Context The question is really two questions, and the polling has started answering the wrong one. Do these cameras work? And who gets to search them? Almost every claim made in public answers the first while the fight is about the second — which is why a technology with a genuine, measurable record keeps losing ground, and why its defenders keep producing evidence that does not go to the charge. Start with the case for them, at its strongest. A plate reader finds a specific car fast. The Arizona randomized trial — 117 patrol routes in Mesa, three arms, thirty weeks — found readers beat the same officers checking plates by hand on stolen-car hits (26.7% against 13.3%), stolen plates, occupied recoveries and theft arrests, every one of them significant. Colorado police recovered six abducted children in five months on plate alerts. In Grand Blanc Township, Michigan, two carjacking suspects were in custody five minutes after a 3 a.m. alert. These are not press releases; they are the kind of result the technology was sold on, and they are real. Now the claim it cannot support. Neither trial found the cameras reduced crime — and Mesa found something sharper than nothing. The manual arm, the same officers checking the same plates by hand, DID cut auto theft; the camera arm did not. The authors' explanation is the whole argument in a sentence: a roving officer is unpredictable and a fixed camera is avoidable. Their section heading was "Efficiency Does Not Equal Effectiveness." The Virginia trial came back null on everything. The Justice Department's own evidence clearinghouse rates plate readers "ineffective" as a deterrent. Cleveland's clearance rate for robbery, assault and burglary rose from 15.5% to 18.7% after a hundred cameras went in — a real move — and the department told its own council it could not say how many of those cases the cameras had solved. (TS: the "70% of crime involves a vehicle" line the industry uses everywhere is the founder repeating what a police major once told him. There is no study behind it.) The best modern evidence for the other side is newer than the trials, and it deserves its numbers. Mourtgos and Adams looked at staggered Flock adoption across 216 agencies against seven years of federal crime data and found motor vehicle theft down 11.0% (confidence interval −17.3 to −4.2) and vehicle-theft clearances by arrest up 15.9%, with no effect on other property crime. Four things are true about it at once: it is the strongest study the cameras have; it is not peer reviewed; Flock supplied the deployment-timing data and promotes it as independent; and the clearance rate was already climbing three months before the cameras went live, which the authors themselves name as the main threat to their finding. Take it seriously and it moves the efficacy question from settled to open — for vehicle theft specifically, which is the one crime a plate reader is built for. So the honest summary is narrower than either side wants: they are an evidence tool, and they may be a shield against exactly one crime. They help you find the car after the crime. They do not stop the crime. Everything that follows is an argument about what that is worth, and what it costs. San Francisco is the case both sides cite, and the dates refute the version you have heard. Car break-ins fell from 30,701 in 2017 to 5,329 in 2025, a 22-year low. The collapse happened between August and November 2023: five months before the first Flock camera went in, four months before the surveillance ballot measure passed, sixteen months before the current mayor took office. What did coincide with autumn 2023 was bait cars and saturating the tourist hot spots with plainclothes officers on overtime. Meanwhile Oakland's 293 cameras logged 639 million plates and 1.1 million alerts in 2025 — a 0.17% hit rate — and the department kept stolen-vehicle alerts switched off because it lacks the staff to answer them. The cost is the search log, and this is where the record is worst. In May 2025 a Texas sheriff's deputy searched 83,345 cameras across 6,809 networks nationwide for a woman who had ended her own pregnancy, entering the reason as "had an abortion, search for female." Flock called the reporting false; court records later showed deputies had considered charging her. One Illinois police department's logs showed more than 4,000 nationwide lookups run for federal immigration agents over a year, as informal favors; Illinois found Flock in violation of state law, and California sued El Cajon for sharing with 27 states. A Georgia police chief ran his ex-girlfriend's plate about 500 times. An Indianapolis officer ran his wife's 3,759 times in ten months. EFF's audit of 12 million searches found hundreds tied to protests and 80 agencies logging searches with ethnic slurs; a Flock training webinar taught departments how to prepare for the "No Kings" marches. Aurora, Colorado paid $1.9 million after a misread plate — a stolen Montana motorcycle sharing digits with her Colorado SUV — put a woman and four children, aged 6 to 17, face-down on the pavement. And Aurora is also where, in December 2025, a reader flagged a stolen car with a 14-month-old inside as it entered the city, before the AMBER Alert went out, and the child was back in under ten minutes. Same city, same technology, both true, and any account of this that carries only one of them is selling something. The defenders' arithmetic on that is not stupid, and it deserves stating. Charles Fain Lehman of the Manhattan Institute divides roughly 125 abusing officers by the 25,000 with access and gets half a percent, and notes that Flock watches under half a percent of American road miles. If you think the cameras are a tool like any other, that is a supervision problem with a supervision answer: audit logs, written reasons, retention limits — the things Flock in fact shipped on Aug. 13. EFF's reply is that the guardrails were the deal the first time, and one officer's logged search reason in their audit was "hehehe" typed twenty times. The road-mile number is the one worth pressing, because Norfolk's own discovery logs answer it: the city's cameras captured one plaintiff's car between 475 and 