Why Governments Are Quietly Buying AI That Reads Crowds
2,760 cameras watched the 2025 Maha Kumbh Mela, feeding a single AI system tasked with spotting the moment a gathering of tens of millions of pilgrims turns into a stampede. Nobody voted on that system. No agency held a public hearing about whether software should decide, in real time, which clusters of moving bodies count as a threat. The contract got signed, the cameras went up, and the public found out after the fact, the way the public usually finds out about this category of purchase. That's the pattern worth noticing. Crowd-reading AI doesn't arrive through legislation. It arrives through procurement — line items in budgets, vendor demos at security trade shows, pilot programs that quietly become permanent. The directives reflect a "speed first" positioning that frames governance requirements as barriers to rapid deployment [Lawfare](https://www.lawfaremedia.org/article/military-ai-policy-by-contract--the-limits-of-procurement-as-governance) , and procurement officers are not equipped, and were never asked, to referee the line between "detect a stampede" and "track a protest." Most coverage of this technology treats it as a security story or a privacy story. It's actually a story about who gets to make a decision nobody elected them to make. This piece names the vendors, the contracts, and the failure rates that the marketing copy leaves out. It tells you what these systems can verifiably do, what they're quietly sold to do, and where the gap between those two things has already produced a wrongful arrest. Table of Contents 1. What "Reading a Crowd" Actually Means 2. The Vendors Nobody's Heard Of 3. The Accuracy Number That Should Worry You 4. The Loophole the Law Left Open 5. Who Is Buying This and Why 6. The Failure Mode Everyone Downplays 7. Who This Technology Is For 8. Verdict
What "Reading a Crowd" Actually Means
Strip away the marketing language and crowd-reading AI does three distinct things that get sold as one product. The first is density and flow — cameras counting bodies, modeling how a mass of people moves, flagging the physics of a crush before it happens. At the 2025 Maha Kumbh Mela in India, authorities used AI-based software and 2,760 CCTV cameras to monitor crowd density and detect surges, aiming to prevent stampedes [Advantage Technology](https://www.advantage.tech/leveraging-ai-cameras-for-efficient-crowd-management/) . This part works. It's mostly math, and the math is checkable.
The second is behavior classification — software guessing at intent from movement. By analyzing human movement and activity patterns, AI systems can automatically identify loiterers, suspicious shoplifting behavior, crowd formation, and policy violations [Sirix Monitoring](https://sirixmonitoring.com/blog/ai-powered-video-surveillance-for-security/) . This is where things get slippery, because "suspicious" is not a physical property. It's a judgment, and the judgment is now automated.
The third is the one vendors talk about least: identity and sentiment layered on top. Cameras that don't just count the crowd but try to determine who's in it and how they feel about being there.
The Vendors Nobody's Heard Of
The Camera Companies
You won't see these brands on a billboard. Verkada AI Cameras offer facial recognition, people counting, and behavioral analysis, making them useful for monitoring entry points and detecting unusual activity in real-time [Advantage Technology](https://www.advantage.tech/leveraging-ai-cameras-for-efficient-crowd-management/) , sold to transit hubs and event venues as a safety upgrade. Actuate markets itself the same way, pitching AI video analytics that go beyond simple gun detection to include weapon detection, fire detection, and crowd detection [Actuate](https://actuate.ai/ai-crowd-detection-enhancing-safety/) , while telling retail clients the same hardware can track customer movement and behavior to maximize profits [Actuate](https://actuate.ai/ai-crowd-detection-enhancing-safety/) . Same sensor, same model, two entirely different customers buying two entirely different products from the seller's perspective — public safety to one, foot-traffic monetization to the other.
The Sentiment Layer
Then there's the company most people have never heard of until a FOIA request surfaces its name in a procurement record. U.S. Customs and Border Protection partnered with AI firm Fivecast to deploy social media surveillance software that detects "problematic" sentiment and emotion, reporting flagged users to law enforcement [Substack](https://emilytvproducer.substack.com/p/american-thinker-theyre-spying-on) . That's not a camera reading a crowd in a plaza. That's software reading the emotional temperature of a crowd that exists only online, which raises the uncomfortable question of where "crowd" stops meaning a physical gathering and starts meaning anyone posting about one.
I once sat through a vendor pitch for a "public safety analytics" platform that took four slides to admit it was facial recognition with a friendlier name. Nobody in that room asked the obvious question.
The Accuracy Number That Should Worry You
Vendors love a clean accuracy figure. Studies report that surveillance-based anomaly systems have demonstrated accuracy rates exceeding 85% in identifying loitering behavior [visionplatform](https://visionplatform.ai/loitering-and-crowding-detection-in-malls/) — a number that sounds solid until you do the arithmetic at scale.
