Misidentification & facial recognition technology

Wrongful arrest

On 25th February 2026 the Guardian reported on the wrongful arrest of Alvi Choudhury in my home city of Southampton, for a burglary committed in a city he had never visited.

The basis of the arrest appears to be an incorrect ‘identification’ from a retrospective facial recognition (RFR) search.

Over the last few years there has been a steady stream of articles about similar wrongful arrests in the US (here & here). But the incident with Alvi Choudhury is perhaps the first widespread media reporting of a wrongful arrest in the UK based on an erroneous RFR search. This is significant moment in the UK and it is critical that lessons are learnt on how such errors can be avoided in the future.

There has already been much discourse on the performance of FRT algorithms used by the police, and a pervasive argument in some corners that the technology is not ‘accurate enough’. Here, I want to take a different perspective. The technology is as accurate as it is and this should be established through independent testing. To ensure the safe and responsible use of FRT it is incumbent upon policing to know the limitations of the technology and ensure FRT is operated by sufficiently trained and competent operators, following documented policies with effective quality assurance.

In this blogpost, I will explain how human factors have a huge part to play in ensuring RFR results are accurate, and how effective procedures and processes can be used to mitigate errors. In particular, I will focus on the critically overlooked issue of how RFR is used by human operators, and what impact the human operator has on the performance of the FRT system.

So, what is RFR, how is it being used by police in the UK, and how did its use lead to a wrongful arrest? Read on to learn more.

Facial recognition technology in policing

Facial recognition technology (FRT) uses computer algorithms trained on very large datasets of facial images. The algorithms detect and encode faces, then compare these encodings, resulting in a score of how similar or dissimilar the two faces are. How reliable this score is will depend on the accuracy of the algorithm and the extent of its training data, as well as the quality and types of facial images being compared.

The accuracy of FRT algorithms has increased rapidly in recent years, due to technological advances in computing power and access to massive facial image datasets. However, FRT is not infallible. Accuracy rates decline with lower quality images and performance varies across different demographic groups. These two factors are closely related and low quality images can serve to further exacerbate demographic effects. The operation of FRT is typically done under human supervision.

In the England & Wales, the police utilise FRT in three ways:

• Retrospective facial recognition (RFR): All police forces in England & Wales use RFR to search images of unknown persons of interest (the probe image) against large databases of known persons (also known as the gallery) to attempt to identify the person. The gallery typically contains mugshot custody images, whereas probe images could be sourced from anywhere (e.g. social media, CCTV or identity documents). The FRT system compares the probe to every image in the gallery and ranks the results by similarity score. The system will then present a ranked list of potential candidates to a human operator for review. The candidate list may be a fixed size (e.g. the top twenty) or candidates with a similarity score above a preconfigured threshold.

• Live facial recognition (LFR): The use of FRT to overtly identify individuals based on their face in public places. Faces of passersby are scanned by a camera and then compared using FRT to a watchlist of wanted individuals. If the similarity score exceeds a preconfigured threshold the FRT system issues an alert, which is verified by a human operator before any further action is taken.

• Operator initiated facial recognition (OIFR): Where police officers use FRT on a mobile device to identify a person they have come into contact with.

LFR has received the most attention in the UK media and is considered particularly controversial by privacy groups like Liberty and Big Brother Watch. However, RFR is the most common use of FRT by police forces, with over 20,000 searches of the national RFR system undertaken every month.

All police forces in England & Wales have access to the facial search capability within the Police National Database (PND), which was introduced in 2013. Some forces also have access to local FRT systems to carry out RFR on specific databases.

A recent public attitudes survey undertaken by the Home Office found that two thirds of individuals were supportive of the police use of FRT, and of the three use cases RFR was seen as the most acceptable (97%). However, cases like the mistaken arrest of Alvi Choudhury based on an incorrect RFR search are likely to negatively errode the public’s perception of trust in FRT.

Some could argue that given the technology has been in use for over 12 years, with tens of thousands of searches completed every month, this was just a freak occurence. But this would provide little reassurance to an individual who has been wrongfully arrested in their home, in front of their family. Although such errors are not commonplace (the actual rates of wrongful arrest from RFR are not known) the impact on individuals is significant.

There is an extensive literature of international guidance, good practice and research on the safe and effective use of facial recognition technology for RFR in policing, from groups such as the European Network for Forensic Science Institutes, the Facial Identification Scientific Working Group, the Biometrics Institute and most recenty, the International Forensic Science Alliance, as well as in academic publications. Despite this wealth of useful information, there is currently no national guidance or policy in England & Wales that dictates how the police should undertake and action RFR searches.

The case of Alvi Choudhury is a clear example of what happens when FRT systems are operated without effective policies to safeguard their use.

Human factors in RFR

In response to Alvi Choudhury’s wrongful arrest, a Thames Valley Police spokesperson stated:

“While we apologise for the distress caused to the complainant in this case, their arrest was based on the investigating officers’ own visual assessment that the individual matched the suspect in CCTV footage following a retrospective facial recognition match, and was not influenced by racial profiling.”

This statement implies that a primary cause of the erroneous identification was an investigator’s visual assessment of the facial images. For anyone who has researched human ability in facial identification, the fact that this is problematic will not come as a surprise.