526 times, and another's 849 times, in about four months. Roughly four sightings a day, for two people with no warrant against them and no charges. Half a percent of American road miles is not half a percent of any particular American's driving, because cameras are not sprinkled over the map — they sit where people actually drive. The rest of the stack repeats the pattern with worse evidence. Ring's doorbell network never actually retreated: Amazon told the Senate in November 2025 that 2,723 police departments were on its platform, about 600 more than in the year it supposedly pulled back, and its Community Requests pipe now clips straight into Axon's evidence system. No randomized trial of home cameras exists anywhere; an NBC survey of 40 partnered departments found 13 with zero arrests from Ring footage and none measuring crime impact. Gunshot detection and drone-as-first-responder programs are the two pieces of the stack this desk has not yet read the literature on, and rather than wave at them: that is the next piece. Britain has the longest record and makes the smallest claim, which is the tell. A 40-year meta-analysis puts CCTV's crime effect at about 37% in car parks, about 12% on residential streets, zero on violence, and zero across American studies. But a London study found robbery solve rates jump from 8.9% to 55.7% when usable footage exists. Cameras solve; they do not deter. On facial recognition the Met scanned about 3.1 million faces in a year for 2,077 alerts, 10 of them false, and 962 arrests — and the campaigners' famous "98% false positive" figure divides false alerts by alerts while the police divide them by faces seen. Both numbers are arithmetic on the same data, and citing either without the denominator is a way of not having the argument. Speed cameras are the exception, and it is worth being precise about why. A speed camera keeps no searchable retroactive record of where you have been; it observes one act, at one place, and issues one ticket. There is no log for a deputy in another state to query, which removes at a stroke every objection in the four paragraphs above — and leaves only the objection this desk actually has, which is about discretion. They are also where that instinct loses to the evidence. A Cochrane review puts fatal and serious crashes down 11% to 44% near cameras. A December 2025 PNAS study of New York's staggered rollout — 700,000 collisions, 18 million tickets — found collisions down 30% and injuries down 16% in the seven months after a camera went live. New York's design also answers the revenue charge: a flat $50, no points, no vendor cut, and 74% of ticketed vehicles get no more than two a year. The desk's view, stated as the desk's view: speed laws should be enforced by human beings with discretion — someone who can tell a driver weaving through every narrow gap from a driver doing 15 over on an empty stretch of straight highway at no risk to anyone. And the best answer to that came from the other side, in a thread on Aug. 30: under a camera the discretion does not vanish, it moves upstream, to whoever decides where the camera goes and whether it is signposted 300 meters out, as French law requires. New South Wales tested the opposite in 2020 by stripping the warning signs off its mobile cameras; fine revenue rose roughly tenfold, the road toll rose too, and the signs were back within 26 months. Chicago, which tickets at 6 mph over, is the case for the prosecution: a study the city itself commissioned found the fines fell hardest on Black and Latino neighborhoods even with race-blind camera placement. The technology is not the variable. The rules around it are. The strongest argument for enforcement at this level is worth meeting head on. Matt Yglesias made the case for fare enforcement in 2022 with a Bill Bratton anecdote and one unusually honest sentence: without low-level stops, "the only way to get illegal guns off the street is by stopping people at random, which realistically means racial profiling." Notice what that argument depends on — the stop being a pretext an officer chooses, which is exactly the thing a camera removes. When Nashville measured whether low-level traffic stops caught serious offenders or cut crime, with the NYU Policing Project and Stanford in 2018, the answer was no. By this month Yglesias had moved to firmer ground: fare enforcement makes the system "pleasant for normal people and safe for kids," which is a claim about order rather than about clearance rates — and the evidence for it is the kind that exists. None of this is left against right, which is the most interesting fact in the file. Bernie Sanders wants the cameras banned; so does Thomas Massie. Ron DeSantis and the ACLU landed in the same place in the same week. The real split is order against liberty, and it runs through both parties. Yglesias is "favorably inclined" and thinks a pro-camera Democrat plays well against type. The heterodox account @xwanyex puts the pro case at its bluntest, on a case that had video: "There's all these arguments about how we need to get it right, how we need to be sure, that seem oblivious to the existence of video surveillance. There's no ambiguity here. What could we possibly be waiting for?" And then he concedes the other side's point — a panopticon, he wrote in June, "makes the old ways newly risky." What the November ballots are actually deciding, then, is not whether the cameras work. It is whether a search of everywhere a car has been for the last week should require a warrant, a supervisor, or a written sentence an officer can type "hehehe" into. The courts have started answering it without waiting for the ballots: in June the Supreme Court held 6-3, in Chatrie, that cellphone location data is protected by the Fourth Amendment, which knocked out the precedent Norfolk's win rested on, and the Fourth Circuit appeal now lands in a different landscape. Florence, Oregon has already written the answer a city can write — cameras kept, access restricted to in-state agencies, a specific justification required for every search, retention capped. The places losing their cameras are, almost without exception, the places that never wrote one.