> daily_signal(2026_07_14)

The plate cameras we have watched cities rip out all summer just got audited by LAPD's own watchdog, and the number is 161 innocent drivers.

PickBits Daily Signal · Tuesday, July 14, 2026

By Mark Pickering · 10 min read · July 14, 2026

// tl;dr

LAPD's inspector general counted 161 innocent drivers pulled over on false "stolen vehicle" hits in a single two-month window, and not one of them was a misread. The cameras matched the plates correctly; the shared hot list was stale, because other jurisdictions had recovered the vehicles and never removed them. Against 337 alerts that did recover a genuinely stolen car, that is roughly one bad stop in every three an officer acted on. LAPD has disconnected its 138 Flock cameras and the Police Commission takes the audit up today.
Twenty-six current and former Meta workers sued on Monday, alleging Meta assembled its layoff list with internal AI fed by keystroke-and-screen monitoring. The mechanism they allege is precise: the models scored people on output, and every one of the 26 had taken, requested or been approved for legally protected leave. A person who is lawfully absent produces less output. Meta says decisions "were and are made by people, not AI." A class-action waiver forces all 26 into individual arbitration.
A Los Angeles Uber driver filed a class action alleging Uber profiles drivers with their own location and biometric data to compute the lowest fare each will personally accept. Edwin Carranza calls it a "surveillance wage." He is not relitigating employee-versus-contractor status; he is suing on privacy and consent, and asking for an order forcing Uber to disclose what data sets driver-facing pay.
The FDA cleared iPredict-DR, an AI that screens for diabetic eye disease from a retinal photo a nurse can take, with no eye specialist in the loop. Diabetic retinopathy is silent while it is treatable, and the annual specialist exam is the appointment most people with diabetes never make. The clearance moves the screen into the primary-care visit they were already attending. It is on sale across the US now.

We have spent six weeks watching cities pull Flock's license-plate cameras out of the ground. On June 4, we counted thirty of them, all reacting to the same thing: federal agencies were quietly searching the data. Those were misuse stories. Somebody ran a search they had no business running. Today's is not a misuse story, and I think that makes it worse. LAPD's inspector general audited the cameras doing exactly what they are supposed to do, and counted 161 innocent drivers pulled over in two months on plate reads that were every one of them correct. Nobody hacked anything. A shared list went stale and nobody was checking it. Then look at stories two and three. Meta's models scored people on output, and 26 plaintiffs say that is how everyone who took protected leave ended up on the layoff list. An Uber driver says the app worked out what he was desperate enough to accept, and paid him that. Nothing in there is broken software either. Story four is the good one, and there is usually one. Today it is a nurse with a camera.

Today: LAPD's watchdog counted 161 innocent drivers stopped on correct plate reads against a stale list, 26 Meta workers sued to force the company to prove an AI did not build its layoff list, an Uber driver alleged the app computes the least he will personally accept, and the FDA cleared an AI that screens for diabetic blindness in a primary-care office.

1. LAPD's own watchdog counted the cost of its license-plate cameras: 161 innocent drivers, pulled over in two months, on plate reads that were all correct.

The machine was right. The list it checked was three years out of date.

The LAPD Office of the Inspector General audited the department's automated license-plate reader program over August 1 to September 30, 2025, and found 161 alerts where officers confirmed the camera had matched the plate correctly and the vehicle still was not stolen. That distinction is the whole story. The optical character recognition worked. The system failed because the shared "hot list" it checks against was stale. Other jurisdictions had recovered vehicles and had not removed them in time, so the machine confidently matched a real plate to a record that should no longer have existed. In the same window, 337 alerts did recover a genuinely stolen vehicle and ALPR data contributed to 74 arrests, which means that of the 498 stolen-vehicle alerts officers acted on, roughly one in three sent them after somebody who had done nothing. The cameras read 210.5 million plates in those two months.