The math doesn't survive contact with a real city. Even at 99.9 percent accuracy, a system checking faces against a database will produce about a dozen false positives or negatives — a tolerable cost for a trade fair of 10,000 registrants, but a number that multiplies into real danger once police start running the same software across a city of a million people [IEEE Spectrum](https://spectrum.ieee.org/facial-recognition-gone-wrong) .
The number that should worry you isn't the accuracy rate. It's the population it gets multiplied against.
On the best-performing system as of March 2026, calibrated so a false positive match happens only one time in a million, a false negative — missing a real match — still occurs one time in 500 [Federation of American Scientists](https://fas.org/publication/face-recognition-bias/) . That's the trade-off nobody puts on a slide: tune out the wrongful flags, and you let more real threats slip through. Tune up sensitivity, and the false-flag count climbs with it. There is no setting where both numbers are good. As of March 2026, there were at least nine documented U.S. wrongful arrests tied to face recognition misidentification, mostly involving Black people [Federation of American Scientists](https://fas.org/publication/face-recognition-bias/) . Christopher Gatlin spent 17 months in jail for an assault he didn't commit after the software said he matched the suspect; clearing his name took two years [Federation of American Scientists](https://fas.org/publication/face-recognition-bias/) . In 2026, U.S. immigration agents misidentified a detained woman as two different people — not once, but twice [IEEE Spectrum](https://spectrum.ieee.org/facial-recognition-gone-wrong) .
Vendors will tell you accuracy has improved. It has. NIST has found error rates falling by roughly a factor of two every two years under controlled conditions [Federation of American Scientists](https://fas.org/publication/face-recognition-bias/) . But "controlled conditions" is the tell. Performance degrades under poor image conditions — blur, pose variation, reduced resolution — and that degradation is not evenly distributed across demographic groups, with false positive and false negative rates increasing disproportionately for marginalized race and gender groups as image quality drops [arxiv](https://arxiv.org/pdf/2505.14320) . A crowd is, by definition, never a controlled condition. People move, they're backlit, they're partially obscured by other people. The error rate you should care about is the one in the rain, at dusk, in a crowd of thousands — not the one in a lab.
The Loophole the Law Left Open
Europe passed the law everyone points to when this comes up, and the law is narrower than its press coverage. The EU AI Act's prohibitions, which took effect in February 2025, ban emotion recognition — but only in workplaces and education institutions [European Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) . Your boss can't scan your face to decide you're disengaged. Your school can't flag students who look stressed during an exam [Stateofsurveillance](https://stateofsurveillance.org/news/eu-ai-act-august-2026-biometric-surveillance-explainer/) .
Nothing in that list covers a public square.
Crowd-emotion software watching a protest, a transit station, or a stadium falls entirely outside the ban — it was never aimed at that use case in the first place. Post-event facial recognition, run on footage after the fact rather than live, is merely classified "high risk," not banned outright [Stateofsurveillance](https://stateofsurveillance.org/news/eu-ai-act-august-2026-biometric-surveillance-explainer/) , and individual EU member states can pass national laws authorizing exactly what the Act nominally prohibits [Stateofsurveillance](https://stateofsurveillance.org/news/eu-ai-act-august-2026-biometric-surveillance-explainer/) . Read the regulation as written, not as summarized, and the protection most people assume exists for public gatherings simply isn't there.
The American picture is governance by handshake, not statute. Over the past year, the rules governing AI's use by the military have increasingly come not from statutes or regulations but from bilateral agreements between the government and individual vendors [Lawfare](https://www.lawfaremedia.org/article/military-ai-policy-by-contract--the-limits-of-procurement-as-governance) — agreements that bind only the parties who sign them and were never designed to provide the democratic accountability or institutional durability that statutes provide [Lawfare](https://www.lawfaremedia.org/article/military-ai-policy-by-contract--the-limits-of-procurement-as-governance) . Draft federal procurement language has gone further, proposing that the government be granted rights to use AI tools for "any lawful" purpose, language that emerged directly out of a public dispute between a major AI vendor and the Pentagon over surveillance restrictions [Nextgov.com](https://www.nextgov.com/acquisition/2026/04/trade-and-industry-groups-warn-risks-gsas-draft-ai-procurement-guidance/412614/) .
Who Is Buying This and Why
Three buyers, three different appetites.
- Event organizers and transit authorities buy density and flow tools because a stampede is a measurable, insurable, politically catastrophic failure they need to prevent, and the crowd-physics layer of this technology genuinely helps with that.
- Police departments buy the behavior-classification layer because "proactive" is the word every chief wants in a budget hearing, and a system that flags loitering before a theft happens sounds like prevention rather than surveillance.
- Federal agencies buy the sentiment and identity layer because the legal exposure of monitoring a named individual is higher than the exposure of monitoring "the crowd," and crowd-level analysis is, conveniently, exempt from most of the protections built for individuals.
