Privacy Erosion

On a Thursday afternoon in January 2020, Robert Williams was pulling into the driveway of his home in Farmington Hills, Michigan, when Detroit police handcuffed him on his front lawn, in front of his wife and two young daughters. He spent the night on the floor of an interrogation cell. The next day, detectives slid a piece of paper across the table — a grainy still from a store's security camera showing a man shoplifting watches — and asked him to explain himself. Williams held the photo next to his own face. "This is not me," he said. "You think all Black men look alike?" One of the detectives looked at the image, then at Williams, and said something close to an admission: "The computer says it's you."

The computer was wrong. A facial recognition algorithm had matched the surveillance still to an old driver's-license photo of Williams, and no human being had bothered to check the match before an arrest warrant was issued. Williams — believed to be the first person in the United States wrongfully arrested because of a facial recognition error — was released after thirty hours. His case became a landmark not because it was exotic but because it was ordinary: a demonstration of how a single database query can reach out and pull a specific human being out of his own driveway.

What happened to Robert Williams is the visible tip of something much larger and quieter. Most of the time the infrastructure that identified him produces no arrest, no interrogation, no story — just a log entry, a face matched to a name, a movement recorded. The remarkable thing about surveillance in 2026 is not the dramatic error. It is the routine, silent, comprehensive functioning of a system that has been built, largely without public debate, to identify almost anyone, almost anywhere, at almost any time. This chapter is about that system: how big it has become, why it grew, who it serves, and what it will do to a society that lives inside it.

The Ceiling That Used to Exist

For most of human history, walking through public space carried a natural kind of anonymity. You could cross a city, join a crowd, or step into a shop without leaving behind a permanent, searchable record of having been there. This was not because anyone protected your privacy. It was because surveillance was expensive. A police officer could follow one suspect. A store detective could watch one corner. Tailing a single person around the clock took a rotating team of agents; monitoring an entire population was simply impossible, no matter how much a government might have wanted to. Human labor was the ceiling, and that ceiling did more to protect privacy than any law ever did.

This is the single most important thing to understand about what AI has changed. The protection was never really legal — it was practical, an accident of arithmetic. The Fourth Amendment's guarantee against unreasonable searches, and the "reasonable expectation of privacy" that courts read into it, rested implicitly on the assumption that continuous surveillance of everyone was infeasible. You had an expectation of privacy in your movements not because the street was private but because no institution could afford to watch every street.

Artificial intelligence removes the arithmetic. A camera feeds video to an algorithm that can compare a face against billions of stored images in a fraction of a second, tirelessly, at a marginal cost approaching zero. The task that once required hundreds of officers and months of work — reconstructing where a person went, whom they met, what they did — collapses into a database query. This is why the change is qualitative rather than merely quantitative. It is not that we now have more surveillance than before, in the way that a city has more cameras than a village. It is that the constraint which defined what surveillance could ever produce has been lifted. When the cost of watching one more person falls to nearly nothing, the natural equilibrium shifts from "almost no one is tracked" to "everyone can be, and the only question is who bothers to look."

The Scale of the Machine

Consider how thoroughly the tracking infrastructure has already been assembled, using the three figures that anchor this chapter.

Facial recognition now operates at roughly sixty-five U.S. airports under Transportation Security Administration programs, where a traveler's face can serve as a boarding credential in place of a physical ID. Law enforcement agencies in at least thirty-two states use facial recognition technology, according to work by the Georgetown Law Center on Privacy and Technology, for purposes ranging from identifying suspects in security footage to mapping who appears alongside whom at public gatherings. And Clearview AI, a single private company, has scraped more than thirty billion facial images from social media and the open web into a searchable biometric index — a database built entirely from photographs of people who never agreed to have their faces turned into identifiers, and which the company has said it intends to grow further.

Read individually, each of these is a discrete deployment with a plausible rationale — airport throughput, criminal investigation, a commercial product. Read together, they describe something no one voted for: a country in which the ability to identify a specific person from their face, and to reconstruct where that face has been, has been built out across the places people pass through as a condition of ordinary life. The airport, the street, the store, and the social-media photo have been wired into a common capability. The infrastructure of individual tracking is not a proposal. It largely exists.

