You can’t turn off the cameras — even the ones you bought.

Outside · the mechanism

A camera used to be a camera. Now it’s a searchable memory.

The change that matters isn’t the lens. It’s that what a camera sees now flows into a database that is retained, networked, and searchable months later.

Four kinds of sensor are spreading fastest. Automated license-plate readers — the dominant vendor, Flock Safety, runs more than a hundred thousand of them — photograph every passing car and log where and when it was seen. Face-recognition systems match people against databases. Acoustic sensors, originally sold to detect gunshots, are being pointed at human voices: in October 2025 Flock announced it would expand its microphones to listen for “human distress,” including screaming.¹ And camera-equipped drones are increasingly launched as first responders.

Any one reading is mundane. The problem is the aggregate. Oakland offers an unusually clear picture, because its police department publishes the numbers: in 2025 its 293 readers recorded more than 638 million license-plate scans, which generated about 1.1 million “hotlist” alerts — and only a fraction of those led anywhere.² That is the real mechanism: a running record of the movements of an entire city, almost all of it belonging to people no one suspects of anything.

Outside · security

The people collecting this can’t reliably keep it safe

A system that records everyone’s movements is only as trustworthy as its weakest camera — and the cameras have not held up well to scrutiny.

Across late 2024 and 2025, independent researchers catalogued a long list of security flaws in Flock’s devices — one white paper documented 51 findings, 22 of them assigned formal vulnerability identifiers.^5,6 A musician-turned-researcher, Benn Jordan, working with others, showed that with brief physical access to a roadside camera it could be fully compromised, and reported that some older units ran outdated software with unencrypted storage.⁶ Worse, dozens of cameras were found sitting on the open internet with no password at all — live feeds of parking lots, arguments, and children in parks, viewable by anyone who knew where to look, many of them the type of camera that tracks people rather than plates.⁷

Flock disputes the framing. The company says it has registered the reported vulnerabilities with the national CVE database, continues to publish findings, and maintains that none of them undermine its customers’ public-safety work.⁵ It initially argued the attacks required physical access and inside knowledge; researchers responded that publicly mounted and cloud-connected cameras showed the same weaknesses.⁶ Lawmakers have since asked the Federal Trade Commission to investigate, pointing to stolen customer credentials and the absence of mandatory multi-factor authentication.¹ The through-point is simple: this is a database of where people go, and it has repeatedly turned out to be less secure than the companies selling it claim.

Outside · misuse and error

A tool this powerful gets used for more than catching criminals

Two things follow from mass collection: the data gets searched for purposes far from the original pitch, and the systems make mistakes with real consequences.

On the misuse side, the Electronic Frontier Foundation obtained records of more than 12 million searches run by over 3,900 agencies in less than a year, and found searches tied to political protests, searches targeting Romani people, and searches related to women seeking reproductive healthcare.¹ Reporters separately documented Texas officers searching plate data to find a woman who had traveled for an abortion, and federal immigration authorities reaching one state’s data in apparent violation of its privacy law — enough to trigger a state audit.^1,4 These systems have also been misused by individual officers to track specific people, including a documented case we cover on our advocacy page.

On the error side, face recognition has contributed to more than a dozen known wrongful arrests, and the pattern is disturbing: nearly everyone misidentified has been Black, and the technology is measurably less accurate for darker skin.^9,10 One man was handcuffed in front of his children; a woman was interrogated while eight months pregnant; another man was jailed over a holiday weekend for a crime in a state he had never visited.⁹ A recurring flaw makes it worse: police often build a photo lineup around the algorithm’s pick, so a witness “confirms” the same face the machine chose.⁹ Departments reply that a match is only an investigative lead and that some, like Detroit, have tightened their rules — a fair point, except the record keeps showing the lead treated as proof.⁹

Outside · the evidence

Does it actually work? The honest answer is: unclear

The case for these systems deserves a fair hearing, and it has real points in its favor. It also has less evidence behind it than the marketing suggests.

