Introduction
On 1 April 2025, during a session of the House of Lords, the UK Government confirmed that trials were underway for a remote AI weapons detection (AWD) system designed to identify concealed weapons such as knives from a distance. In response to a question about knife crime and stop and search, Lord Hanson of Flint, Minister of State for the Home Office, stated that the Home Office, working with industry partners, is developing ‘… the technology to detect from a distance knives concealed on the person …’. In giving more light to this assertion, Baroness Doocey claimed that the results were so far ‘excellent’ and would hopefully ‘eliminate the need for stop and search, as weapons hidden under clothing will be visible and the police will not have to do anything’.Footnote 1 This initiative comes at a time of renewed political and legislative focus on violent crime, particularly knife crime. According to the Office for National Statistics (ONS), police in England and Wales recorded 54,587 offences involving knives or sharp instruments in the year ending December 2024, a 2% increase from the previous year.Footnote 2
In 2018, the UK Government held a competition urging proposals for technological advancements that could detect individuals carrying knives.Footnote 3 Policymakers have also argued that stop and search powers are essential for tackling such violence.Footnote 4 Simultaneously, new legislation such as the Terrorism (Protection of Premises) Act passed on 3 April 2025, widely known as Martyn’s Law,Footnote 5 places duties on public venues to assess and mitigate the risk of terrorist attacks.Footnote 6 It is within this confluence of legislative pressure and technological promise that AWD systems are being advanced as tools for public protection.
Remote weapons detection represents a subtype of algorithmically mediated surveillance. We use the term ‘surveillance’ because these tools can be deployed, overtly or covertly, to monitor, observe or scan systematically from a distance at a group population level (which might be referred to as bulk or mass surveillance), as well as such monitoring being directed, or targeted, at particular individuals because of a specific reason or investigation. As the Constitution Committee explained:
A literal definition of surveillance as ‘watching over’ indicates monitoring the behaviour of persons, objects, or systems. However, surveillance is not only a visual process which involves looking at people and things. Surveillance can be undertaken in a wide range of ways involving a variety of technologies [including] sensing devices, body scans, technology for tracking movement, and many others.Footnote 7
Unlike conventional metal detectors, which function through direct physical proximity and the detection of metallic signatures, AWD systems typically integrate multiple non-invasive technologies – such as millimetre-wave imaging, radio frequency sensors, magnetometers, and AI-based pattern recognitionFootnote 8 – to identify threats at a standoff distance.Footnote 9
Proponents of AWD highlight its ability to distinguish between genuine threats and innocuous everyday items, pinpointing the precise location of potential weapons on a person or within their belongings in real-time.Footnote 10 While some systems primarily focus on detecting ‘brandished weapons’ – those visibly held in hand – to minimise false alarms,Footnote 11 other advanced AI models, particularly those integrating Convolutional Neural Network (CNN)-based architectures and terahertz imaging, demonstrate potential for identifying more complex improvised and concealed weapons.Footnote 12 The claimed advantages extend to proactive threat detection, significantly accelerated response times (with alerts generated in mere seconds), and a reduction in false positives through continuous learning and contextual analysis.Footnote 13
However, recent history cautions against taking such claims at face value. The experience of Evolv Technologies in the United States illustrates how similar systems may not deliver on their promise and can introduce new forms of risk. Evolv’s AI-powered scanners, deployed in schools, stadiums, and theme parks, were marketed as ‘touchless’ and ‘frictionless’ security solutions.Footnote 14 However, investigative reporting by the BBCFootnote 15 revealed major reliability issues, including high false positive rates that triggered unnecessary secondary searches,Footnote 16 and worrying false negatives where real weapons, including knives and bombs, passed undetected.Footnote 17 As an example, in Utica, New York, a teenager was repeatedly stabbed by another student in a school that had installed Evolv Express – Evolv’s AI-based weapons detection system – only months earlier.Footnote 18 The system had failed to detect the hunting knife, notwithstanding Evolv’s marketing claims that it had ‘detected and stopped over 30 thousand guns and 27 thousand knives’ from entering its customers’ venues.Footnote 19 As well as in many large stadiums and schoolsFootnote 20 across the US, the BBC further revealed that Evolv’s scanners are used in the UK, despite not having been tested by the UK Government’s National Protective Security Authority.Footnote 21 As of January 2024, the Federal Trade Commission (FTC) was taking action against Evolv technologies for alleged false advertising and misleading claims of product accuracy regarding the products’ ability to detect weapons by using AI.Footnote 22 Questions about reliability have continued to surface. In October 2025, an AWD system developed by Omnilert triggered an alert in a Baltimore school after misidentifying a teenager’s packet of crisps as a firearm.Footnote 23 These examples highlight the risks of deploying high-stakes, proprietary AI systems without adequate independent scrutiny. Furthermore, factors such as the dynamic nature of crowded environments, varying environmental conditions, and the limitations of current AI technologies contribute to the complexity. Even established static detection systems, such as those deployed in airport security, encounter difficulties in reliably identifying objects.Footnote 24 Transitioning to AI-powered remote detection is likely to amplify these challenges, potentially leading to increased instances of misattribution and subsequent public distrust.