What the Machine Gets Wrong

The Robert Williams case points to a second, uncomfortable fact: these systems are not as accurate in the field as their marketing implies, and they fail unevenly. In 2019, the U.S. National Institute of Standards and Technology (NIST) tested nearly two hundred facial recognition algorithms and found that most produced far higher rates of false positives for Asian and African American faces than for white faces — in some algorithms, by a factor of ten to one hundred. False matches were also more common for women, the elderly, and children. A year earlier, the researchers Joy Buolamwini and Timnit Gebru had shown in their "Gender Shades" study that leading commercial vision systems misclassified the gender of darker-skinned women up to about a third of the time, while erring on lighter-skinned men less than one percent of the time.

Laboratory error rates translate into human damage in the field. Beyond Williams, at least several other Americans — Nijeer Parks in New Jersey, Randal Reid arrested in Georgia for a crime committed in a Louisiana parish he had never visited, Porcha Woodruff detained while eight months pregnant, Michael Oliver in Detroit — have been wrongfully arrested after facial recognition misidentifications. Nearly all of them were Black. The pattern is not incidental. A technology that fails most often on the faces of the people already most heavily policed does not distribute its errors randomly; it concentrates them. And because officers tend to treat an algorithmic match as strong evidence rather than a lead requiring corroboration — "the computer says it's you" — the error propagates through the justice system with a false authority attached.

There is a deeper epistemic problem underneath the demographic one. Vendors report accuracy figures from controlled tests with high-quality images. Operational conditions — a low-resolution camera, a bad angle, poor light, a moving subject — degrade performance in ways that independent audits, and the deployments themselves, rarely measure or disclose. We know the systems misfire, we know roughly whom they misfire against, and we do not know the true operational error rate of most systems in most places, because the agencies running them are under no consistent obligation to test or publish it.

When Public and Private Surveillance Merge

It is tempting to picture government surveillance and corporate data collection as two separate worlds. In practice they have fused into a single apparatus, and the connective tissue is the data broker.

Data brokers are companies that buy information — location pings from smartphone apps, purchase records from retailers, registrations and public filings, browsing and app-usage data — and package it for resale to anyone willing to pay. Because they operate through commercial transactions rather than compulsion, they occupy a gap in constitutional law. In Carpenter v. United States (2018), the Supreme Court held that police generally need a warrant to obtain a person's historical cell-phone location records from a carrier, recognizing that such records reveal "the whole of a person's physical movements." But Carpenter addressed data the government compelled a company to hand over. When an agency instead purchases essentially the same location data on the open market from a broker, it argues that no compulsion, and therefore no warrant, is required. Reporting has documented federal agencies buying phone-location data and other commercial datasets from vendors precisely this way.

This is the "data broker loophole," and it reveals something about the Fourth Amendment as it actually operates in 2026: a warrant requirement that can be sidestepped by pulling out a checkbook is not much of a requirement. The formal legal line between state surveillance (constrained by the Fourth Amendment) and private surveillance (largely unconstrained) has become, in practice, a line the state can cross whenever it prefers to buy what it would otherwise have to justify to a judge. Congress has debated closing the gap — the proposed Fourth Amendment Is Not For Sale Act passed the House of Representatives in 2024 — but as of this writing the market remains open, and the combined system it enables is more comprehensive than either government or industry could build alone. The government gains reach without warrants; industry gains a customer with subpoena power and deep pockets.

The Engine: Why Industry Keeps Building It

Government demand explains part of the surveillance economy, but the deeper engine is commercial, and it runs whether or not anyone intends harm. The dominant business model of the consumer internet is advertising, and advertising is, at its foundation, a business built on behavioral data. To sell an ad impression, an ad-tech system wants to know who you are, what you want, what you fear, what you can afford, and what will move you — and it improves its margins by knowing more. Every incremental data point about a person is, quite literally, worth money.

This is why surveillance expands continuously even absent any villain. No executive has to decide to erode privacy; the incentive gradient does the work. Companies track users across websites, apps, and physical locations because each connection between data points makes the resulting profile more valuable, and AI is what turns scattered signals — a location ping, a search, a purchase, a newsletter subscription — into a coherent, predictive portrait. Those profiles are auctioned through automated exchanges thousands of times a day among firms most consumers have never heard of. The system's expansion is not a conspiracy; it is a thermodynamic tendency of a market that pays for data. Understanding this matters because it means the problem cannot be solved by identifying and stopping bad actors. The pressure is structural, and it points in one direction: collect more.

