The Face That Follows

That is the unsettling shift beneath Joseph Cox’s September 28 introduction to a 404 Media interview with Kashmir Hill. Cox describes viral accounts taking clips of people, running them through facial-recognition software, and publishing names and other personal information. A passing appearance becomes the beginning of a dossier rather than the end of an encounter.

The introduction is not a prevalence study. It does not establish how often these searches succeed, how many people use them, or which safeguards individual services provide. Nor does its warning establish that everyone is now identifiable. What it supplies is a concrete practice worth examining: recognition used not merely to see a person, but to move information about that person between contexts.

The face has become a possible lookup key. The important question is who gets to turn it.

Recognition is not one operation

“Facial recognition” can describe significantly different tasks. NIST’s December 2019 report on demographic effects distinguishes one-to-one verification of an asserted identity from one-to-many identification searches through a photographic database.

That distinction is more useful than treating every appearance of the technology as the same invasion. Checking whether a face matches an identity someone has presented is not the same operation as asking a database who an unknown person might be. Both need scrutiny, but they pose different questions about enrollment, consent, access and consequences.

A voluntary verification system can begin with someone choosing to identify themselves. An identification search can begin with someone else choosing to identify them. That reversal matters even before the algorithm produces a result.

Neither operation creates a name from facial geometry alone. For a search to return a named person, some part of the surrounding system must associate an image or record with that identity. For it to expose further personal information, another association must connect that identity to those details. The model’s comparison is one step; the assembled information service does the rest.

This is why arguing solely about the quality of the recognition algorithm misses the architecture of exposure. A comparison can be technically impressive while the database behind it is unjustifiably assembled, the query is abusive, or the result is distributed without a defensible purpose. Those are separate failures. A higher accuracy score cannot average them away.

The encounter loses its boundary

Consider the practice Cox describes, without assuming every search works. Someone appears in a clip. Another person searches their face. A candidate identity is attached, then circulated alongside other information. The original audience and the original purpose of the recording no longer contain what happens next.

There is nothing new about strangers recognizing someone in public. The change is that recognition can be attempted by a person who has no prior relationship with the subject, using a database rather than acquaintance. The image can travel to the recognizer instead of the recognizer having to be present.

That gives us a better privacy question than whether the face was already visible. Visibility answers whether someone could see you. Searchability asks whether that sight can be connected to records elsewhere. Publication asks whether those connections can be broadcast. Treating all three as equivalent makes the most consequential transformations disappear into the word “public.”

A person may accept appearing in a crowd without accepting a stranger’s attempt to connect that appearance to their home or workplace. That is an ethical distinction, not a claim that every jurisdiction recognizes the same legal boundary. It remains a useful distinction even when the technology ignores it.

The problem also does not require a futuristic wearable. The clip-and-search practice already described by 404 Media separates capture from identification. Speculation about what might eventually sit on someone’s glasses should not distract from what can happen after an ordinary image leaves its first setting.

The camera does not have to know your name. Someone downstream can ask.

Wrong answers and unwanted right answers

There are two distinct ways this arrangement can go badly. It can identify the wrong person. It can correctly identify someone who should not have been exposed. Better matching addresses the first problem. It can intensify the second.

NIST’s 2019 study is useful here precisely because it refuses to reduce performance to a single reassuring number. It evaluated 189 algorithms from 99 developers using 18.27 million images of 8.49 million people. The report found wide differences between algorithms and demographic differences across many, but not all, of those tested.

It distinguishes false positives—incorrect associations of different people—from false negatives, failures to associate images of the same person. It also reports that some highly accurate identification algorithms had undetectable false-positive demographic differentials in its testing. That matters: neither “all systems fail equally” nor “the technology is accurate” is an adequate account of the evidence.

These are dated results from a defined evaluation, not a scorecard for services operating in September 2026. They do not establish the performance of the tools behind the accounts Cox describes. Importing an old laboratory result as a current product verdict would be its own identification error.

The durable lesson is methodological. NIST recommends knowing the particular algorithm and measuring it on operational image data. It also warns against reporting accuracy without specifying the error being measured, and calls for reporting false-positive and false-negative rates at the configured threshold for demographic groups.