Whilst we are, on the whole, very good at recognising faces of people we know, there is a well established body of scientific research demonstrating that we are not very good at identifying faces that we don’t know. As summarised in my 2021 written evidence to the House of Lords Justice and Home Affairs Committee on FRT:

“Humans are surprisingly poor at comparing faces of people we do not know, with large individual differences in ability. This is in stark contrast to the relative ease by which we can recognise people who are familiar to us (see [11] for a thorough explanation on the differences between familiar and unfamiliar faces).

Even when comparing faces in images taken on the same day, people can be mistaken, on average, one-fifth of the time [12]. Human performance at comparing unfamiliar faces further declines when images are taken under suboptimal conditions, e.g. images from CCTV systems [13]. Poor face-matching performance has also been observed for some professional groups, including passport officers, where years of experience bared no correlation to performance [14], and operators of facial recognition technology [15]. There are however some people who demonstrate consistently superior face-matching ability, namely ‘super-recognisers’ who are people with a high natural face recognition ability [16], and trained forensic examiners [17].”

Performance of a human operator can have a significant impact on the accuracy and reliability of results from a face recognition system [15]. The interaction between the algorithm and the human operator and its impact on accuracy is often overlooked when evaluating the performance of face recognition technology [18].”

Further complicating the issue is the fact that many of us think we are far better at identifying faces than we are (this is a well known cognitive bias called the Dunning-Kruger effect).

Demographic effects are also not isolated to FRT algorithms. Demographic biases have been found in human face-matching ability, the most well researched being the other-race effect, where people are worse at recognising and matching faces of a different ethnicity to their own.

This all has real world implications for RFR use, with a 2015 study finding that human performance in face matching can curtail the accuracy of FRT systems. Moving forward to 2024 and, paradoxically, as FRT systems become more accurate research found that humans actually got worse at reviewing RFR results, as the algorithms present more convincing lookalikes.

International guidance, such as the ENFSI Guideline for Facial Recognition System End Users are unequivocal that people running RFR searches should be:

“specialist FR operators that have received formalized training and should be able to demonstrate ongoing competency and proficiency in FR review”

Further international guidance from FISWG outlines the requirements for the training and mentorship of specialist FR operators.

This is even alluded to in a recent guide on FRT from the Home Office, published in December 2025:

“The images are compared against the Police National Database, a library shared by police forces of the custody photos of known people. The system then produces a list of the most similar images and these images are reviewed by a specially trained operator who decides whether there are any matches.”

But in England & Wales, any police officer with access to the police national database can feasibly run RFR searches. There are no national training requirements, officers do not have to demonstrate proficiency in reviewing RFR candidate lists and there is no requirement to follow any particular process when processing RFR searches. Compare this to fingerprints, another biometric widely used in policing. Fingerprints are compared by specially trained examiners, following documented processes subject to quality assurance by peer review. Why isn’t this the case for RFR?

Reporting RFR results

Concerningly, the Thames Vally Police spokesperson claims that the officers’ visual assessment of the images was the basis of the arrest. This contradicts a widely held international consensus that RFR results are an investigative lead and not, in of themselves, a basis for arrest. According to ENFSI:

“When a potential candidate is reported, it is recommended that the following information is provided:

A potential candidate does not indicate a positive identification of the individual, and a summary of the reasons why.

Important decisions such as fundamental rights limitations (for example arrests, crimes imputation, freedom of movement, etc.) should not be adopted based exclusively on the potential candidate reported.

The investigative lead report is not intended as evidence in court proceedings. The main purpose for the investigative lead report is to provide criminal investigation with intelligence.”

Given the inherent risks and limitations of RFR results, potential candidates should be corroborated and this should be reflected in relevant policies for the reporting of RFR results.

Mitigating errors

In the case of Alvi Choudhury, it is not clear whether the investigating officer undertaking the RFR search had received any training, followed a specific process or had their findings reviewed. Regardless, it is clear that there were insufficient safeguards in place to detect their mistaken decision.

Implementing effective quality management practices in RFR, such as training and competency testing, standard operating procedures and quality assurance would provide effective mitigation of errors for the human review of RFR results and align with international guidance and good practice.

According to the Guardian:

“The Home Office said guidance and training to minimise error and maintain public confidence in retrospective facial recognition was under review by the Police Inspectorate. It said a new national facial matching system is under development, with an improved, independently tested algorithm.”

These are welcome but long overdue measures and present a much needed step towards the professionalisation and standardisation of FRT use across police forces in England & Wales. I iterated this in a response to the recent Home Office consultation on a new framework for law enforcement use of biometrics, facial recognition and similar technologies:

“we do consider it necessary to ensure that RFR is operated within an effective quality management system, whereby searches are documented and recorded, follow standard operating procedures, are conducted by trained and competent operators and undergo quality control and assurance. We believe that having a quality managed RFR service would be an ideal fit for the proposed National Police Service, mirroring models used internationally (e.g. FBI CJIS, Norwegian Criminal Investigation Service, Spanish National Police, Swedish National Forensic Centre etc.).”

But with no timescales for the delivery of a national framework for FRT use, and with over 20,000 searches being conducted a month nationally, there are still significant risks of further errors occuring. Mitigations should be put in place as soon as possible, at the very least ensuring that police forces on a local level adhere to recommended minimum requirements for FRT use, including:

• Effective training and, where relevant, testing of FRT users.

• Detailed policies and procedures on how FRT is to be used.

• A clear and documented understanding of the risks associated with FRT.

• Reporting requirements for FRT results.

Doing so will provide effective measures to minimise the risk of further error and maintain public trust and confidence in the police use of FRT.