Facial Recognition and Surveillance — What the Law Has Not Decided

When What Was Brought In as a Clue Becomes the Reason

There is a paper that reads side by side four facial recognition cases brought to American courts. They are the cases of Dillon, Williams, Patel, and Miles. The author writes that the recurring problem across the four cases is not a single flawed algorithm, but a chain of interlocking risks (DOI: 10.2139/ssrn.7356558, preprint, not peer-reviewed).

The links in the chain are these. The low quality of the query photo put into the system to find someone. The performance that differs by demographic group. Misunderstanding a similarity score as the probability that it is the same person. The habit of believing the machine's answer first before starting. A follow-up confirmation procedure that hints at a specific person. The investigative side not turning over all materials to the defense side. A proprietary system where the manufacturer does not disclose the inner workings. An excessively broad face database.

This paper points to the Williams case as an example showing what a misidentification does to a person. A misidentification is an incident where the wrong person is pointed out as the same person. This case also shows why it is valuable to make corroborating investigations confirmed by other evidence mandatory, it says. In the Dillon case, a different thing is seen. It is that a seemingly precise score can keep an investigator stuck right there. That is true even when facts that contradict it are right in front of them.

A paper that writes about the same problem in the language of procedure goes one step further. The author states at the outset that technical errors and uneven performance are important, but they are not the main concern of this paper. What the author is looking into is the risks of institutions and procedures. They look at when a tool brought in to generate investigative clues is treated as if it has already established identity. Then a provisional candidate can be promoted into an actionable suspicion. A warrant is a document where a court authorizes an arrest or a search. That promotion can be used to prop up a warrant application before the grounds are sufficiently verified. When that happens, what failed is not just a single wrong output. It is an unlawful arrest and unjust detention, and the author calls this a constitutional failure (DOI: 10.2139/ssrn.6505861, preprint, not peer-reviewed).

Biometric information is information pointing to that person extracted from bodily features like the face, fingerprints, and irises. If a password leaks, you can change it, but you cannot change your face. This is precisely the nature that a paper in the pages of Korean constitutional law points out. Biometric information including the face, unlike existing personal information, does not change. Therefore, the infringement resulting from leakage is massive (DOI: 10.35901/kjcl.2022.28.4.535).

Before reading, three things are clarified. First, this piece does not deal with the technology of reading emotions from faces. What it deals with is identity identification, namely, whose face this is, and that alone. Second, the grades of the cited studies are uneven. Law review papers often go up on open repositories before peer-reviewed pages. Several citations propping up the conclusions are like that. All the figures below are values observed by those reports under those conditions, so the conditions are written together.

Third is which country's law is being discussed. The judgments and statutes cover those of the United States (court cases and California legislation), the United Kingdom (the Bridges judgment), the European Union (the Artificial Intelligence Act), the European Court of Human Rights (Glukhin v. Russia), and South Korea (the Personal Information Protection Act). Brazil and Canada appear only as cases held by papers. When the abstract clarifies which country's story it is, it is written, and when it does not clarify, it is written that it is not clarified.

There Are Four Blanks in the Word 'Accuracy'

The U.S. National Institute of Standards and Technology (NIST) runs facial recognition algorithms submitted by developers under identical conditions. The results are published on a publicly accessible web page that is updated regularly. Something stands out first in the NIST reports cited by this article: there is no single number called 'facial recognition accuracy'. The reports split depending on what task is being performed and what photographs are used.

1-to-1 and 1-to-Many Are Not the Same Task

A 1-to-many search is the task of comparing a single face against numerous faces in a database in sequence to retrieve candidate matches. By contrast, 1-to-1 verification is the task of merely confirming whether two photographs are of the same person. Checking if a passport photo matches your current face at an airport gate is 1-to-1. Matching a single face captured on a camera against millions of driver's license photos is 1-to-many.

NIST reports covering 1-to-many searches divide this task into two. Positive identification is when the person being looked for is expected to be in the database. An employee entering an office building is an example. Negative identification is the opposite, where it is presumed that the person is not there. Sorting out whether a compulsive gambler is entering a casino is an example. The reports explain why this distinction is useful: the two applications tolerate errors of pointing out the wrong person and errors of missing an existing person to different degrees (DOI: 10.6028/nist.ir.8381). Even with the same algorithm, what counts as a failure changes.