Fundamentally, outputs from these tools are probabilistic – an alert or ‘hit’, on the face of it definitive, is in fact an indication that a combination of measurements has reached a threshold, ie certain measurements have a certain level of similarity to the measurements produced by a particular weapon profile that the system has been trained on. The probabilistic nature of AWD is a crucial factor in the considerations discussed below regarding whether such outputs could satisfy the legal test justifying police interference with liberty, particularly where there are more than minimal false positive rates in operational use.
In addition to these technical concerns, the legal and ethical implications of a new mode of surveillance-led, AI-powered scanning and detection tool are not negligible. First, they introduce ambiguity about whether a remote (and possibly covert) scan constitutes a ‘search’ in law. Secondly, the uncertainty of the tool’s output creates confusion as to whether the legal threshold for stop and search, or other police action, is met, and makes it unclear whether procedural safeguards apply. Thirdly, they raise concerns about indirect discrimination, particularly if the AI system is more likely to flag individuals based on group characteristics (for example, clothing styles, physical gait, or objects or medical devices carried or used by those with disabilities). Fourthly, these systems, similar to other technologies such as facial recognition, raise questions about the proportionality in human rights terms of new forms of group surveillance and the sensitive scanning under clothing, and the implications for the policing-by-consent model in England and Wales.
In the discussion that follows, this paper is structured around five key questions. Before turning to these questions, Section 1 provides the necessary foundation by briefly discussing how AWD systems operate. In Section 2, we address the first two questions: whether AI-generated alerts satisfy the legal requirement of ‘reasonable grounds for suspicion’ that is necessary for a lawful stop and search and whether the act of remote scanning itself constitute a ‘search’ in law. Here, we start by examining the legal foundations and safeguards of stop and search under PACE and associated legislation before proceeding to draw out the nature of these AI outputs and whether they can be assimilated to the standard of individualised, objective suspicion required by PACE and the common law. We then consider whether a scan that reveals concealed objects is functionally equivalent to a physical search, and the significance of this classification for legal safeguards. Section 3 explores these implications under equality law and police procedural duties. This includes the public sector equality duty (PSED) and the risk of bias or disproportionate impact on protected groups, as well as the duties under PACE Code A (such as the requirement to inform a person of the grounds of a search). Section 4 focuses on how this technology engages surveillance law and the right to privacy under Article 8 of the European Convention on Human Rights (the Convention). Here, we assess whether population-level, preemptive scanning amounts to an interference with private life and, if so, whether it can be justified under Article 8(2). We also reflect on broader issues of public consent and the Peelian principle of policing by consent in the context of an algorithmically enhanced policing model. In the concluding section, we suggest potential legal and policy reforms that would be required to align remote weapons detection with principles of legality, accountability and non-discrimination.