For a person named online, however, the consequences extend beyond the match. A candidate result may be presented as certainty. Other people may copy it. A correction may have to reach audiences the original publisher cannot fully identify. These are risks of the proposed information chain, not outcomes quantified by the NIST study.

Even perfect matching would not settle whether that chain should have been started. Accuracy asks whether the system found the person. Legitimacy asks whether the search and disclosure were justified. The second question is not a bug report for the first.

A restriction is more concrete than a reassurance

There is a documented alternative to treating searchable faces as an inevitable feature of the atmosphere: constrain access to the database.

In May 2022, the ACLU announced a settlement filed in its litigation against Clearview AI under the Illinois Biometric Information Privacy Act. Its account described a nationwide prohibition on making Clearview’s faceprint database available to most private entities, subject to exceptions, and a separate five-year restriction on access by Illinois state and local government entities, including law enforcement.

The announcement also described an opt-out mechanism for Illinois residents, restrictions on how submitted opt-out photographs could be used, and an end to free trial accounts offered to individual police officers without their employers’ knowledge or approval.

Those details matter more than a generic promise to use AI responsibly. They identify particular actors, particular forms of access, and particular operations that must stop or change. The individual-officer provision, for example, concerns organizational authorization rather than the accuracy of a face comparison.

This is the ACLU’s account of the settlement as filed, from a party to the litigation. The announcement said court approval was still required at that point. It is not an independent audit of subsequent compliance, and it should not be read as a universal prohibition on facial recognition. The national restriction it described had exceptions; the Illinois provisions had specified durations and scope.

Nor does restricting one database establish that a person is unsearchable elsewhere. A company-specific remedy is not an internet-wide eraser. Its value is narrower and more concrete: it shows how the question can become one of enforceable access conditions rather than voluntary restraint by whoever happens to possess an image.

The ACLU said its plaintiffs included groups representing survivors of domestic violence and sexual assault, undocumented immigrants, and current and former sex workers. Their inclusion makes the stakes less abstract. The cost of being identifiable is not evenly distributed, and the person making the query is not necessarily the person who bears it.

An appeal needs somewhere to land

The coherenceist test is not whether the system produces a tidy identity record. It is whether the people it affects have a meaningful way to constrain its use and correct its consequences. Order that depends on excluding their objections is simply a well-organized imbalance of power.

That suggests a set of requirements, not a description of safeguards currently available across the industry.

First, an affected person needs a route to learn that a consequential search occurred. A right to contest a result is difficult to exercise when the query, operator and downstream decision are all invisible. Notice can raise legitimate investigative and safety questions, but those questions require bounded rules, not permanent invisibility by default.

Second, inspection should distinguish the stages. What image prompted the search? What candidate record came back? Who attached a name? Who decided the result justified action or publication? If these steps collapse into “the AI identified you,” responsibility becomes as opaque as the machinery.

Third, a correction must address the use made of the result, not merely a record inside one service. If an organization acts on a false identification, fixing the candidate entry alone does not reverse that action. A meaningful process needs an accountable decision-maker with authority to reconsider the consequence.

Fourth, there must be limits on searches that produce accurate answers. Otherwise an appeals system protects only the misidentified. Someone correctly exposed to an abusive former partner has no recognition error to report. The question is whether the search was permitted and whether the information should have been disclosed.

These requirements will not all fit every setting in the same way. They also do not prove that mass identification becomes acceptable once enough forms are added. Some uses may warrant prohibition rather than a more elegant complaint portal. Accountability should include the power to say that an operation must not occur.

The useful measure of progress, then, is not simply how many faces a system can recognize. It is whether access is justified, results are handled as evidence rather than verdicts, and affected people can challenge both mistakes and unwanted uses of correct information.

A searchable face does not automatically end privacy. But it creates a durable temptation to treat a visible person as an available record. The technology industry can keep improving the lookup. Someone still has to defend the difference between being seen and being made searchable.

Otherwise the future arrives with excellent recognition and no one responsible for answering the complaint.