If you hear the phrase "99 percent accuracy," there are at least four blanks you must fill in first. Is it 1-to-1 or 1-to-many? What photographs were used for measurement? Where was the threshold set to declare a match? How large is the database being compared against?

A Score Is Not a Probability

Even if you fill in all the blanks, one misunderstanding remains. That is what the system actually outputs. A similarity score is merely a number assigned by the system indicating how closely two faces resemble each other, not the probability that the two faces belong to the same person. This is just like how the scale reading on a bathroom scale is not a health score. Exactly this is one of the items that the previously examined paper placed in the chain of risk. The numerical value of resemblance is mistaken for the probability of identity.

What the machine judges is resemblance, not identity. Turning resemblance into identity is done by humans.

Errors Are Not Evenly Distributed

In 2019, NIST released a report quantifying performance variation across demographic groups. A follow-up public draft summarizes what has been uncovered in the meantime. This section must be read with the utmost care. It states that the causes differ depending on the direction of the error.

The rate at which two different traces are mistakenly said to have come from the same source—this is the false positive rate. In facial recognition, this is the rate of saying two different people are the same person. Conversely, the rate of missing someone and saying they are a different person when they are actually the same—this is the false negative rate. This is the failure of not finding a face even though it is there.

NIST's summary is as follows. The inequality among groups on the false negative side largely stems from specific groups being poorly captured. In particular, this is a problem where darker-skinned faces are captured with underexposure. There are two solutions. One method is to use algorithms that are more resilient to poor image quality. A better method is to fix the capture stage itself with better cameras and shooting environments. On the other hand, the much larger variation on the false positive side appears even in high-quality photos. For this side, the creators of the algorithm must reduce it (DOI: 10.6028/nist.ir.8429.ipd, public draft). There is one condition. The algorithms whose metrics are tabulated in this report are those submitted to the one-to-one verification track.

Cut this section in half and it becomes false either way. Leaving only the front half turns it into "we just need better cameras," and leaving only the back half turns it into "facial recognition is inherently racist." What the report stated is that the two problems are different and must be fixed in different places.

Improvements Are Also Being Reported

Attempts to reduce bias are also producing numbers. Not including those numbers alongside is also clipping the context. One study notes that it refutes the widespread assumption that accuracy and fairness are in a trade-off relationship. The method is as follows. Uncertain predictions are selected to receive ground-truth label annotations from human experts and are then trained further. In this way, both were improved simultaneously.

On the FairFace dataset, accuracy rose from 92.58 percent to 96.47 percent. The metric indicating the degree of bias dropped from 3.86 to 1.16, a reduction of about 70 percent. In particular, accuracy for Black women rose from 80.2 percent to 90.86 percent (DOI: 10.37284/eaje.9.1.4850). This is a value produced by a single study in a specific way on a specific dataset. It is not "bias has been solved," but rather "under these conditions, it decreased by this much."

There are also reports that state their limitations more clearly. This is the case for a preprint that incorporates a technique (fairness regularization) to reduce performance differences between groups during the training process. The max-min accuracy gap between racial subgroups decreased by 18 to 27 percent, and the standard deviation fell by 12 to 20 percent. Even so, the authors conclude that they could not completely eliminate the gap. There are also counterintuitive results. Between the two baseline models compared, the deeper network actually had a larger variance between groups. The authors suggest that model capacity can amplify demographic imbalances (DOI: 10.21203/rs.3.rs-8330446/v1, preprint).

We apply the rules established in the previous section to these numbers as well. The abstracts of these reports do not specify the task for which accuracy is measured. There is no mention of whether it is a one-to-one verification or a one-to-many search. Nor is there any mention of where the threshold for declaring a match was set. The only blank filled in is a single dataset name. The questions we asked about "99 percent accuracy" remain just as valid for the side reporting improvements.