1. Understanding AWD systems
(a) The sensing technologies
AWD systems use several distinct methods to detect concealed objects. Millimetre-wave sensing, one of the extensively studied modes of concealed weapons detection, exploits high-frequency electromagnetic waves that readily penetrate common clothing material and are reflected from the human body and any concealed items.Footnote 25 These systems rely on two primary methods to achieve this: passive imaging, which simply receives naturally occurring radiation to detect the differences in emissivity and reflection between human skin and hidden objects; and active imaging, which acts like radar by transmitting its own millimetre-wave energy and analysing the returning echoes.Footnote 26 A second family of technologies, radar-based systems, transmit radio frequency signals to analyse the returned echoes by relying on the fact that these radio waves pass through the thin layers of clothes, and are reflected highly by the body, plus any hidden objects between the two.Footnote 27 A third approach uses magnetometer-based walk-through systems to detect metallic weapons by inducing and measuring secondary magnetic fields, the spatial behaviour of which is determined by a target’s ‘electrical conductivity, magnetic permeability, shape, and size’.Footnote 28 Meanwhile, infrared thermography systems rely on natural heat emissions, capturing the temperature gradient between the concealed object area and the clothing surface that occurs when body heat is absorbed by a hidden weapon before radiating through the clothing.Footnote 29
(b) What is being scanned
At their core, AWD systems’ primary sensing is directed at objects on or near the body to serve as a ‘discrete, privacy-preserving modality’.Footnote 30 In the case of infrared scanning, for example, Khor et al noted that ‘images of faces captured using passive thermography are difficult to recognise, thus providing an additional layer of privacy’.Footnote 31 AWD systems are object-focused in that they ask whether a body is carrying something that matches the profile of a weapon, not who the person is. That said, the distinction is not entirely clean. While the primary sensors do not capture facial images, several systems incorporate conventional cameras as part of their alert workflow which presupposes that even though the detection process may not involve faces, the response process often does.Footnote 32 AWD systems scan every person who enters a detection zone, continuously and without differentiation. This potentially creates a two-stage detection architecture in which the first stage automatically screens everyone in an area, and the second stage subjects flagged individuals to a ‘threat isolating or removal procedure’ on the basis of an algorithmic output.Footnote 33 Moreover, several sensing methods, even though they are not identifying individuals by face, inherently interact with the body itself, not merely with objects upon it.Footnote 34 Some AWD systems additionally analyse how a person moves, hinging this on the likelihood that a person carrying a concealed weapon may walk slightly differently, and that these changes can be detected using, for instance, radar-based motion analysis.Footnote 35
(c) What output is produced
The feature that most clearly distinguishes AWD from a conventional metal detector is the use of machine learning algorithms (typically, deep CNNs) – trained on datasets containing examples of both weapons and everyday items – to interpret sensor data. The algorithm learns to identify patterns in the sensor data that tend to distinguish one from the other and calculates a probabilistic ‘confidence score’Footnote 36 to estimate the likelihood of a weapon’s presence, frequently filtering out anomalies with low confidence levels.Footnote 37 The system does not make a yes-or-no determination that a weapon is present, rather, it calculates a probability. Kotter et al described this in the context of magnetometer-based concealed weapon detection as algorithms that ‘perform pattern recognition and calculate the probability that the collected magnetic signature correlates to a known database of weapon versus non-weapon responses’.Footnote 38 The same logic underpins modern AWD systems. Typically, operators apply an adjustable threshold to the confidence score to discriminate between actual threat items and benign personal belongings; however, this creates a direct and unavoidable trade-off that can increase the false alarm level or let genuine threats slip through.Footnote 39 The internal reasoning of the algorithms is not readily interpretable by human operators, a problem widely recognised as the ‘black box’ problem.Footnote 40 This matters because a police officer acting on an AWD alert may be unable to explain why the system flagged one person and not another. This is fundamentally different from the cognitive process by which an officer forms reasonable suspicion on the basis of observation and experience,Footnote 41 where the assessment can in principle be articulated and scrutinised.
2. Stop and search, legal safeguards and remote weapons detection
(a) Brief history of stop and search legislation in England and Wales
The history of stop and search in British policing can be traced back to the Vagrancy Act 1824, which was referred to as ‘sus law’ due to the powers in the legislation to search and arrest suspected individuals.Footnote 42 In particular, section 4 allowed the policing of individuals who were suspected of ‘committing offences to be deemed rogues and vagabonds’.Footnote 43 This legislation became a significant source of conflict between the police and the public, particularly due to its disproportionate application against Black individuals. In 1980, the Commission for Racial Equality (predecessor to the Equality and Human Rights Commission) published a report urging the UK government to amend or repeal section 4 of the Vagrancy Act due to the disproportionate number of arrests of Black individuals, particularly by London police forces. It was reported that 43.5% of the Metropolitan police force’s ‘sus law arrest[s]’ were of Black individuals:
We recognise that this is largely a London problem, because the ‘sus’ law is little used by most police forces elsewhere. There are widespread and genuine feelings, particularly among the minority communities, that some police officers abuse the law by ‘picking on’ young blacks.Footnote 44
When the repeal of the sus law was discussed in Parliament, many were of the opinion that a repeal without adequate replacement would leave a significant gap in criminal law.Footnote 45 Nonetheless, the sus law was repealed following the Brixton riots in 1981, which were driven by rising conflicts between the police and Black communities in London.Footnote 46 Following this repeal, PACE was enacted to replace a confusing patchwork of archaic lawsFootnote 47 with a single, comprehensive statute and its associated Codes of Practice. The main aim of the framework was to standardise and professionalise police work across England and Wales,Footnote 48 carefully balancing the powers of the police against the rights and freedoms of the public. This legislative intent sought to control police behaviour, provide clear statutory rights to individuals, and protect privacy. To achieve this aim, PACE contains three main concepts: the notion of reasonable suspicion;Footnote 49 the regulation of use of force by the police;Footnote 50 and the regulation of police behaviour when it comes to collection of evidence.Footnote 51 In the context of stop and search powers of the police,Footnote 52 the first concept is pertinent for our discussions in this article.
