The Accountability Inversion

Every surveillance tool needs a story that makes people want it, and the body camera arrived with the best story of all: it was going to watch the watchers. After a decade of demands for police accountability, a small camera on an officer’s chest was offered as the answer, a neutral witness, a check on power, a thing the public had asked for. That story did its work. The cameras were bought, mandated, normalised. And now, with the cameras already on every chest and the public already sold on the principle, the part of the device that was always pointed outward, at us, has been handed to an artificial intelligence that reads faces. The accountability camera and the surveillance camera are the same camera. Only the direction of attention has changed, and the story that justified the first now quietly cloaks the second.

§01 · The sale: a camera to watch power

Trace how the camera got here, because the sequence is the point. The body-worn camera was adopted in Canada in the language of accountability and transparency: a way to document police interactions, to settle disputes about what happened, to rebuild trust after high-profile killings. Toronto became the first Canadian service to put cameras on its officers in August 2020, in the months after a global wave of protest against police violence. The RCMP began its national rollout only on 18 November 2024, with more than ten thousand cameras planned and a majority deployed within months.verified Alberta went further and mandated body cameras for every police agency in the province in 2023, describing it as a transparency measure to document interactions, collect better evidence, and speed the resolution of complaints.

Hold that justification, because it is doing more than it appears. Every one of those rationales points the camera at the police: document their conduct, settle disputes about their actions, resolve complaints against them. That is the sale. But a camera does not know which way it is pointed. The same device that records an officer’s conduct records everyone the officer encounters, the bystanders, the complainants, the faces on a sidewalk, and it captures them in far greater number. The accountability story justified buying a tool whose larger output was always a stream of footage of the public. The watchdog was, by construction, also a recorder of everyone the watchdog passed.

Counter: a camera that records the public as well as the officer is just how cameras work, and the accountability use is real. Both true. The question this case asks is what happens when someone notices that the device is already on every chest, the public stream is already flowing, and the only thing missing is a way to make the stream searchable by who is in it.

§02 · The lens turns around

That missing piece is artificial intelligence, and it inverts the device. A body camera with face recognition does not wait to be reviewed after an incident to settle a dispute about an officer. It watches, in the moment, everyone in front of it, and compares each face against a list. The direction of attention reverses: the camera sold to capture the conduct of the one person wearing it is repurposed to identify the many people standing before it. The watchdog becomes a bloodhound. Gideon Christian, an associate professor of artificial intelligence and law at the University of Calgary, named the reversal exactly: instead of a police body camera watching the police for the public, he said, “you now have that camera being retrofitted to watch the public on behalf of the police,” which he called “a very concerning reversal of the purpose of that tool.”

And the police force says so in its own filing. In the privacy assessment it submitted to Alberta’s regulator, the Edmonton Police Service calls automated facial recognition a “significant adaptation” of the body camera that creates a “more sensitive category of use of personal information,” and states it plainly: “The continuous scanning of faces for comparison against a watchlist constitutes proactive surveillance.” The device sold as a record of police conduct is described, in the force’s own words, as proactive surveillance of the public.

This is the laundering move stated cleanly. Take an instrument of outward surveillance, wrap it in a story about watching power, get it bought and normalised on that story, then activate the outward-facing capability the device always had. The reputational wash is “accountability”; the product that comes out the other side is a face-recognition network already mounted on thousands of officers, paid for and accepted before its real reach was switched on.

§03 · Edmonton crosses the line

This is not a forecast. On 2 December 2025, the Edmonton Police Service announced a pilot, switched on that month, of artificial-intelligence facial recognition running on officers’ body-worn cameras, the first police agency in Canada to try it. About fifty officers carried it on patrol. The system was trained to match faces against a watch list of roughly seven thousand people: by the acting superintendent’s account, 6,341 flagged with cautions such as “violent or assaultive,” “armed and dangerous,” “weapons,” “escape risk” and “high-risk offender,” plus a separate 724 with at least one serious criminal warrant.verified It ran only in daylight, only once an officer started recording, and in “silent mode”: the officers were not told whether a match was made, the results reviewed afterward at the station. A second phase, if approved, would push near-real-time match notifications to officers in the field.