How Errors Lead to Arrests

What the papers we look at next point to as a problem of the law is not "how often algorithms are wrong." The answer the law gives is that a human should check it at the end. One chapter of the research book we will look at next writes as follows: in algorithm-based policing, meaningful human intervention is a legal requirement (DOI: 10.1093/9780191979927.003.0010). Which country's law it is does not appear in the abstract.

There is research that went inside to see how that requirement operates in the field. Based on rare field access granted for the operation of police facial recognition, this is an ethnography written by directly observing the field. The author writes that not only whether human intervention happens, but also what the nature of that judgment is, remains a core unresolved question. The observed decision points were sometimes the intelligence unit in the control room and sometimes the intervention team on the street. Across those diverse deployments, one thing was clearly identifiable. The author calls this automation presumption. It means that discretion has shifted over to the algorithm's suspicion verdict. Discretion refers to the room to see a situation and decide for oneself. It was that way regardless of whether the verdict was accurate (DOI: 10.1093/9780191979927.003.0010).

Let us give it a name. The habit where a human treats the answer put out first by a machine as the starting point and default for review is called automation bias. If you have ever turned the steering wheel in the direction a navigation system pointed even when it indicated a strange route, that is it.

The conclusion of the same research book is stronger. Repeated claims regarding efficacy, racial fairness, and public safety lacked evidence when examined closely. The role of the safeguard itself—that a human makes the final check—was being neutralized. However, it does not end with diagnosis. The author separately creates and presents a human rights due diligence framework tailored to the police use of facial recognition. It is a framework that examines the risk of rights violations in advance (DOI: 10.1093/9780191979927.001.0001).

Let us see what remains at the end of that. A briefing summarizing city-level police agency research for practitioners writes as follows: agencies that used facial recognition routinely had a larger arrest gap between Black and white people, and especially so in adult arrests. And it immediately qualifies this. This research does not show that facial recognition software caused that gap on its own. A more plausible concern is that this technology deepens pre-existing gaps. Depending on how and where it is used. The briefing's recommendation is short. It is to treat facial recognition as an investigative lead rather than standalone evidence, and to make corroborating investigation, recording, supervision, training, and regular auditing rules (DOI: 10.22215/apb.v3i3.6097).

The argument of a human rights law paper taking Brazil as a case study is even more uncomfortable. Even if it is a technically accurate system, it is bound to produce discriminatory outcomes. That is so when placed in a context of policing that has targeted specific races and of structural inequality. The authors cite scale and the difficulty of challenging algorithmic verdicts as grounds. The method of picking out people in public spaces in real time based solely on faces therefore creates a distinct kind of discrimination harm (DOI: 10.1017/bhj.2026.10047).

And a Report in the Opposite Direction

If we stop here, we have only looked at one side. There is research covering 508 U.S. city police agencies from 2016 to 2020. This research reports that the use of facial recognition is directly linked to a significant reduction in police-related deaths. Especially among white residents. Fixed camera density alone showed almost no independent association. When combined with facial recognition, it was projected that as cameras increased, deaths involving Black residents steadily decreased.

The authors also note a concerning trade-off. The use of facial recognition was correlated with an increase in the rate at which police officers were assaulted. Adding more fixed cameras in environments where facial recognition is used increased that risk further. And when police body-worn cameras reached saturation, that death-protection effect seen among white residents weakened. The conclusion is that the public safety benefits of this technology depend on governance choices regarding camera configuration and deployment (DOI: 10.2139/ssrn.6753771, preprint).

This is an observational study, and the authors also write in the language of association. Still, the fact that such a report exists at one end of the debate cannot be erased.

Consent — Can You Ask for Consent from Someone Passing in Front of a Camera?

There is one door that personal privacy legislation has long leaned on. It is consent. This is the principle that if my information is to be used, I must permit it. What The Right Be Forgotten — The Internet Never Forgets also covered was whether information once made public could be reeled back in. Yet, discussions continue that this door does not work as intended in front of faces.