Inspector General Matthew Barragan recommended the department stop installing new cameras and sign no new ALPR contracts pending public input, require annual audits of who accesses the data, and state plainly that misusing it is a disciplinary offense. And look at what he was writing into. LAPD had not conducted a thorough ALPR audit since 2022, and when a California Public Records Act request asked for audit records in 2025, the department answered that it had none. It has since declined to renew its Flock Safety agreement and disconnected the company's 138 pole cameras; the Axon contract expires at the end of July; and the Board of Police Commissioners takes the findings up today. Flock's counter is that its cameras capture point-in-time images of vehicles in public view rather than continuously tracking a person, which is a fair answer to a different question. Nothing in this audit turns on following anyone around. It turns on a database nobody was maintaining, and on the fact that the department could not produce its own error rate for three years.

404 Media report July 2026 LAPD Office of the Inspector General audit found 161 alerts where officers confirmed the license plate reader matched correctly but the vehicle was not stolen during August to September 2025 stale hot list from other jurisdictions 337 alerts recovered genuinely stolen vehicles 210.5 million plate reads LAPD disconnected 138 Flock Safety cameras Police Commission takes up findings — 404media.co · July 13, 2026

Why this matters: If you drive, stop worrying about a camera misreading your plate. Your actual risk is a database three states away that nobody bothered to update, and it reaches you as a felony stop with the guns out. And this came from the department's own watchdog, not a privacy group, which is why it is hard to wave away. Action this week: Find out whether your town runs plate readers and who it shares them with, which is a five-minute lookup almost nobody does. The EFF's Atlas of Surveillance (atlasofsurveillance.org) maps which US agencies have deployed ALPR by city, and the crowd-mapped DeFlock project (deflock.me) plots individual cameras and vendors. The sharing question matters more than the camera count, because LA's false hits came out of other jurisdictions' stale records, so your exposure is set by the worst data hygiene of any agency your department shares a list with. If you were ever pulled over on a plate-reader hit, an alert record exists: request it, and ask specifically which hot list the alert came from and when that entry was last validated. And if you operate any system that syncs a flag list across organizations, this audit is about you. Measure how long it takes a de-flag at the source to actually reach the thing making the decision, and alarm on the worst case rather than the average. Then give every entry an expiry the enforcement point will refuse to act on. Ask whether you could produce your own version of 161 out of 498 if somebody demanded it this afternoon. LAPD could not, for three years.

404media.co: LAPD regularly pulled over innocent people because license plate readers flagged their cars as stolen (July 13, 2026)
nbclosangeles.com: LAPD watchdog urges reconsideration of automatic license plate readers (July 2026)
futurism.com: LAPD abandons Flock contract after false alarms (July 2026)
therecord.media: License plate cameras may be next target after Supreme Court reins in location tracking (June 2026)

2. Twenty-six Meta workers are suing to force the company to prove an AI did not build its layoff list.

Score people on output, and somebody who was legally on leave looks exactly like somebody who was slacking.

On May 19, the day before Meta cut roughly 8,000 people, we ran the argument plainly: if Meta said an AI made the call, no federal law would require it to prove otherwise. That was a story about a legal vacuum, with no litigation anywhere on the board. Twenty-six of those workers are now attempting exactly the proof we said no law compelled. They filed on Monday, July 13 in the Northern District of California, before Judge William Orrick, in a 71-page complaint alleging Meta "used a constellation of internal artificial-intelligence systems" to score, rank and select employees for the termination list: a system called Metamate, employee-trained "second-brain" agents, keystroke- and activity-monitoring data, AI-token-usage dashboards, and algorithmically assisted performance ranking.

The allegation is sharper than "a robot fired me," and the mechanism is testable. The models scored on performance, productivity and output metrics. An employee on legally protected leave produces fewer metrics because they were lawfully absent, so the model, doing exactly what it was built to do, ranked them below peers who were at their desks. Every one of the 26 plaintiffs shares one characteristic: all took, requested or were approved for protected leave in the past 24 months, and their claims run under the FMLA, the Pregnancy Discrimination Act, the ADA and state leave laws. The monitoring layer feeding those scores is its own story, and one we have been on since April, when we covered Meta's Model Capability Initiative recording employee keystrokes and screens to train agents; the complaint now alleges that same data was used to build the AI tools. Meta's denial is a single sentence and it is the sentence the case turns on: "Workforce management and organizational decisions were and are made by people, not AI." That may hold up. These are allegations, nothing is proven, and discovery is where it gets settled, on one question: whether a human who signs off on a machine-ranked list is a decision-maker or a signature. Note the third input in that list, though, before you decide it is exotic. If a dashboard measuring how much you use the company's AI tools feeds your ranking, then AI adoption has quietly become a performance metric, and nobody on parental leave is burning tokens.