The Aggregation Problem

The gravest privacy danger is not any single database but their combination. Individually, most of the data being collected looks harmless. That you bought sunscreen, drove down a particular road, searched a medical term, subscribed to a magazine, and stood in a certain aisle for ninety seconds — none of these facts, alone, seems sensitive. Aggregated and linked to a single identity, they compose a portrait of a life more detailed than any secret police force in history could have assembled, and they do it from records freely generated in the course of ordinary living.

AI is what makes aggregation dangerous rather than merely voluminous. Machine-learning systems are built to find patterns in exactly this kind of high-dimensional data, and they specialize in inferring things that were never directly observed. From purchasing patterns, an algorithm can estimate a health condition; from browsing behavior, a political leaning; from location traces, financial stress or a pregnancy; from the metadata of who contacts whom and when, the shape of a person's relationships. These inferences are probabilistic and frequently wrong. But they are produced at scale, applied to real decisions, and almost never disclosed to the person they describe. When an inference is wrong — a mistaken health flag, a false risk score — the cost lands on someone who cannot see the input, cannot see the model, and has no way to contest a conclusion they don't know exists. Insurers, for instance, increasingly use algorithmic analysis of social-media activity, purchases, and inferred movement to inform risk and pricing, drawing on data sources and methods they rarely reveal — so that a person's associations or hobbies, none of them legitimate actuarial factors, can quietly shape what they pay.

And aggregated data is nearly impossible to unwind. A broker typically has no relationship with the people in its files, no portal for them to review what it holds, and, under most U.S. law, no strong obligation to correct errors or honor deletions. Even where deletion rights exist, the record may already have been copied to dozens of downstream buyers. Once your biometrics, associations, and behavior have been captured and dispersed through the commercial data ecosystem, meaningful control over that information is, for practical purposes, gone.

The entire edifice is legally defended by a single word: consent. Users agreed to the terms of service. Shoppers were notified by a sign near the door. Airport passengers may request alternative screening. In this account, surveillance is something people choose, and anyone uncomfortable can simply opt out.

The account does not survive contact with reality. Terms-of-service agreements run to tens of thousands of words of non-negotiable legal language that can be changed unilaterally; study after study finds that far fewer than one in a hundred people reads them. More fundamentally, refusing them often means giving up services that are no longer optional in any real sense — the communication platforms, marketplaces, maps, and work tools that have become the infrastructure of modern economic and social life. A choice whose only alternative is exclusion from ordinary participation is not a free choice.

Physical space makes the fiction starker still. A sign at a store entrance announcing that faces are being scanned offers no genuine opt-out: you can turn around and leave, but you cannot un-be-identified in the instant of walking in. At the airport, the alternative to biometric processing is a slower line and the prospect of secondary screening — a penalty on refusal, not a real choice, and one most travelers on a schedule cannot afford to make. On the public sidewalk, in the park, on the train platform, there is no notice and no alternative at all. The consent model was designed for a world where surveillance meant a bounded, deliberate interaction: a company you chose to approach, a form you chose to sign. It cannot govern surveillance that is continuous, ambient, and inescapable. Genuine consent in such an environment would require plain-language disclosure of what is collected and why, alternatives that do not cut a person off from essential activities, and a real ability to decide before the fact. Present systems supply none of these — which is why "you consented" functions less as a description of what happened than as a legal ritual performed after the fact.

The Chilling Effect, and How Well We Can Measure It

Surveillance changes behavior even when it never produces an arrest. When people know — or merely suspect — that they are being watched and recorded, they become less willing to attend a protest, research a stigmatized topic, voice a dissenting opinion, or associate with a group that might attract scrutiny. This is the chilling effect, and it strikes at the heart of civil liberties, because the freedoms of the First Amendment depend on a measure of the freedom the Fourth was meant to protect. You cannot fully exercise the rights to speak, assemble, and petition if every exercise of them is identified, logged, and stored against an unknown future. Privacy, in this light, is not a shield for wrongdoing. It is a precondition for free thought, free association, and dissent.

How well is this documented? Better than one might expect, though not perfectly, and it is worth being precise about the evidence. The most rigorous studies exploit the natural experiment of Edward Snowden's June 2013 revelations about mass surveillance. The legal scholar Jonathon Penney found that traffic to Wikipedia articles on terrorism-related topics dropped measurably and durably after that date, consistent with people avoiding searches they feared might be monitored. Alex Marthews and Catherine Tucker found similar declines in privacy-sensitive Google search terms across multiple countries. In a controlled experiment, the communication researcher Elizabeth Stoycheff showed that participants primed with a reminder of government surveillance became less willing to express minority opinions online — a laboratory demonstration of the "spiral of silence." Survey work by PEN America documented writers self-censoring their subjects and research after the Snowden disclosures.