Two facts about the supplier matter. The body camera is Axon’s, the dominant maker that also won the contract for the RCMP’s cameras; but Axon does not build the face-matching model, and publicly declined to name the third-party vendor it used. From Edmonton police emails, CBC News identified that vendor as Corsight AI, an Israeli company whose technology, the New York Times reported in 2024, was used in part in Israel’s mass surveillance of Palestinians in Gaza (CBC said it had not independently confirmed that reporting).verified The University of Calgary’s Gideon Christian put the obvious point: “the track record of who is providing the technology matters.” And Axon’s own framing names the larger purpose: its chief executive, Rick Smith, called the pilot “early-stage field research,” and wrote that testing “outside the U.S.” would let the company “apply those learnings to future evaluations, including within the United States.” A University of Alberta criminologist called it plainly: “Edmonton is a laboratory for this tool.”

Now watch the oversight, because it splits in a telling way. Alberta’s Information and Privacy Commissioner, Diane McLeod, received the privacy impact assessment on the same day the program was announced and was critical of the service for launching before her review; months later her office said it was still working through it. The service maintained it was not required to wait. The police commission by contrast, the body meant to hold the service to account, reviewed the proposal in September and raised no objections, its chair calling it “proof of concept testing only.” Against that clearance, Kate Robertson of the University of Toronto’s Citizen Lab, a former justice-system lawyer, called the pilot “likely the most high-risk algorithmic surveillance program” she had seen in Canada, and said flatly that “there is no legal authority that would justify” it, while the force’s own assessment argues only that the project “can be crafted and framed in a way that may be legally viable.” The safeguard ran after the act, the commission cleared it, and whether it is lawful at all is, on the record, contested.

A camera bought to watch the police is now reading seven thousand faces off the street. The review of whether it should arrived the day it already had.

Counter: it is a limited pilot against a defined high-risk list, not blanket surveillance of everyone. Granted, and the case marks it as a pilot. But a pilot is how a taboo becomes a tool: it establishes that the camera can be pointed this way, normalises the sight of it, and leaves the list’s size and contents to grow by the same administrative discretion that set it at seven thousand.

§04 · The line was drawn for a reason

The reason this reads as a line being crossed is that the line was drawn deliberately, and recently, by the industry’s own ethicists. In 2019, Axon’s artificial-intelligence ethics board concluded that putting face recognition on police body cameras could not be ethically justified given the state of the technology, citing its “unequal performance across races, ethnicities, genders and other identity groups,” and the company stepped back; amid the 2020 protests the major technology firms paused selling the technology to police entirely. The former chair of that board, Barry Friedman, now a law professor at New York University, told the Associated Press he is concerned Axon is moving ahead without enough public debate, testing or vetting. “It’s not a decision to be made simply by police agencies and certainly not by vendors,” he said. “A pilot is a great idea. But there’s supposed to be transparency, accountability. None of that’s here. They’re just going ahead. They found an agency willing to go ahead and they’re just going ahead.”verified (Friedman and seven others had resigned from the board in 2022 over a separate Axon plan for Taser-equipped drones.)

The reluctance is not Friedman’s alone, and the regulators drew the line too. Motorola Solutions, which lost the RCMP camera contract to Axon, says it can build facial recognition into body cameras but has “intentionally abstained from deploying this feature for proactive identification.” Canada’s privacy commissioners found the face-scraping company Clearview AI to be conducting mass surveillance that was unlawful, and found the RCMP’s use of it a breach of federal privacy law.legal The RCMP’s body-camera policy, today, specifically forbids facial recognition. So the Edmonton pilot does not arrive into an empty field. It steps over a boundary that the technology’s own ethicists, a rival vendor, the country’s privacy commissioners, and the national police force had all, for stated reasons, refused to cross. The accountability frame is what lets that step be taken as an experiment rather than read as the reversal it is.