(b) Key stop and search provisions under PACE
Section 1 of PACE confers on a police officer the power to search a person or vehicle in a public place for certain prohibited items (stolen property, offensive weapons, etc) which would otherwise amount to assault, but only if the officer has reasonable grounds for suspicion that such items will be found.Footnote 53 This statutory requirement reflects a fundamental balance that seeks to empower the police to tackle crime while protecting individuals from arbitrary intrusion.Footnote 54 It operationalises the underlying ‘policing by consent’ model built on the Peelian principles,Footnote 55 that the authority of the police relies upon the common consent of the public as opposed to the power of the state.Footnote 56 The requirement of reasonable suspicion is meant to ensure that physical intrusions on personal liberty and privacy are justified by concrete and objective factors, not hunch or bias.Footnote 57 Code A of PACE includes guidance to ‘make clear what constitutes “reasonable grounds for suspicion”’Footnote 58 and establishes a two-part test in this regard. First, the officer must hold a genuine suspicion that they will find the object for which the power allows a search. Secondly, that suspicion must be reasonable, meaning it must rest on an objective basis derived from facts, information, or intelligence relevant to the likelihood of finding the object, such that a reasonable person would reach the same conclusion.Footnote 59 Information or intelligence can include reports from the public describing a person carrying a relevant article, or committing a crime in which such an article would constitute relevant evidence.Footnote 60 Suspicion may also be based on reliable intelligence that certain gang members habitually carry knives and can be identified by distinctive clothing or symbols,Footnote 61 or based on a person’s behaviour, for example trying to hide something at night.Footnote 62 Searches should be conducted using ‘[e]very reasonable effort … to minimise the embarrassment that a person being searched may experience’.Footnote 63
In general, there is no power to require a person to remove any item of clothing other than an outer coat, jacket or gloves,Footnote 64 but an officer may feel inside pockets of outer clothing, collars, socks or shoes if reasonably necessary to search for the item in question.Footnote 65 Any more thorough searches, involving the removal of further clothing, must be done out of the public view and by an officer of the same sex.Footnote 66 The Code further emphasises that personal factors can never be the basis of reasonable suspicion (unless there is a description of a person carrying articles for which there is a power to stop and search). Such factors would include a person’s physical appearance with respect to protected characteristics such as age, disability or race, or ‘[g]eneralisations or stereotypical images that certain groups or categories of people are more likely to be involved in criminal activity’.Footnote 67
From the foregoing, several key principles can be distilled. Stop and search must be conducted with courtesy, consideration, and respect, minimising embarrassment and observing limits on the removal of clothing. Reasonable suspicion must be genuine, grounded in relevant facts, information, or intelligence, and linked to articles covered by the power. It cannot rest on generalisations or stereotypes. It should be noted that under section 60 of the Criminal Justice and Public Order Act 1994 (CJPOA 1994), discussed below, an authorisation may be put in place to permit police officers to stop and search anyone in a specific area without the requirement to have reasonable grounds, if there is a belief that serious violence may take place in that area or that offensive weapons will be carried. The authorisation must be limited in both time and place.
(c) AI outputs and the legal standard of ‘reasonable grounds’ for stop and search
Prior to conducting a stop and search under section 1 of PACE, an officer must have genuine suspicion that is reasonable for suspecting that they will find a specific object. The question here is whether the output from an AWD system can meet this standard. In respect of algorithmic individual risk assessment tools, Oswald has argued that:
[p]robabilistic outputs of risk or harm based on comparison with other people in the past cannot … satisfy the requirement for reasonable grounds, as they would fall within the exclusions of generalisations, category-based suspicion, and suspicion based on general association …Footnote 68
Berman (in the context of the US probable cause standard) explains the dangers of relying on probabilistic analytics to justify a search for drugs: a model ‘will identify correlations and indicate that a certain combination of characteristics, when present, predict at some rate of statistical probability that an individual is a marijuana user’. It will not be based on a causal claim or anything unique to that individual.Footnote 69
While acknowledging that all evidence is probabilistic to some degree, Berman argues that the ‘individualized suspicion’ requirement serves fundamental values including ‘safeguarding human dignity, preserving individual autonomy, and guaranteeing procedural justice’.Footnote 70 An AI’s probabilistic output, derived from identifying correlations in data, ‘denies the procedural outlets for factual explanation and legal justification that are so critical to fulfilling the individualized-suspicion requirement’s purposes’.Footnote 71
Although algorithmic risk assessment tools are structurally different in terms of function and data analysis to AWD, both produce probabilistic outputs. This inherent probabilistic nature creates an immediate tension with the legal requirement that suspicion must relate to the likely discovery of a specific item. In critiquing the act of stop-and-searching an individual solely based on an algorithmic generated similar score, Kotsoglou reflected that ‘[s]cientific propositions including algorithmic output which is necessarily statistical in nature refer to a target system (reference class) and not directly to an individual case or event’.Footnote 72 As previously highlighted,Footnote 73 AWD generates a probabilistic match or ‘hit’ which indicates that certain measurements have a certain level of similarity to the measurements produced by a particular weapon profile that the system has been trained on.