A paper that directly examines consent in privacy law places two methods side by side and writes that both fail (DOI: 10.2139/ssrn.4333743, preprint). One is the notice-and-choice approach used mainly in the United States. Companies post privacy policies. If people continue transactions or do not opt out, they are considered to have consented. The other is the explicit consent approach of the European Union's General Data Protection Regulation (GDPR). A person must voluntarily and actively consent.

On the notice-and-choice side, there is no evidence of actual consent, and people are pressured or manipulated. The explicit consent side also collapses under scale. Even experts do not know what algorithms will do with that data, and requests coming from thousands of places cannot be handled. The author's conclusion is that consent in privacy is mostly a fiction. However, the author does not close the door there. They suggest demoting consent to a very weak and limited license, and stacking four duties on top of it. These are the duty to obtain proper consent, the duty not to dash reasonable expectations, the fiduciary duty, and the duty to avoid unreasonable risks.

There are also papers arguing that this failure is structural when it comes to faces. A paper dealing with the aftermath of facial recognition being embedded in public infrastructure uses the expression "structural collapse of the notice-and-consent paradigm." It argues that because of asymmetric power and function creep, consent in such an environment has become a sham.

This paper also puts forward an alternative. It is a public-private partnership model that views biometric information as a quasi-public good—an asset close to a public good—and ties risk distribution and profit sharing together. The author notes themselves that it is a hypothesis (DOI: 10.62517/jel.202614123).

Where Does South Korea Stand?

South Korea protects biometric information as sensitive information through the Personal Information Protection Act. Sensitive information is information whose processing is restricted more stringently than general personal information. The constitutional law paper cited earlier acknowledges this framework while viewing improvement as necessary. It points out three things. Regulations allowing the exceptional processing of sensitive information must be more specific. Regulations on damages must be improved. Regulations on intelligent CCTVs equipped with facial recognition must be newly established. The author cites the Biometric Information Privacy Act of Illinois in the United States as a contrast case. It also introduces the fact that Facebook faced $650 million in liability for damages for violating that law (DOI: 10.35901/kjcl.2022.28.4.535).

Regarding automated decisions in general, Article 37-2 was newly established in South Korea's Personal Information Protection Act. This is a clause dealing with decisions made by systems without human intervention. A paper comparing this clause to Article 22 of the European General Data Protection Regulation points out that the level of protection is lower in three places.

First, South Korean law grants a right to refuse rather than a general ban. Instead of blocking things in principle, it works only when the party concerned raises an objection. Therefore, there are limits to protecting people who cannot exercise that right. Second, it regulates a state of complete automation at the level of the entire system. Thus, a regulatory blind spot arises for multi-step profiling. Profiling is analyzing and evaluating an individual's characteristics. Third, limits remain on the right to demand an explanation (DOI: 10.1057/s41599-024-03470-y).

This clause was not made targeting facial recognition. However, the structure where a camera picks out a face and a person attaches behind it is precisely multi-step. This article reads that the pointed-out blind spot can also reach facial recognition. Yet, whether this clause applies to investigative agencies' facial recognition is not covered by the cited abstract.

The UK Bridges Ruling — A Won Lawsuit and the Room Left by the Ruling

A paper introduces the Bridges ruling from the UK as the world's first successful legal challenge to this technology (DOI: 10.52214/stlr.v22i2.8666). It was a case contesting South Wales Police's operation of live facial recognition. It is an operation that instantly matches the faces of passing people. The UK Court of Appeal judged that operation unlawful in August 2020 for three reasons.

First, it infringed on the right to respect for private life guaranteed by Article 8 of the European Convention on Human Rights. The European Convention on Human Rights is a human rights treaty made by European countries. The reason it was seen as an infringement was because there was inadequate basis in law. Second, the data protection impact assessment pursuant to Section 64 of the UK Data Protection Act 2018 was inadequate. This is because it did not assess risks to the rights and freedoms of individuals processed by the system. Third, it failed to fulfill the public sector equality duty imposed by Section 149 of the UK Equality Act 2010. This is because it did not assess whether the software was biased regarding gender and race (DOI: 10.1111/1468-2230.12623).

The third reason is worth reading again. The basis on which the court judged it unlawful was not "it is biased," but "they did not check whether it is biased." Here is the shape of what the law demanded of the technology in this ruling.