These studies converge, which is what gives the chilling effect its credibility as a finding rather than an intuition. But their limitations should be named honestly, because Question of methodology is a fair one. Observational studies that track search or page-view traffic can establish that behavior changed after a surveillance event, yet cannot fully rule out other causes — a news cycle, changing interests, unrelated events — that moved the same trend. Experiments like Stoycheff's cleanly isolate the effect of a surveillance prime, but in an artificial setting whose intensity may not match the diffuse, ambient awareness of real life. Self-report surveys capture what people say they do, which is not always what they do. The best current answer is that multiple independent methods, each imperfect in different ways, point the same direction — strong evidence that documented surveillance suppresses protected activity, and weaker evidence about exactly how large the effect is in any given setting.

The harm, in any case, radiates outward. When an activist self-censors, a source goes quiet, or a defense attorney grows cautious about which public gatherings to attend, the loss is not confined to that person. It thins the whole ecosystem of democratic participation — the journalism, advocacy, and legal representation that a free society runs on.

The Time Bomb in the Archive

Here is the argument that should trouble even someone untroubled by present-day authorities: the danger of a surveillance archive is not fixed at the moment of collection. Data compiled today, under a government that respects civil liberties, becomes an inheritance passed to every government that follows — including ones with different values. A database recording who attended which demonstration, who associates with whom, who expressed which view, is an instrument whose destructiveness depends entirely on who holds it and what they want. It sits quiet under one administration and becomes a targeting list under another. The infrastructure of repression can be, and historically has been, built in advance of the repression itself.

History supplies the pattern with grim consistency. The Netherlands before the Second World War maintained one of Europe's most thorough and well-intentioned population registries, recording residents' religion for benign administrative purposes; after the German occupation, that registry became a map for locating and deporting Dutch Jews, and the country's death rate among them was among the highest in Western Europe. The Nazi regime's censuses and identification systems were organized with the help of tabulating machines supplied through IBM's German subsidiary, Dehomag — a case, documented by the historian Edwin Black, in which cutting-edge information technology built for administration was repurposed for genocide. In the United States, Census Bureau data, legally protected against disclosure, was nonetheless used to help identify Japanese Americans for internment during the same war. In each case, the data was gathered by a functioning government for ordinary reasons, and turned to persecution when circumstances changed. The distinguishing feature of what is being built now is scale and permanence: not a paper registry in one country but billions of biometric records, distributed and durable, waiting.

Lock-In: Why This Is Harder to Undo Than Most Policy

Most policy mistakes are reversible. A bad tax can be repealed; a failed program can be shut down; a law can be rewritten by the next legislature. Biometric surveillance is different in a way that deserves emphasis, because it changes the stakes of getting the regulation late.

You cannot delete a face from the world. Once a person's biometric template and behavioral history have been captured and copied across distributed commercial and government databases, no statute can reliably recall them. The algorithmic techniques that make identification possible cannot be uninvented. The practical anonymity that existed before these systems were built cannot be restored by passing a law, because the law can stop future collection but cannot reach into the thousands of systems that already hold the past. This is what makes the regulatory window narrowing rather than stable: every year of unrestricted collection adds permanently to a stock of data that no future rule can clear. A privacy law in 2030 governs only what is gathered after 2030; everything gathered before it remains. The choices being made right now — to permit deployment without oversight, aggregation without consent, construction without authorization — are therefore unusually consequential, because they are unusually difficult to take back.

Normalization and the Right to Move Anonymously

There is a slower, subtler danger than any single database: that we simply get used to it. Norms track circumstances, and the circumstances of public life are shifting fast enough that a generation is now growing up for whom ambient identification is not an intrusion but a background condition, as unremarkable as weather. Face-scan boarding is convenient. Recognition-based store entry is frictionless. Doorbell cameras that identify visitors are sold as safety. Each is pitched, and often experienced, as a small gain — and each quietly retires a piece of the anonymity that used to be free. The worry is not that people will be forced to accept surveillance but that they will cease to notice it, and that the very idea of a right to move through the world unidentified will come to seem as quaint as anxiety about telephone-operator eavesdropping.