Start with the strongest version of the argument. Flock’s own analysis estimates its network helps solve roughly 700,000 crimes a year — on the order of a tenth of reported crime nationwide.⁸ Police can point to concrete results, including cameras that helped identify a suspect in the killings of three women; and some civil-rights voices, such as an Oakland chapter of the NAACP, argue that an automated read is less biased than an officer’s discretionary decision about whom to stop.¹¹ None of that should be waved away.

But the independent evidence is thin and mixed. A review of the research concludes there is little empirical support for the claim that plate readers reduce crime.³ A study of a large plate-reader expansion in Atlantic City found no clear drop in violent crime overall — some associations with fewer shootings and thefts, but with the authors warning that other factors are hard to rule out.⁴ And the yield is lopsided: recall that Oakland logged 638 million reads to produce about 1.1 million alerts, only a sliver of which mattered, and even the police there describe the cameras as “one element,” not the cause of any decline.² The distinction that matters: “this camera helped solve a case” and “recording everyone, all the time, is justified” are different claims — and only the first is well-supported.

Inside · the turn

The same pattern now lives in your house

The devices you bought run the identical play: they record, they phone home, and you don’t control what happens next.

Take the robot vacuum — a useful worked example because so many now ship with a camera, microphones, and a permanent connection to the maker’s cloud. In 2024, security researchers at the DEF CON conference showed that popular Ecovacs models could be taken over through a weak Bluetooth and PIN design, letting an attacker switch on the camera and microphone from outside the home, with no indicator light — and even delete the spoken “camera on” warning to stay hidden.¹² An Australian broadcaster reproduced the attack live, quietly pulling photographs from a (consenting) owner’s device.¹³ The researchers said the company took about ten months to respond; it later attributed real-world incidents to recycled passwords and promised a fix.¹²

The lesson isn’t that vacuums are sinister. It’s that a camera and a microphone on your floor, wired to someone else’s servers, is a surveillance device whether or not anyone meant it to be — and its safety depends entirely on a company you have to trust.

Inside · two more

Your doorbell and your television are watching too

These aren’t exotic hacks. They’re the default behavior — or the default-insecure design — of ordinary consumer gear, and regulators have already acted on both.

The Federal Trade Commission charged Amazon’s Ring with letting employees and contractors browse customers’ private videos and with security so weak that hackers hijacked cameras; by the FTC’s account, one employee spent months viewing thousands of recordings from women’s bedrooms and bathrooms before a colleague noticed. Ring settled for $5.8 million and was ordered to delete algorithms built from improperly viewed footage and to require multi-factor authentication.¹⁴ Ring had also built a police-facing side — a feature letting departments request residents’ doorbell video — which it discontinued in 2024 after sustained criticism.¹⁶

Televisions do it more quietly. Vizio installed “automated content recognition” on roughly 11 million sets, switched on by default, tracking what appeared on screen second by second and selling those viewing profiles — without meaningful consent. It paid $2.2 million to the FTC and New Jersey and was ordered to obtain opt-in consent going forward.¹⁵ Vizio noted the data wasn’t tied to your name; it was tied to your IP address and your household. The common thread across the vacuum, the doorbell, and the TV is that the watching is built in, not bolted on.

The thread

You bought the device. You don’t control it.

Both halves of this page share one root, and it’s a legal one as much as a technical one.

The thing recording you — on a pole or on your shelf — runs someone else’s software, on someone else’s terms. And under Section 1201 of the Digital Millennium Copyright Act, bypassing the digital locks on a device can itself be unlawful, so it is often the manufacturer, not you, who decides whether the camera can be turned off, whether the data leaves, and when the product simply stops working.¹⁷ The Copyright Office grants narrow repair exemptions every few years, but the default runs the other way — which is why the right-to-repair fight and the surveillance fight keep turning out to be the same fight.

Whether it’s a city camera you never agreed to or a vacuum you paid for, the shape is identical: you are surrounded by devices that see you, and you have the least say over them of anyone involved.

Where we come in

What you can actually do

You can’t opt out of a camera on a pole. But you can shrink your own exposure, and you can push on the parts that are decided in public — which is the whole reason this club has two sides.