However, there is a paper that views the room left by the ruling more broadly. According to this paper, the court viewed facial recognition in public spaces as an acceptable means to achieve law enforcement purposes. The paper writes that this includes cases of identifying and tracking a very large number of people. It is also said not to have demanded transparency in the technology used by the police. The author concludes that while this ruling is a reassuring judgment requiring changes in police discretion in the short term, it rather justifies expansion in the long term (DOI: 10.52214/stlr.v22i2.8666).

The European Court of Human Rights has also given an answer. It is a court that deliberates on whether the European Convention on Human Rights has been violated. In Glukhin v. Russia, the court recognized violations of Article 8 (private life) and Article 10 (freedom of expression) of the Convention. A paper analyzing this case confirms that using facial recognition in criminal procedures creates interference with the right to private life. Whether that interference is justified can yield different answers depending on the case, it says. However, the paper writes that neither the constitution of the country covered nor the European Convention on Human Rights bans the use of facial recognition in criminal procedures in and of itself (DOI: 10.59295/spd2024j.10, conference proceedings). This is a general theory that use itself is not banned, and the point that a violation was recognized in this case remains unchanged. The abstract does not reveal which country's constitution it is. It is not a judgment on the South Korean Constitution.

Article 5 of the European Union Artificial Intelligence Act — Bans and Exceptions Attached to the Bans

Among the laws this article covers, the one that explicitly specifies bans in its articles is Article 5 of the European Union Artificial Intelligence Act. Let us first unpack one of the acts prohibited by this provision. Remote biometric identification refers to figuring out who someone is by comparing their physical traits against a database, usually from a distance, even if the other party is not cooperating. Fingerprints are taken only when you press your finger down, but faces are captured even without touching anything.

Article 5 prohibits eight practices that are incompatible with human dignity, autonomy, and fundamental rights. A paper analyzing this provision introduces them as an absolute ban, regardless of whatever safeguards are attached. The same abstract states that the analysis also covers applicable exceptions and boundary cases (DOI: 10.2139/ssrn.6973579, preprint). Both statements coexist within a single abstract.

A chapter of a book dealing with the Artificial Intelligence Act separates the two. This chapter divides the bans under Article 5 into total bans and conditional bans. Real-time remote biometric identification conducted in publicly accessible spaces for law enforcement purposes is not an absolute ban. It is a "complex regime" accompanied by conditions, safeguards, and member state approval procedures (DOI: 10.1093/law/9780198925705.003.0003). What those conditions specifically are is found in neither of the two abstracts cited by this article. All that can be confirmed is the fact that conditions are attached.

Among the eight, four apply to facial or bodily characteristics. They are indiscriminate scraping, emotion recognition in the workplace and educational institutions, biometric categorization based on sensitive attributes, and real-time remote biometric identification. Among them, two are directly tied to identity identification. One is scraping images indiscriminately to build a facial database. The other is the real-time remote biometric identification we just looked at. These bans have applied since February 2, 2025. Sanctions became enforceable starting August 2, 2025. The cap is 35 million euros or 7 percent of global annual turnover, whichever is higher (DOI: 10.2139/ssrn.6973579, preprint).

Even the regulation specifying the bans contains both bans and exceptions. How to view those exceptions is where the prescriptions of the papers diverge.

One side says to eliminate the exceptions. The Brazilian paper we saw earlier says not to copy Europe's exception clauses. The author's conclusion is that countries with high structural inequality should introduce a total ban. The scope is limited to real-time remote biometric identification for law enforcement purposes. The reason is that certain technical practices are inherently incompatible with constitutional guarantees of equality and non-discrimination (DOI: 10.1017/bhj.2026.10047).

The other side opposes a total ban. A paper reviewing real-time facial recognition by British law enforcement agencies proposes a gradual approach. Instead of a sweeping overhaul of regulatory change, the idea is to patch the existing legal framework bit by bit. Specifically, it establishes national-level rules governing the list of faces to be matched. Spatial, temporal, and contextual limits are placed on deployment. The basis is a proportionality and necessity assessment, meaning whether it is not excessive relative to the purpose and whether it is strictly necessary. To this, transparency procedures are added.