That third category is where the architecture gets clever in a way that should bother you. Neighborhood doorbell cameras, license plate readers, and hyperlocal social media sites create a crowdsourced record of people's movements in public spaces [Salon](https://www.salon.com/2026/04/23/us-government-ramps-up-mass-surveillance-with-help-of-partner/) — data nobody consented to feeding a surveillance pipeline, gathered by tools people bought to watch their own porch.
The Failure Mode Everyone Downplays
You are at a protest. The software watching the crowd has decided, on the basis of clustering and movement, that this gathering merits a "behavioral anomaly" flag. You did nothing. You are simply standing in the part of the crowd the algorithm weighted heavily. Cultural differences in emotional expression can lead to inaccuracies and biases in AI models trained to read crowd behavior, and these models perform poorly outside controlled environments, risking misjudgments that result in unfair treatment of certain individuals [PubMed Central](https://pmc.ncbi.nlm.nih.gov/articles/PMC12103326/) . A study interviewing ten professional crowd managers found that, at the time, AI played little to no role in their actual day-to-day practice [PubMed Central](https://pmc.ncbi.nlm.nih.gov/articles/PMC12103326/) — the people whose job is reading crowds for a living were not yet relying on the tool being sold to replace their judgment.
Here is the failure mode advocates skip past in the demo: the system doesn't know the difference between a riot forming and a flash mob proposing marriage. It knows density, velocity, and clustering. Everything past that is inference, and inference is where the false positive lives.
Three studies in this piece use three different accuracy baselines — 85 percent for loitering detection, 99.9 percent for face verification at small scale, one-in-500 for face verification at national scale — and none of them are measuring the same thing the marketing copy implies they're measuring.
Who This Technology Is For
A transit authority managing a single, predictable, ticketed event — a stadium, a parade route, a religious pilgrimage with known entry points — gets real value from the density and flow layer, and the Maha Kumbh Mela deployment is a legitimate case for it.
A police department considering the behavior-classification layer should know it is buying a tool that will flag innocent people at a rate determined by population size, not by the vendor's accuracy slide, and that the department — not the vendor — will own the wrongful stop that results.
A federal agency or city council weighing the identity and sentiment layer should know it is operating in the exact regulatory gap the EU AI Act left open and the U.S. has not yet closed by statute, which means the only constraint right now is whatever the contract says, and contracts get renegotiated.
Verdict
Buy the crowd-physics layer. Don't buy the rest without a public vote on it first.
The density and flow technology has a clean track record and a clean use case: prevent crushes, model bottlenecks, save lives at scale events. That part of this market deserves to exist and deserves to grow. The behavior-classification and sentiment layers are a different product wearing the same marketing language, sold on accuracy figures that evaporate the moment you do the population math, and currently governed — in the U.S. — by vendor contracts rather than law, and — in the EU — by a statute that explicitly didn't cover this use case. Any agency buying the full stack because it came bundled with the part that actually works is buying surveillance infrastructure it never had to justify to anyone.
The thing nobody quite says out loud: this technology was never approved by anyone who has to answer to the people it watches.
FAQ
Is it legal for police to use AI to monitor crowds at protests?
In the U.S., there is no federal statute specifically barring it, and current oversight runs through individual procurement contracts rather than law. In the EU, the AI Act's emotion-recognition ban applies to workplaces and schools, not public gatherings, so protest monitoring generally falls outside it.
What companies sell crowd surveillance AI to governments?
Verkada and Actuate sell camera-based behavioral analytics to both public agencies and private venues. Fivecast has sold sentiment and emotion-detection software to U.S. Customs and Border Protection for monitoring online activity tied to travelers.
How accurate is AI crowd behavior detection really?
Vendor-cited figures for loitering and anomaly detection sit around 85 percent under tested conditions. Facial recognition systems used at city scale produce a false non-match roughly one in 500 times even when tuned to minimize false matches, and accuracy drops further with blur, poor lighting, or partial obstruction — all normal conditions in a real crowd.
Does the EU AI Act ban emotion recognition AI?
Only in workplaces and education institutions. Public-space and law-enforcement use of emotion-reading systems is not covered by that specific prohibition, though related biometric rules may still apply depending on the system.
Can facial recognition lead to wrongful arrest?
Yes, and it already has. Documented U.S. cases include a man jailed for 17 months over a false match and immigration agents misidentifying one detained woman as two separate people in 2026.
Why do governments buy surveillance AI through procurement instead of passing laws first?
Procurement moves faster than legislation and doesn't require public hearings. Contracts between agencies and vendors currently function as the de facto rulebook for how this technology gets used, which means the terms can change without a vote.
What's the difference between crowd density AI and crowd behavior AI?
Density and flow systems count people and model movement patterns to prevent stampedes — a largely mathematical function. Behavior and sentiment systems attempt to infer intent or emotion from that same movement data, which is a judgment call dressed up as a measurement.