Skip to content
Max Weber Programme for Postdoctoral Studies - Department of Law

Facial recognition in EU policing: A critical look, with Diya Sarkar Ghosh

In this #EUIResearch interview, Max Weber Fellow Diya Sarkar Ghosh examines the use of facial recognition technology by law enforcement authorities in the European Union, exploring its potential risks and the regulatory frameworks governing its use.

18 November 2026 | Research story

18.11.26_Web Banner Diya_EUIResearch

Facial recognition technologies are increasingly used by law enforcement agencies across the European Union. Many view them as valuable tools for identifying criminal suspects and supporting public safety. At the same time, their use has raised concerns about privacy, potential misuse, and the disproportionate surveillance of marginalised communities.

To address these concerns, Diya Sarkar Ghosh, a Max Weber Fellow at the EUI Department of Law, critically examines the use of facial recognition technology by EU law enforcement and its implications for fundamental rights. Her central argument is that the problem is not a lack of regulation, but the gaps and ambiguities in existing rules and the difficulties in enforcing them. She identifies shortcomings in the current EU regulatory framework and proposes four concrete reforms: greater transparency and independent oversight of retrospective facial recognition; tighter controls on national security exemptions; stronger judicial authorisation as well as data-erasure requirements for emergency deployments; and legally binding fundamental rights assessments at the procurement stage.

Many of us have heard concerns about facial recognition and its implications for privacy and surveillance, but we may not have a clear picture of how these technologies work. Could you walk us through how facial recognition works in practice, from capturing an image of someone's face to identifying or verifying their identity?

With facial recognition technology (FRT), we essentially look at the translation of human biology into computational data. At its core, the technology uses a camera to capture a face and extract distinct nodal points such as the distance between a person's pupils, the depth of the eye sockets, and the width of the nose. These measurements are converted into a mathematical code called a faceprint, or biometric template, that can be used for identification or verification of one’s identity.

Once generated, the facial template can serve two distinct purposes: verifying an identity or identifying an unknown person. One-to-one (1:1) verification asks a simple question: Is this the person who they claim to be? To answer it, the system compares the person's face with a reference image associated with their claimed identity, calculating how closely the two match. This is the technology that, for example, allows users to unlock their smartphones with facial recognition. One-to-many (1:N) identification asks: Who is this person? Rather than checking a face against one specific reference image, the system searches a larger database for similar faces. It produces a list of potential candidates, accompanied by similarity scores indicating how closely each face matches the unknown person. Law enforcement personnel, including investigators, can then examine these candidates to establish the person's identity.

Ultimately, it is important to understand that facial recognition technology does not produce infallible results. Rather, it relies on statistical comparisons governed by a set of human-devised parameters, making it a probabilistic tool for identification and verification, not an unquestionable determination of identity.

Your work explores the use of different types of facial recognition technology in law enforcement. What are the main differences between these technologies, and how are they used in practice?

The regulatory and technical framework governing facial recognition rests on a single critical variable: temporal immediacy. In other words, the central distinction is when facial recognition occurs. Real-time remote biometric identification (RT-RBI), also known as live facial recognition, operates as people move through publicly monitored spaces. For example, a CCTV camera at a train station can continuously scan the faces of passers-by and compare them against law-enforcement watchlists to identify potential suspects. This means that everyone entering the monitored space is subjected to facial scanning, often without their knowledge and consent, regardless of whether they are suspected of any wrongdoing. Because of its potential for indiscriminate surveillance, the EU AI Act classifies RT-RBI as an unacceptable risk and establishes a general prohibition under Article 5(1)(h) of the 2024 Act. The exemptions are narrowly defined and include locating missing persons or victims, addressing imminent threats such as terrorism, and identifying or tracking suspects of serious crimes. Deployment requires prior authorisation from a judicial or independent authority, with an emergency provision allowing for authorisation within 24 hours after deployment in exceptional law-enforcement circumstances.