Courthouse News Service report July 13 2026 twenty-six current and former Meta employees sued in Northern District of California before Judge William Orrick 71-page complaint alleging Meta used internal AI systems Metamate second-brain agents keystroke and activity monitoring AI token usage dashboards to score rank and select employees for layoff list plaintiffs all took protected leave FMLA ADA Pregnancy Discrimination Act arbitration class action waiver — courthousenews.com · July 13, 2026

Why this matters: If you have ever taken medical, family or parental leave, the defect alleged here is one you cannot see and did not consent to, and it is almost certainly in more systems than Meta's. Any model that scores on output volume (commits, tickets, tokens, hours active) will rank a person who was on twelve weeks of leave below an identical peer who was not, because the metric cannot tell "produced less" apart from "was lawfully not there." Your fairness dashboard will not catch this, because leave status is probably not even a feature in the model. Nobody is checking the denominator. Action this week: Dig out your arbitration agreement and read it, because it is the most consequential document in this story and most people have never opened theirs. These 26 plaintiffs have counsel and a real theory and they still cannot bring a class action, since Meta requires a mutual arbitration agreement with a class-action waiver that forces them into individual arbitration one at a time. If you were on protected leave and then selected in a reduction in force, read the EEOC's guidance on AI and the ADA (eeoc.gov), which establishes that algorithmic tools can violate the ADA when they screen out people with disabilities, including in selection for termination. And if you build or approve performance-ranking tooling, fix this deliberately: score per active period and exclude protected-leave windows from both the numerator and the denominator, then run the disparate-impact test on leave-takers as a cohort. If nobody in your organization can produce an example of a manager moving someone off the model's list and why, you do not have human review. If you took protected leave and were then cut, send this to whoever is helping you read your severance.

courthousenews.com: Meta employees sue over use of AI in workforce reduction (July 13, 2026)
eeoc.gov: Artificial intelligence and the ADA
npr.org: Meta layoffs and AI jobs (May 20, 2026)
nbcnews.com: Meta layoffs, AI (2026)

3. An Uber driver says the app worked out the least he would personally accept, and then paid him that.

Personalized pricing, run backwards, against the people doing the work.

In June, we ran the story of Amazon's software deciding when a delivery driver was allowed to turn on the air conditioning. That was an algorithm setting working conditions. This one sets pay. Edwin Carranza, a Los Angeles Uber driver, filed a putative class action in San Francisco Superior Court on Wednesday, July 8, alleging Uber unlawfully collects drivers' personal data and "uses it to algorithmically set fares based on predictions of how little drivers will accept." The claim is not that Uber pays badly. It is that Uber pays each driver a different, individually computed amount for the same trip, derived not from the ride's value or the driver's performance but from a machine-learned prediction of that specific person's reservation price. Carranza says the data he was required to hand over to work at all (precise location, trip history, acceptance and rejection patterns, cancellation history, driving times, app-use behavior, identity verification and biometric facial-verification data) was fed back into a profile that then set his pay. He calls the result a "surveillance wage."

The alleged behaviors are specific enough to test in discovery, which is what makes this more than a complaint about vibes: offers that got worse after he declined low-value rides, lower fares on trips heading in the direction of his home at the hour he usually stopped for the day, and surge notifications he says were frequently false, pulling him away from where he wanted to be. His line for it is the good one. Daily use of the app is a slot machine whose algorithms "distribute higher-value fares and lucrative bonuses unpredictably," a "sporadic reward system" that "preys on hope." The legal architecture is the clever part, and it is why this one is worth watching. Carranza is not relitigating employee-versus-contractor classification, the swamp that has consumed gig-work litigation for a decade. He is suing on privacy and consumer-protection grounds, and the core of it is consent: he says he was never told, and never agreed, that his data would be used to build a model of his own price sensitivity. That routes around the classification fight entirely and it is a far easier case to win. Uber did not respond to a request for comment, and its long-standing position is that upfront fares are calculated from base fare, estimated trip length and duration, pickup distance and surge. Discovery is what will establish whether anything else is in there. What Carranza asks for is the remedy that actually matters: an order forcing Uber to disclose what data and algorithmic factors set driver-facing trip amounts.