The same paper also offers a path that can be chosen instead of transparency procedures. It is to limit public use exclusively to crime prevention and ensuring public safety, only when transparency does not compromise effectiveness. Confined this way, law enforcement agencies would primarily use covert surveillance in investigations. This is surveillance authorized in advance by a judicial authority, except in urgent cases (DOI: 10.1017/s0008197323000454).

What happened in between is also recorded. A paper analyzing media reports in Canada sets down the starting point like this. In January 2020, an American newspaper reported on Clearview AI. Clearview AI was an American company that built a database with billions of images scraped from the internet. Without any public review, it had already secured over 600 police agencies as customers in the United States alone. Deployment had preceded public review.

The authors also jot down the aftertaste. The complex relationship between the media and the police softened the criticism. The result was that space opened up to justify other, less controversial facial recognition use cases in its place (DOI: 10.2139/ssrn.7308163, preprint).

California — Why the Ban on Facial Recognition on Body Cams Shrank from Permanent to Three Years

This is a case from the legislative process in the state of California. In 2019, California passed a law banning police from using facial recognition on body cam footage for three years. This temporary ban expired on January 1, 2023, and was reportedly not extended. The bill was originally proposed as a permanent ban. There is a study that tracked why it was reduced to three years using legislative documents, industry reports, and media coverage.

Early on, lawmakers and civil society criticized that the technology could worsen discrimination against minorities and people of color. However, critics later shifted the rationale for the ban toward the low accuracy of the technology itself. This allowed proponents, including manufacturers and police lobbyists, to push back. Their counterargument was that the technology is more accurate than humans and will only get better. The authors argue that this counterargument weakened the need for a permanent ban.

The authors conclude that opposition to algorithmic governance—meaning the practice of managing people through algorithms—is effective only when it is biopolitical rather than tech-centric. In plain terms, this means that the argument is more effective when it focuses on how bodies and populations are governed rather than how well machines get the right answers (DOI: 10.16997/book55.m).

In this case, accuracy was a poor axis to use as a weapon in the debate. This is because it is an axis that the opposing side can parry by saying, "It will get better soon."

Public opinion research offers a curious clue. There is a survey using a national sample of US adults. It reports that support for facial recognition is influenced by both belief in public safety and privacy concerns. The effect of privacy concerns varies depending on the policing context and trust in the police (DOI: 10.1177/00111287221150172).

An experiment from a study using a crowdsourced sample—meaning participants gathered online—also stands out. When 496 participants were shown information about demographic bias in facial recognition, their opposition increased. That effect was hardly swayed by ideological orientation. However, the same paper also reports that ideological orientation predicts initial acceptance. It had a positive direction with right-wing authoritarianism and a negative direction with libertarianism (DOI: 10.1177/09636625221113131).

The conclusion from reading these two records together in this article is as follows. There is one experiment showing that exposure to bias information can increase opposition. However, when the information given is about "accuracy," that attitude is held hostage by technological improvement. This means that as technology gets better, opposition loses its ground to stand on.

What Has Not Yet Been Decided

We can now list what the law has not yet decided.

First, what status machine-made clues hold within procedures. Here lies the conclusion of the paper synthesizing the four cases. The legitimacy of facial recognition in U.S. courts depends much less on whether the algorithm can generate candidates than on something else. It is whether, after that candidate is created, legal procedures preserve it in a state where it can be cross-examined. The author proposes a conditional use model. Facial recognition remains an investigative clue and does not become the sole basis for arrest or guilt. The practical details of the search are disclosed, and independent evidence must exist prior to the warrant. The author describes when errors become most dangerous: when clues contaminate subsequent investigative stages and disappear from view before trial.