What would that lost thing actually be? The "right to move anonymously" is not a right to do wrong unobserved. It is the ordinary, unremarked freedom to attend a meeting without it being logged, to visit a clinic or a lawyer or a house of worship without a record, to walk somewhere without generating a permanent, queryable trace. It is the precondition that lets a person try on an idea, seek help for a stigmatized problem, or join a cause before they are certain of it — all without the act being filed away against them. A society that normalizes comprehensive identification does not merely lose privacy as a private comfort. It loses a structural condition of open democratic life, and it may lose it without ever deciding to.

Regulation Against the Clock

Some governments have tried to draw lines. The European Union's AI Act — which entered into force in August 2024, with its prohibitions on certain practices applying from early 2025 and its high-risk obligations phasing in through August 2026 — is the most comprehensive framework yet, and it sharply restricts real-time remote biometric identification in public spaces, permitting it for law enforcement only in narrowly defined, judicially supervised circumstances. In the United States, absent federal legislation, the action has been local: San Francisco, Boston, and dozens of other cities have banned municipal use of facial recognition; some states now require warrants for sustained facial recognition surveillance; and a few have moved to restrict police purchases of commercial data, aiming squarely at the data-broker loophole.

These efforts are real, and they prove that course correction is possible. But three structural features blunt them. First, they are geographically fragmented, and data does not respect jurisdiction: a face captured by a camera in a permissive state can end up queried in a state that formally bans the technology. Second — and this is the deep problem — regulation moves on a legislative clock while deployment moves on a product clock. By the time a harm is identified, studied, debated, drafted, enacted, and enforced, the infrastructure has typically been running for years and the databases have grown to a scale that makes remediation more theoretical than real. The lag is not a bug that better lawmaking could eliminate; it is a structural mismatch between two processes with fundamentally different speeds, and it means regulation tends to arrive to govern a world that no longer exists. Third, even perfectly timed rules cannot reclaim data already collected.

International coordination, which the problem plainly demands, remains mostly aspirational — and the reason it is urgent is precisely that data flows across borders while law does not. Surveillance is a global supply chain: hardware built in one country, software written in a second, data stored in a third, used against the residents of a fourth. In such a chain, the effective standard is set by its most permissive link, because collection and storage can be routed to wherever the rules are loosest and the results imported anywhere. A strong regional regime like the EU's sets standards for its own market but cannot govern exports or foreign storage. Without an architecture that binds the permissive links — baseline cross-border rules on biometric collection, retention, and transfer, with real enforcement — stronger protections elsewhere can simply be circumvented.

What Would Democratic Authorization Actually Require

Behind the practical questions sits a normative one that a serious book should not dodge: should privacy in public be treated as a right that must be affirmatively protected, or merely a preference that individuals forfeit by not opting out? The argument of this chapter is that it is the former, and the reason is structural rather than sentimental. If anonymity in public is a precondition for free assembly, free inquiry, and dissent — and the chilling-effect evidence suggests it is — then it is not the kind of good that should evaporate through mass inaction, any more than the right to a fair trial should lapse because most people never invoke it. A precondition for democratic participation cannot coherently be left to be waived by default, because its erosion damages everyone, including those who never chose to give it up.

If that is right, then surveillance infrastructure of this scope requires democratic authorization commensurate with its power — and most of what exists today was never authorized in any meaningful sense. It was procured. Police departments and municipal IT offices acquired identification systems under generic purchasing authorities never designed to sanction population-scale surveillance, and residents of entire cities were enrolled in searchable databases without a vote, a hearing, or a notice. Adequate authorization would look very different. It would require, at minimum, that such systems be established by legislation rather than purchase order — a deliberate decision by an elected body, not an administrative one; that their existence and scope be subject to mandatory public notice, so that a community knows it is being watched and by what; that they operate under genuinely independent oversight with the power to audit and to stop; and that they carry sunset provisions, forcing periodic, affirmative reauthorization rather than indefinite drift. None of these is exotic. Together they mark the difference between a surveillance capability a democracy has chosen and one it has merely acquired.