Courthouse News Service report July 9 2026 Los Angeles Uber driver Edwin Carranza filed putative class action in San Francisco Superior Court alleging Uber uses drivers biometric and location data to algorithmically predict the lowest fare each driver will accept surveillance wage upfront fares suing on privacy and consumer protection grounds not employee classification seeking order requiring Uber to disclose algorithmic factors setting driver pay — courthousenews.com · July 9, 2026

Why this matters: Companies have wanted to charge each of us a different price for years. This complaint says Uber got there first, and ran it backwards, on the people who work for it. The FTC's own surveillance-pricing study found that everything from your precise location to your mouse movements to what you abandoned in a cart is already used by intermediaries to set individualised consumer prices; Uber's drivers are simply the population where the practice is most measurable and now most litigated. Action this week: If you drive for a platform, request your own data, because it is the only way to see what the model sees and almost no driver does it. Uber runs a privacy center (myprivacy.uber.com, reachable in the app under Privacy), and California drivers have an additional statutory right under the CCPA and CPRA to request the specific pieces of personal information collected, the sources, the purpose and the third parties it was shared with, explicitly including information used for automated decision-making and profiling. Pull it and keep it, because the pattern Carranza describes is only provable against your own logged history. Keep an independent record of offered versus completed fares too, since the platform's number is the thing in dispute. And if you work in pricing, growth or ML, ask your own organization the question this complaint is built on: does any model in our stack take an individual's behavioral or biometric data as an input to the price or pay we offer that specific person, rather than a segment or a market? If yes, find out where exactly the user consented to that use, and whether there is a feature in there that is a proxy for desperation. Someone chose to put it in, and discovery will find the commit. If you know somebody who drives for a living, send this to them.

courthousenews.com: Uber drivers claim personal data used to manipulate fares (July 9, 2026)
ftc.gov: Surveillance pricing study indicates wide range of personal data used to set individualized consumer prices (January 2025)
columbialawreview.org: Veena Dubal, On Algorithmic Wage Discrimination
hrw.org: The Gig Trap, algorithmic wage and labor exploitation in platform work in the US (May 2025)

4. The FDA cleared an AI that catches diabetic blindness from a photo a nurse can take, with no eye specialist in the loop.

It is not trying to beat an ophthalmologist. It is trying to exist in a building where there is not one.

Four days ago, we covered a seventeen-year-old whose AI flags autism and ADHD from a single retinal photo; three weeks before that, a triage tool that cut the wait from a suspicious mammogram to biopsy from over two months to under ten days. Retinal photos keep showing up in this newsletter, because a camera pointed at the back of your eye is cheap and a specialist is not. On July 9 the FDA granted 510(k) clearance to iHealthScreen, a small medical-device company in Richmond Hill, New York, for iPredict-DR: software that analyzes color retinal fundus images captured on an iCare DRSplus camera and detects more-than-mild diabetic retinopathy in adults with diabetes who have not been diagnosed with it. The clearance rests on a clinical validation trial covering diagnostic performance, safety and usability, and the product is commercially available across the US now.

Diabetic retinopathy is one of the leading causes of preventable blindness, and the cruelty of it is that it is both silent and treatable: by the time vision changes are noticeable the damage is often permanent, while catching it early makes it manageable. The recommended defense is an annual dilated eye exam, and an enormous share of Americans with diabetes never get one. Not because they refuse, but because it means a separate appointment, with a separate specialist, in a separate building, at a separate cost, and that friction is doing more damage than the disease. The line that actually matters is the company's own: the system is non-invasive and "easy to use by any minimally skilled healthcare worker or nurse." That sentence is the entire point, because it means the screening can happen where the patient already is, in the primary-care visit they were going to attend anyway, instead of at a referral they were never going to make. Founder and CEO Alauddin Bhuiyan frames the mission as making retinal screening accessible in primary care and community settings. Two things I want you to have, though. This is a screening tool, not a treatment and not a substitute for ophthalmic care: a positive result is a referral, and a specialist still treats you. Second, the published coverage of the clearance does not disclose sensitivity and specificity, and a 510(k) is a substantial-equivalence determination rather than evidence of superiority over the autonomous screening systems already on the US market. Those numbers live in the FDA's 510(k) summary, which is the document a deploying clinic should read instead of the press release. It is the same machinery as the other three stories. A model, scoring a human being, off a pile of data. The difference is that a regulator looked at it, the purpose is written down, and at the end of it somebody gets told to go see a doctor.