Second, how to count the harms of surveillance. International human rights law evaluates measures that restrict rights through three steps. It asks whether there is a legal basis, whether they pursue a legitimate purpose, and whether they are necessary in a democratic society. A chapter applying this evaluation in the ethnographic research book mentioned earlier points this out. This technology touches not only privacy but simultaneously freedom of expression, freedom of assembly, and non-discrimination. In doing so, it proposes the concept of compound human rights violations. The chilling effect is the effect where people voluntarily reduce their exercise of rights because of the feeling of being watched. Because this effect touches multiple rights at once, it cannot be captured by counting them one by one (DOI: 10.1093/9780191979927.003.0004). The same chapter also introduces a newly asserted right. It is the right to be judged by a human.

Third, what happens to the very act of gathering in public spaces. There is a chapter dealing with the chilling effect of public-space facial recognition on peaceful assemblies and political protests. This chapter points out the absence of oversight and accountability mechanisms, noting that state power has been greatly expanded. It argues that the gap should be filled with legally binding human rights obligations. These are obligations that bind both the government and the technology companies assisting its surveillance (DOI: 10.2139/ssrn.5333821, preprint, not peer-reviewed). Here, the victim is not the person wrongly caught, but the person who chose not to go out. And the person who did not go out is difficult to catch in statistics.

The question running through these three is the same as the question "When an Algorithm Sets the Sentence — Is AI Recidivism Prediction Fair?" asks about recidivism prediction tools. Exactly when does a machine's output come to be treated as fact within judicial procedures, and who can overturn it? In facial recognition, that moment arrives much earlier than the courtroom, in front of the investigator's screen.

So we return to the sentence with which this article opened. When clues become reasons, it is not the algorithm that fails. The verification step that must exist between clue and reason fails. That step does not form automatically even if cameras improve, datasets find balance, or the false positive rate drops. The conclusion of the paper synthesizing the four cases is that procedures must preserve that step.

And what the law has not yet decided is not whether this technology is accurate. It is what we are allowed to do once it becomes accurate. If we hang the grounds for opposition on accuracy, the day accuracy becomes sufficiently good, that opposition will collapse on its own. The question that must remain on that day is different. Do we want this country to be a place where people walking down the street can pass by without a name? That is not a question an algorithm can answer.

Source Papers

Grades (preprints, public drafts) are indicated in the text each time.

