The Camera That's Wrong Seven Times Out of Ten (And We're Still Paying For It)

Saw a story out of California this week that made me put my coffee down. One town’s automatic licence plate readers, the Flock system, misread plates in 71% of the alerts sent to police. Not a rounding error. Not a bad week. Seven out of ten times, wrong.

I work in tech. I’ve spent twenty-odd years watching systems get shipped before they’re ready, watching “good enough for now” quietly become “good enough forever” because nobody wants to spend the budget fixing it once it’s live. So none of this surprised me. What got me was the gap between what the company claims and what the maths actually says. Flock’s line is 96% accuracy per character. Sounds great, until you remember a plate has six or seven characters, and the errors compound. Do the sums properly and you land somewhere near the failure rate this town reported. That’s not a rare glitch. That’s the expected outcome, printed in the fine print the whole time.

I don’t know Flock’s engineers, and I’m sure most of them are decent people trying to solve a real problem. But there’s a pattern here I recognise from close range: sell the dream to a customer who doesn’t have the technical grounding to interrogate the claim, let the marketing slide do the talking, and let some overworked local government sign the contract without asking the one question that matters, which is “wrong how often, and what happens when it is.” Councils here do the same thing with facial recognition trials and “smart city” pilots. Nobody in the room understands the false positive rate, and by the time anyone does, the contract’s signed and the cameras are bolted to the poles.

The bit that actually worries me isn’t the technology, it’s what happens downstream of it. A misread plate isn’t just an inconvenience, a wrong fine, a bit of admin. In the US, some of these alerts lead to felony stops: guns drawn, people pulled from cars, treated as dangerous suspects because a computer said so and nobody double-checked before officers rolled up hot. That’s not a software bug. That’s a person on their knees on the road because a camera got confused by a smudged plate at dusk. We don’t have Flock here, not that I’ve seen, but we’ve got speed cameras, toll gantries, and a state government that’s never met a surveillance camera it didn’t like the sound of. Melbourne’s CBD alone has enough number-plate recognition gear to make a decent stab at tracking anyone’s daily movements without much effort. I’m not against all of it. I like that stolen cars get found. I like that hit-and-run drivers get caught. But “useful sometimes” and “trustworthy enough to justify a gun pointed at someone’s head” are very different bars, and we keep letting vendors blur the line between them.

There’s a tension I can’t tidy away here. I’m genuinely fascinated by what this kind of pattern-matching technology can do well, when it’s built properly and checked by people who know what they’re looking at. I’m also fairly sure that most of the places rolling it out have no real appetite for the oversight required to make it safe. Both things are true at once. The industry’s incentive is to ship first and fix reputation later, because contracts get signed on promises, not on audited failure rates. That’s not unique to policing tech, it’s the same logic that gave us dodgy telehealth apps and self-driving demos staged for investors. But when the failure mode is a teenager getting dragged out of a car at gunpoint over a licence plate that had one wrong digit, “we’ll get it right eventually” isn’t good enough.

What I keep coming back to is that the fix here isn’t really technical. Better algorithms will shave the error rate down, sure, and maybe in five years it’s 98% instead of 71%. That still leaves real people getting stopped over nothing, and it still leaves the question nobody in these procurement meetings seems to ask: should this exist at all, regardless of how well it works? I don’t have a tidy answer. I like the idea of catching the bloke who’s stolen forty cars this year. I don’t like the idea of a system that treats every driver as a suspect worth scanning, twenty billion times a month, on the promise that it’s mostly fine. Somewhere in there is a genuinely useful tool. Somewhere in there is also a surveillance apparatus with a failure rate nobody would accept from a lift or a smoke alarm. We built it anyway, and we’re still arguing about which one we ended up with.