What We Still Don't Know

Intellectual honesty requires marking the edge of the evidence. We can measure a great deal — deployment counts, error rates, short-run chilling effects — but the largest question is the one we can least answer: what happens to a society's democratic character when its population lives not for months but for decades under conditions of practically unavoidable identification? We have no clean empirical method for that. The natural experiments we rely on, like the Snowden disclosures, capture a shock and its aftermath over a few years, not the slow settling of a norm across a generation. We cannot run a controlled trial on a democracy. Historical analogies from authoritarian surveillance states are suggestive but confounded by everything else that made those states authoritarian. So the long-run political consequence of comprehensive surveillance infrastructure — whether it corrodes democratic vitality gradually, or whether societies adapt in ways we cannot foresee — sits genuinely in the register of the feared rather than the known. That uncertainty is not a reason for complacency. Given the lock-in problem, it is the opposite: we are being asked to make an irreversible bet whose downside we cannot measure, which is precisely the situation in which caution has the most value.

Summary

AI has removed the practical ceiling — human labor — that once made mass surveillance of public space impossible. That is a qualitative change, not merely more of the same: when the cost of tracking one more person falls to nearly zero, the default shifts from "almost no one is watched" to "anyone can be." The infrastructure to exploit this has largely been built. Facial recognition operates at roughly sixty-five U.S. airports and in the hands of law enforcement across at least thirty-two states, and Clearview AI has assembled more than thirty billion scraped facial images into a searchable index — together describing a country wired for individual tracking that no one voted to create.

These systems are also imperfect and unevenly so: NIST and independent audits show markedly higher false-match rates for Black, Asian, female, and elderly faces, and real people — Robert Williams among them — have been wrongfully arrested as a result. Government and private surveillance have effectively merged, with data brokers as the connective tissue and the "buy it instead of subpoena it" loophole hollowing out the Fourth Amendment's warrant requirement in practice. The engine is commercial: an advertising economy that profits from behavioral data expands surveillance continuously, with no villain required. Aggregation turns individually harmless records into portraits more revealing than any historical regime could assemble, and AI mines them for undisclosed, often-wrong inferences that people cannot see or contest. The consent that supposedly authorizes all this is a legal fiction — unread agreements and opt-outs that penalize refusal or don't exist at all.

The consequences are a documented chilling of protected activity, an archive that becomes a weapon in the hands of any future government that inherits it (as population registries and census systems have before), and a lock-in problem that makes late regulation nearly useless, since collected biometrics cannot be recalled. Regulation — the EU AI Act, municipal bans, warrant rules — is real but fragmented and structurally slower than deployment, and it cannot reach data already gathered or foreign links in a global surveillance supply chain. The strongest normative claim is that privacy in public is a precondition for democratic participation, not a personal preference to be waived by inaction, and that surveillance of this scope demands authorization by legislation, public notice, independent oversight, and sunset provisions rather than quiet procurement. What we cannot yet know — because there is no method to measure it — is what decades of unavoidable identification do to a democracy. Given that the bet is irreversible, that uncertainty argues for constraint now, while constraint still means something.

Key Takeaways

The change is qualitative, not quantitative. AI removed human labor as the ceiling on surveillance. The protection anonymity once enjoyed was never legal but arithmetic, and that arithmetic is gone.
The infrastructure already exists. Sixty-five airports, thirty-two states, and thirty-billion-plus scraped faces are not a proposal — they are a built capability for tracking individual movement, assembled largely without public authorization.
The systems fail, and they fail unevenly. Independent audits show far higher false-positive rates for Black, Asian, female, and elderly faces; documented wrongful arrests, almost all of Black men, are the result.
State and market surveillance have merged. Data brokers let law enforcement buy what it would otherwise need a warrant for, turning the Fourth Amendment's protection into one that can be sidestepped with a purchase order.
The engine is commercial and needs no villain. An advertising economy that monetizes behavioral data makes surveillance expand structurally; aggregation plus AI inference then makes harmless records dangerous.
Consent is a fiction. Unread, non-negotiable terms and penalized or nonexistent opt-outs cannot constitute meaningful agreement to continuous, ambient identification.
The archive is a time bomb, and it is locked in. Data collected under democratic norms becomes an inheritance for any future government; biometric data, once distributed, cannot be recalled, so the window for meaningful regulation narrows every year.
Privacy in public is a structural precondition, not a preference — which is why surveillance of this scope should require legislative authorization, public notice, independent oversight, and sunset provisions, and why the greatest long-run risk is the one we have no way to measure.

Sources

Last updated: 2026-09-08

V2 (in progress) Previous: V1