• "Use of Facial Recognition in Court" (2026, preprint) — DOI: 10.2139/ssrn.7356558 — Synthesis of four cases. The problem is the chain of risk, and legitimacy in court depends on the possibility of post-hoc contestation.
• "When a Lead Becomes Probable Cause" (2026, preprint) — DOI: 10.2139/ssrn.6505861 — The risk that when a lead tool is treated like an identity confirmation, a provisional candidate can prop up a warrant application.
• "Face Recognition Vendor Test Part 7" (2021) — DOI: 10.6028/nist.ir.8381 — Positive identification and negative identification in 1-to-many search have different error tolerances (airport transfer context).
• "Face Recognition Vendor Test Part 8" (2022, public draft) — DOI: 10.6028/nist.ir.8429.ipd — False-negative inequality is largely due to photography, while much larger false-positive variance remains even with high-quality photos (1-to-1 track).
• "Improved AI-Based FR… Demographic Biasness" (2026) — DOI: 10.37284/eaje.9.1.4850 — Refutation of the accuracy-fairness trade-off assumption. FairFace 92.58%→96.47%, bias metric 3.86→1.16, Black women 80.2%→90.86%.
• "Fairness Regularization in CNNs" (2025, preprint) — DOI: 10.21203/rs.3.rs-8330446/v1 — FairFace and AffectNet. The gap decreased by 18 to 27% but was not eliminated, and deeper networks had higher variance.
• "Human–Technology Interaction… Surveillance Operations" (2025) — DOI: 10.1093/9780191979927.003.0010 — 'Presumed intervention'. Discretion was handed over to algorithmic judgments regardless of accuracy.
• "Facial Recognition Surveillance" (2025) — DOI: 10.1093/9780191979927.001.0001 — An ethnographic study with rare access to police operations. Insufficient basis for claims of efficacy and fairness. Proposes a human rights due diligence framework.
• "Human Rights Law Considerations… Public Surveillance" (2025) — DOI: 10.1093/9780191979927.003.0004 — Three-step human rights review, complex human rights violations, and the 'right to be judged by a human'.
• "Facial Recognition in Policing: Racial Disparities" (2026, practitioner briefing) — DOI: 10.22215/apb.v3i3.6097 — The more frequently an agency used it, the wider the Black-White arrest disparity. No direct causality is visible.
• "FRT in the Global Divide" (2026) — DOI: 10.1017/bhj.2026.10047 — Discriminatory outcomes in racialized policing contexts despite technical accuracy. Argues for a total ban on real-time biometric identification for law enforcement purposes.
• "Police FRT… Racial Differences in Police Lethality" (2026, preprint) — DOI: 10.2139/ssrn.6753771 — Observation of 508 cities. Associated with reduced fatalities, but efficacy weakens when assault rates rise or body cameras are saturated.
• "Murky Consent" (2023, preprint) — DOI: 10.2139/ssrn.4333743 — Notice-and-choice specified consent also fails, and consent is mostly a fiction. The alternative is four obligations.
• "Beyond Consent" (2026) — DOI: 10.62517/jel.202614123 — Structural collapse of the 'notice and consent' paradigm. The alternative is a quasi-public good public-private partnership model (noted as a hypothesis by the author).
• "Facial Recognition Technology and the Protection of the Right to Informational Self-Determination" (2022) — DOI: 10.35901/kjcl.2022.28.4.535 — Biometric information does not change, making leak damage severe. Regulations for intelligent CCTV are needed.
• "Automated decision-making in South Korea" (2024) — DOI: 10.1057/s41599-024-03470-y — Article 37-2 is a right to object rather than a general ban. The vacuum of multi-stage profiling and limits of the right to explanation.
• "Automatic Facial Recognition… Police Surveillance" (2021) — DOI: 10.1111/1468-2230.12623 — The three unlawful grounds in the Bridges judgment.
• "Burning Bridges" (2021) — DOI: 10.52214/stlr.v22i2.8666 — The world's first victory. However, the court viewed even mass identification and tracking as acceptable means.
• "Using FRT to prevent and fight crime" (2025) — DOI: 10.59295/spd2024j.10 — Glukhin v. Russia. An interference with the right to private life, but neither that country's constitution nor the convention strictly prohibits its use in criminal procedures.
• "Prohibited AI Practices Under Article 5" (2026, preprint) — DOI: 10.2139/ssrn.6973579 — Two of the eight prohibitions catch identity identification. While speaking of absolute prohibition, it also lists applicable exceptions.
• "Chapter II: Prohibited AI Practices" (2026) — DOI: 10.1093/law/9780198925705.003.0003 — Distinguishes between total bans and conditional bans. Real-time biometric identification is a regime accompanied by conditions, safeguards, and member state approval.
• "Regulating Live FRT in Public Spaces" (2023) — DOI: 10.1017/s0008197323000454 — Gradual adjustment instead of a total ban. Registry rules, space-time-context restrictions, and prior judicial authorization.
• "'Not Ready for Prime Time'" (2021) — DOI: 10.16997/book55.m — How AB-1215 was reduced from a permanent ban to 3 years. It collapsed as the rationale shifted to accuracy. Opposition must be biopolitical.
• "Journalism as Catalyst for Police Governance Reform" (2026, preprint) — DOI: 10.2139/ssrn.7308163 — January 2020 reporting on Clearview AI. Over 600 police agencies as customers in the US alone. As criticism subsides, less contentious uses are justified.
• "FRT in Policing: A National Survey" (2023) — DOI: 10.1177/00111287221150172 — Both public safety beliefs and privacy concerns influence support, with the latter varying by context.
• "The role of ideological dimensions…" (2022) — DOI: 10.1177/09636625221113131 — Crowdsourced sample. Ideology predicts acceptance, but a bias information experiment (496 people) increased opposition regardless of ideology.
• "Power and Protest" (2025, preprint) — DOI: 10.2139/ssrn.5333821 — Chilling effect on assemblies and demonstrations and the absence of oversight and accountability mechanisms. Argues for mandatory human rights duties binding corporations and governments.