Retrospective (post-RBI) analysis, by contrast, takes place after an event has occurred. Imagine, for instance, that a protest is recorded on closed-circuit television (CCTV). Instead of monitoring people in real time, investigators could subsequently analyse the recorded footage by using facial recognition software and compare faces against a database to identify potential matches. A real-world example is France's TAJ database, which catalogues approximately six million facial profiles. In practice, investigators extract an image from crime-scene surveillance footage and compare it against the database, generating a list of potential matches for human validation. Recognising its role in supporting evidence gathering and criminal investigations, the AI Act classifies post-RBI as a high-risk application, subject to rigorous compliance and governance requirements, rather than an outright ban. The TELEFI project found that, prior to 2021, 11 EU member states used retrospective facial matching as a deliberate tool for generating investigative leads rather than as an automated mechanism for arrest.

Ultimately, while real-time systems scan people as they move through public spaces, retrospective systems analyse previously recorded footage, comparing faces against databases to support criminal investigations.

Facial recognition is often presented as a tool for identifying suspects, preventing terrorism and locating missing persons. What risks and unintended consequences can arise from its use in these contexts?

It is true that FRT often serves as a "force multiplier", enabling law enforcement authorities to ensure public safety. However, the use of FRT also comes with significant risks.

One such risk is asymmetric accuracy and algorithmic bias. The accuracy of FRT is not consistent. There exists significant disparities in its accuracy across parameters like race, gender, age, and ethnicity. For example, the seminal Gender Shades paper found a trend of decreased accuracy across marginalised populations based on skin colour. In the context of law enforcement, especially during 1:N identification, the risk of wrongful identification and systemic over-policing of marginalised communities heightens. Likewise, treating imperfect forensic or biometric data as infallible proof has already contributed to discriminatory policing, as seen in Germany’s ‘Phantom of Heilbronn’ DNA investigation. Therefore, improving baseline accuracy does not mend this problem when police agencies routinely lower similarity thresholds during database queries to cast a broader investigative net and identify more potential suspects. This deliberate trade-off means accepting many false matches, potentially exposing innocents to disproportionate scrutiny and discrimination under the pretext of investigative thoroughness.

A second notable risk of post-RBI analysis is its potential to change how people engage with public life by turning everyone into a potential suspect. Because recorded footage can be analysed retrospectively, authorities can use facial recognition to identify everyone present at a protest, regardless of whether they are suspected of any wrongdoing. This reflects a broader shift from investigating specific suspects to scrutinising entire groups of people in public spaces. The prospect of being identified and monitored in this way can have a chilling effect on democratic participation: people who know they may be identified at a demonstration might think twice before attending, potentially affecting their rights to freedom of assembly, expression, and protest, as well as their right to unmonitored anonymity in public spaces. In this sense, the principles of “proportionality” and “necessity” risk being undermined, as potentially innocent people are treated as perpetual suspects whereby they are compelled to justify their lawful presence. Ultimately, lack of individualised suspicion diminishes the foundational constitutional guarantees leading to normalising biometric mass surveillance.

Your work argues that the main issue is not the absence of regulation, but the gaps and ambiguities in how existing rules are designed and enforced. What are the main weaknesses in the EU's current regulatory framework for facial recognition?

The central argument of my research is that the problem is not a lack of regulation of facial recognition technology, rather the gaps and ambiguities in existing rules and the difficulties in enforcing them. Three instruments underpin the regulation of facial recognition technology in the EU. Article 9 of the General Data Protection Regulation (GDPR) safeguards biometric data in the private domain. Article 10 of the Law Enforcement Directive (LED) governs the use and processing of such data by police, requiring it to be “strictly necessary”. Finally, the 2024 EU AI Act prohibits real-time remote biometric identification (RT-RBI) in public spaces, subject to narrow exceptions, and classifies retrospective use as high-risk.

Based on my argument, two gaps are particularly important: the regulatory divide between real-time and retrospective facial recognition, and the differences in the regulatory approaches across EU member states.