Healio report July 9 2026 FDA grants 510k clearance to iHealthScreen for iPredict-DR AI powered screening software detecting more than mild diabetic retinopathy in adults with diabetes analyzes color retinal fundus images from iCare DRSplus camera easy to use by any minimally skilled healthcare worker or nurse enabling screening in primary care and community settings commercially available across the US founder CEO Alauddin Bhuiyan — healio.com · July 9, 2026

Why this matters: If you or someone you love has diabetes, the ask at your next primary-care visit is now a different ask, and it takes one sentence: "Do you do the AI retinal screening here, or do I need a referral?" The most common reason people with diabetes go blind from retinopathy is not that treatment failed. It is that the annual exam never happened, because it meant another appointment with another specialist somewhere else. Action this week: Ask that question, and hold on to two things: a screening result is not a diagnosis, so a positive means you get referred and you should go, and a negative does not retire your eye care forever, it resets the clock on it. If your clinic does not offer it, ask whether they can refer you to a community health center that does, and ask your insurer whether the screening is covered, as it commonly is. If you procure or deploy clinical AI, use this as the template and then do the diligence the coverage skipped: pull the actual 510(k) summary from the FDA database (accessdata.fda.gov) and read the sensitivity, the specificity and, critically, the demographics and image quality of the validation set, because retinal-image models have a documented history of degrading on populations and camera hardware they were not validated against, and your patient panel is not the trial's panel. Then confirm the cleared indication, which here is adults with diabetes not previously diagnosed with retinopathy, a narrower population than "anyone with diabetes." Using a cleared device outside its indication is your liability, not the vendor's. If diabetes runs in your family, send this to them tonight.

healio.com: FDA clears AI-powered screening software for diabetic retinopathy (July 9, 2026)
ihealthscreen.org: iHealthScreen, iPredict-DR
accessdata.fda.gov: FDA 510(k) premarket notification database
pharmashots.com: iHealthScreen reports FDA clearance of iPredict-DR for diabetic retinopathy screening (July 2026)

» What to watch this week

Whether the LA Board of Police Commissioners adopts the inspector general's recommendations today, and what happens to the Axon contract when it expires at the end of July. LAPD has already dropped Flock; Axon is the bigger test, because dropping one vendor is a procurement decision while adopting mandatory annual audits is a governance one. Watch which of the two the commission actually does.
Whether any other department publishes its own false-hit rate. Plenty of departments are almost certainly generating hits like these. What is unusual is that somebody counted. Flock runs on the order of 90,000 to 100,000 cameras nationally, reading roughly 20 billion plates a month, and almost none of those departments publish an error rate. Whether the next one does is what turns this into a standard instead of a one-off.
Whether Meta moves to compel arbitration, and how fast. The class-action waiver is the company's strongest card and the plaintiffs are in court only for an injunction. If Meta compels and the injunction fails, the substantive question (whether a human signing a machine-ranked list is a decision-maker) never gets litigated in public at all, which is the outcome to watch for.
Whether Uber moves to dismiss on consent grounds, and whether the disclosure remedy survives. Carranza's ask is disclosure rather than damages, which is harder to settle quietly. If a court orders Uber to publish what factors set driver-facing pay, that ruling reaches every platform running personalized pricing, and the FTC has an open surveillance-pricing docket that would inherit it.
Whether iHealthScreen publishes the sensitivity and specificity figures, and on what validation demographics. The clearance is real and the design is right; the numbers are the part the coverage skipped. A company confident in its validation set usually says so.

Tomorrow's signal lands here.