Facial recognition surveillance: the debate and what it means for privacy

01The San Francisco facial recognition reversal

In May 2019, San Francisco became the first U.S. city to ban government use of facial recognition technology. The Stop Secret Surveillance Ordinance prohibited city departments from acquiring or using facial recognition technology, citing concerns about civil liberties, racial bias, and the erosion of anonymity in public spaces. The ban was hailed as a landmark by privacy advocates and criticized by law enforcement as shortsighted.

In 2024, the San Francisco Board of Supervisors voted to partially reverse the ban, allowing police to use facial recognition technology for certain investigations, particularly violent crimes and cases involving children. The reversal was driven by high-profile cases where facial recognition was credited with solving crimes that traditional methods could not. The debate in San Francisco encapsulates the national tension: the technology offers real law enforcement benefits but poses real civil liberties risks.

The reversal set a precedent. Other cities that had banned facial recognition are reconsidering their positions. The political dynamics are shifting as public concern about crime intersects with privacy concerns, and as the technology improves in accuracy and reduces demographic bias. The San Francisco case demonstrates that facial recognition policy is not settled law but an evolving debate.

02How facial recognition technology works

A facial recognition system is a technology potentially capable of matching a human face from a digital image or a video frame against a database of faces. Such a system is typically employed to authenticate users through ID verification services, and works by pinpointing and measuring facial features from a given image.

Modern facial recognition systems use deep learning to extract a mathematical representation of a face — a faceprint — from an image or video frame. The system identifies key facial landmarks such as the distance between eyes, the shape of the cheekbones, and the contour of the jaw, then converts these measurements into a numerical vector. This vector is compared against a database of known faces to find matches. The process takes milliseconds, enabling real-time identification from surveillance camera feeds.

The accuracy of facial recognition has improved dramatically with deep learning. Early systems struggled with varying lighting conditions, angles, and expressions. Modern systems can identify faces from partial views, low-resolution images, and even masks or sunglasses. However, accuracy depends heavily on the quality of the input image and the diversity of the training data. A system trained predominantly on one demographic will perform worse on others — the root cause of the racial bias problem.

03Law enforcement use and abuse cases

Law enforcement agencies use facial recognition for two primary purposes: identifying suspects in criminal investigations and monitoring public spaces for persons of interest. The technology has been credited with solving cold cases, identifying fugitives, and locating missing persons. The FBI’s Facial Analysis, Comparison, and Evaluation Services unit processes over 100,000 facial recognition searches annually, accessing a database of over 640 million images including driver’s licenses and passport photos.

However, facial recognition has also produced wrongful arrests. In at least three documented cases in the United States, men were wrongly arrested based on incorrect facial recognition matches. In each case, the arrested individual was Black — the demographic for which the technology is least accurate. These cases demonstrate the danger of treating algorithmic matches as definitive evidence rather than investigative leads. They also highlight the legal due process gap: there are no established standards for how facial recognition evidence should be presented in court, what accuracy thresholds must be met, or what disclosure obligations prosecutors have.

04The privacy and civil liberties concerns

Mass surveillance is the intricate surveillance of an entire or a substantial fraction of a population in order to monitor that group of citizens. The surveillance is often carried out by local and federal governments or governmental organizations, but it may also be carried out by corporations. Depending on each nation's laws and judicial systems, the legality of and the permission required to engage in mass surveillance varies. It is the single most indicative distinguishing trait of totalitarian regimes. It is often distinguished from targeted surveillance.

The civil liberties concerns around facial recognition extend beyond individual misidentification. The technology enables mass surveillance at a scale that was previously impossible. A network of cameras equipped with facial recognition can track the movements of every person in a city, building a database of where individuals go, who they meet, and what they attend. This is not speculative — it is the current reality in parts of China, where facial recognition is integrated into an extensive surveillance system that monitors the Uyghur population in Xinjiang and tracks citizens through their daily routines.

In democratic societies, the concern is that facial recognition chills constitutionally protected activities. People attending protests, political rallies, or religious services may be identified and tracked. The Supreme Court has recognized a right to anonymity in certain contexts, and mass facial recognition surveillance could erode that right. The American Civil Liberties Union and Electronic Frontier Foundation have filed lawsuits challenging government use of facial recognition without adequate privacy protections or public oversight.

05Racial bias in facial recognition systems

Biometrics are body measurements and calculations related to human characteristics and features. Biometric authentication is used in computer science as a form of identification and access control. It is also used to identify individuals in groups that are under surveillance.

The racial bias in facial recognition is well-documented. The National Institute of Standards and Technology (NIST) conducted the most comprehensive study to date, testing 189 algorithms from 99 developers across demographic groups. The study found that African American females had the highest false positive rates — up to 34 times higher than white males in some algorithms. This means a Black woman is far more likely to be incorrectly matched to someone in a database, leading to false identifications and potential wrongful arrests.

The bias has multiple causes: training datasets that are disproportionately white and male, algorithmic design choices that optimize for average accuracy rather than demographic equity, and the fundamental challenge of measuring facial features on darker skin tones where contrast is lower. Some vendors have improved their algorithms, but the NIST study showed that even the best-performing systems exhibited demographic differentials. The problem is not that the technology is inherently racist but that it inherits and amplifies the biases present in its training data and design choices.

06Which cities and countries are banning or adopting it

The global landscape of facial recognition policy is diverging. On the restrictive side, the European Union’s AI Act, which took effect in 2024, bans certain uses of facial recognition including real-time biometric identification in public spaces, with exceptions for law enforcement in narrowly defined circumstances. San Francisco, Portland, Baltimore, New Orleans, and several other U.S. cities have banned or restricted government use. India considered a ban but has instead deployed one of the world’s largest facial recognition systems.

On the adoption side, China leads the world in facial recognition deployment. The technology is integrated into the country’s social credit system, used for everything from boarding subway trains to paying for goods to monitoring attendance at schools and workplaces. Russia has deployed facial recognition at protests, leading to arrests of demonstrators identified through camera networks. The United Kingdom has one of the highest densities of surveillance cameras in the world and uses facial recognition for policing, though courts have ruled some uses unlawful.

The United States lacks a comprehensive federal facial recognition law. The ACLU and other advocates have pushed for federal legislation, but Congress has not acted. This leaves regulation to a patchwork of state and city laws, creating inconsistent protections depending on where a person lives.

07What regulation should look like

Effective facial recognition regulation must address several dimensions. First, accuracy standards — systems should be independently tested for demographic performance before deployment, and agencies should be required to use only systems that meet minimum accuracy thresholds across all demographic groups. Second, use limitations — the technology should be restricted to serious criminal investigations, not routine surveillance or minor offenses. Third, transparency — agencies should publicly disclose when and how they use facial recognition, and individuals should be notified when they have been matched.

Fourth, due process — facial recognition matches should be treated as investigative leads requiring corroboration, not as probable cause for arrest. Fifth, data governance — the databases used for matching should have strict access controls, retention limits, and audit trails. The EU AI Act represents the most comprehensive attempt to regulate facial recognition, but even it has been criticized for its exceptions and enforcement gaps.

The fundamental question is whether facial recognition should be permitted at all in public spaces, or whether the risks to privacy, civil liberties, and democratic participation outweigh the law enforcement benefits. The San Francisco reversal suggests that the answer is not permanent — policy will continue to evolve as the technology improves and as its benefits and costs become clearer through real-world deployment.