera. A police computer displaying unrelated private information should not be released simply because it appeared in the background.
Automation may make those protections faster and more reliable. The accountability problem begins when the only records treated as important are the untouched original and the final redacted video.
Axon’s own architecture shows why that is too simple. Between those two files can sit detections, transcripts, timestamps, object tracks, masks, configuration settings, legal reasons and human corrections.
Those intermediate decisions determine what disappears.
Axon already documents portions of that chain through preserved originals and separate derivatives, linked evidence, activity reports and audit trails. The remaining question is whether enough of the transformation survives for someone to meaningfully examine it later.
A redacted police video is supposed to make some information invisible.
The process that made it invisible should not disappear with it.