First, in practice, the legal distinction between real-time and retrospective facial recognition is deceptive. While live FRT is heavily restricted, retrospective FRT operates under lighter constraints. This creates a dangerous loophole whereby the police can claim an emergency exemption, deploy retrospective tracking without prior judicial authorisation, and take upto 48 hours to seek approval. By reviewing recorded footage after a protest or public event, law enforcement can therefore achieve the same mass-surveillance outcome that the restrictions on live facial recognition were designed to prevent.

The problem is compounded by the delayed introduction of high-risk obligations. Requirements such as fundamental rights impact assessments will not take effect until December 2027. Until then, the regulatory gap between real-time and retrospective facial recognition remains largely unaddressed, leaving the possibility of widespread retrospective surveillance without these additional safeguards.

Second, the framework allows for differences in how member states implement the rules, including through national legislation governing the use of exemptions for real-time facial recognition. Italy and Sweden, for example, have moved to incorporate these exemptions into their national laws, while other countries have adopted a more precautionary approach. This leaves room for broad interpretations of concepts such as “national security” under Article 2(3) alongside wide interpretations of public security exceptions under Article 5 – potentially allowing them to serve as justifications for extensive state surveillance. Consequently, an individual's exposure to biometric surveillance in public spaces may depend on national borders rather than a uniform European standard.

What changes do you recommend to the EU's regulatory and enforcement approach to facial recognition in law enforcement?

To ensure that police use of facial recognition technology meets the strict requirements of necessity established by the Court of Justice of the European Union (CJEU) and genuinely protects fundamental rights under the EU Charter, four structural reforms are needed.

The first priority is to make the use of retrospective facial recognition more transparent and accountable. Article 26(10) should require police forces to keep a standardised record of all requests, approvals, and refusals to use the technology. Independent Data Protection Authorities (DPAs) should be empowered to audit these records and scrutinise how the technology is used. They should also be able to issue binding orders to suspend police use of the technology when they identify repeated or systemic non-compliance.

The second priority is to close the national security exemption loophole under Article 2(3). Member states should not be able to invoke national security as a blanket justification for bypassing EU safeguards on facial recognition. Instead, the use of this exemption should be subject to independent legal scrutiny at EU level. The EU should also establish a clear and exhaustive list of serious crimes and a common definition of what constitutes an “imminent threat”, in line with the standards set by the CJEU. This would prevent individual member states from interpreting these exceptions too broadly or lowering the threshold for their use.

The third priority is to remove procedural-urgency clauses and strengthen requirements for data erasure under Articles 5(3) and 5(4). Emergency exceptions that allow law enforcement authorities to deploy facial recognition technology without prior judicial authorisation should be eliminated, with prior approval required in every case. If these exceptions are retained, authorities must seek judicial approval after deployment and face strict administrative sanctions if they fail to do so. DPAs should also verify that all data collected through unauthorised deployments is permanently erased, helping to mitigate the harm to public anonymity caused by the use of facial recognition without the required safeguards.

Finally, the principle of “Rights-by-Design” should be embedded in the procurement of facial recognition technology. Under Article 27, Fundamental Rights Impact Assessments (FRIAs) should be made legally binding at this stage. This would require police authorities to demonstrate that they have considered less intrusive alternatives, such as privacy-enhancing technologies, and explain why these were rejected. These assessments should be reviewed by independent Data Protection Authorities (DPAs) before procurement is approved. This would ensure that the authorities seeking to acquire facial recognition technology are not responsible for concluding whether its use is justified. The approach reflects the principle of nemo judex in causa sua (no one should be judge in their own cause), a fundamental safeguard against bias.

 

Diya Sarkar Ghosh is a Visiting Max Weber Fellow at the EUI’s Department of Law. Her recent research evaluates the use of facial recognition technology (FRT) by EU law enforcement, arguing that conditional exemptions risk normalising pervasive state surveillance. She also explores accountability in digital public infrastructure (DPI) and the impact of algorithmic decision-making on constitutional sovereignty. Her work has appeared in leading peer-reviewed journals covering deepfakes, the right to erasure, and digital twins. As a Visiting Fellow, she aims to deepen her research on digital constitutionalism and advance her publication projects.

Go back to top of the page