§05 · The machine still needs no reason, now on faces

Set this beside Case 50 and the family resemblance is exact. There, a plate reader checks every passing plate against a hot list with no per-plate reason, lawful because reading a public plate is not a search. Here, a body camera checks every passing face against a watch list with no per-person suspicion. It is the same reason-less machine, moved from the licence plate to the human face, and the face carries far more than a plate does. A plate is registered to a vehicle. A face is the index to a life, and it cannot be swapped, covered, or left at home.

And the face version brings a harm the plate version does not. Facial recognition’s accuracy is not even across people: study after study has measured higher error rates by race, gender and age, and Axon itself concedes that factors like distance, lighting and angle “can disproportionately impact accuracy for darker-skinned individuals.” So the false matches, the wrong person stopped because a machine decided their face was on the list, fall hardest on the people already most policed.verified A reason-less scan that misfires unevenly is not a neutral tool with an occasional glitch. It is a sorting mechanism whose mistakes have a direction, and it is being switched on in a city where, as a University of Alberta criminologist observed, the police service has had a sometimes “frosty” relationship with its Indigenous and Black residents. The accountability camera, turned outward and given a flawed eye, does not distribute its errors at random. It concentrates them.

Counter: facial recognition is improving, and a human officer confirms any match before acting. Improving is not solved, and a human asked to confirm a machine’s confident output is the weakest kind of check, the one Case 49 and Case 50 already found rubber-stamping what the system proposed. The error rate is a measured fact; the safeguard is a hope.

§06 · The strongest case for the other side

The case must survive its best reply. Body cameras have done real accountability work: footage has contradicted false police accounts, supported complaints that would otherwise have been one person’s word against an officer’s, and given families and courts a record where before there was none. Most Canadian services do not run facial recognition on them; Toronto does not, and the RCMP forbids it. Edmonton frames its program as research, with a defined list and a privacy assessment on file. Facial recognition genuinely can help find a missing child or identify a dangerous suspect. None of this is nothing, and the case concedes all of it.

The narrower claim survives untouched: that the body camera was sold and mandated on a story about watching power, and that this story is precisely what now makes it acceptable to point the device the other way, at the public, with an AI eye that needs no suspicion and errs by race. A reader who values body cameras for accountability should be the first to object, because the inversion spends the trust the accountability use earned. The tool that was supposed to watch the watchers is being made to watch everyone else, and the goodwill of the first job is the budget for the second.

Counter: “a valuable accountability tool” and “a public-facing face-surveillance network” can both be true of the same camera. They can, which is exactly why the direction the lens points should be decided in public, on the record, and not switched by a vendor pilot while the oversight body reads the paperwork.

§07 · The inversion, named

Strip it to the structure. A device with two possible directions, inward at the officer and outward at the public, was bought and mandated on the strength of the inward use, the one the public demanded. That use was real, and it built trust and normalised the hardware. Then artificial intelligence activated the outward use that the device always physically had, turning a record of police conduct into a search of the public’s faces, and the accountability story stayed in place as the name on the thing, now describing its opposite. The line the industry’s own ethicists and the country’s privacy commissioners had drawn was stepped over as a pilot, and the safeguard arrived the day the scanning began.

So the honest sentence is the cold one. The surest way to point a camera at the public is to first sell it as a camera pointed at power, because the trust earned watching the watchers is exactly the trust you spend when you turn the lens around. The body camera was the public’s demand answered, and the answer is being repaid by aiming the public’s own instrument back at them. The question this case leaves is the one the accountability frame is built to skip: not “does this footage hold the officer to account,” asked of one recording, but “who is this camera for now,” asked before the watch list grows from seven thousand.