Since AWD is a probabilistic statistical method, its outputs are associated with a degree of inherent mathematical uncertainty.Footnote 74 They provide a likelihood that something is the case, devoid of a narrative of causation or direct evidence. In other words, they tell us that ‘this person has X% chance of carrying a weapon because a shape detected under clothing or other combination of measurements is X% similar to the measurements of the system’s training data’ not ‘this person is carrying a weapon because [specific fact related to the individual or context]’. Viewed through that lens, accepting an AWD alert – on its own – as sufficient for reasonable suspicion could lower the legal standard to a probabilistic calculus. This is problematic in a liberal legal order where police interference with liberty must be justified by articulable reasons, and not by generalisations.Footnote 75 As Kotsoglou and Oswald observe, for instance, a match by an automated facial recognition tool – which is analogous in being a probabilistic algorithmic output –
may on the face of it appear objective and sufficient, but this is belied by the probabilistic nature of the output … raising questions as to the justifiability of regarding the tool’s output as an ‘objective’ ground for reasonable suspicion.Footnote 76
A critical aspect of the law on reasonable suspicion is the continuing role of the human officer. As Lord Justice Purchas highlighted in Castorina: ‘suspicion must arise from reasonable cause. Reasonable cause, it is not disputed, is to be determined as an objective matter from the information available to the arresting officer and cannot have anything to do with the subjective state of the officer’s mind’.Footnote 77 That suspicion may end up being ‘ill-founded [but this] does not in itself necessarily establish that the police officer’s suspicion was unfounded’,Footnote 78 demonstrating the strong support and respect which the officer’s suspicion possesses.Footnote 79 While the meaning of ‘objective basis’Footnote 80 has not been precisely defined, case law confirms that reasonable suspicion must rest on sufficient factual, intelligence or evidential grounds. Information from external sources may support an officer’s belief, but that belief must remain rooted in the suspect’s actual behaviour and capable of objective justification.Footnote 81 As PACE Code A makes clear, a mere ‘hunch or instinct which cannot be explained or justified to an objective observer can never amount to reasonable grounds’.Footnote 82 What is ‘reasonable’ depends on the circumstances and the existence of information that would satisfy an objective observer.Footnote 83 In this regard, another important consideration relates to the influence of ‘automation bias’ on an officer’s reasonable suspicion. Automation bias is a type of cognitive bias in which individuals lean towards the automated output rather than human judgement.Footnote 84 In particular, human-decision making could be influenced by ‘automated decision aids’.Footnote 85 In the context of AWD, this could be the situation of an officer being influenced by the automated output to conduct the searches in a particular location or stop and search certain individuals, trusting the automated output over their own belief.
Could the output from AWD then satisfy the requirement for an officer to hold such reasonable belief? Although such belief may be partially informed from other sources of information, it cannot arise from a wholly unrelated matter.Footnote 86 This raises questions about whether technological alerts or the use of a scanner to determine the possession of weaponry could be appropriately regarded as an alternative source of information or whether reliance on that information aloneFootnote 87 would erode the constitutional requirement for an officer’s individual judgement.Footnote 88 In effect, the AI should assist the officer’s judgement, not replace it. The courts have repeatedly reinforced this human element. In Castorina v Chief Constable of Surrey,Footnote 89 the lawfulness of arrest was held to depend on the officer’s state of mind,Footnote 90 assessed against the modifiedFootnote 91 Wednesbury Footnote 92 standard of reasonableness, which requires the belief to rest on an objective evidential basis. In O’Hara v The Chief Constable of the Royal Ulster Constabulary,Footnote 93 information from a senior officer’s briefing was capable of creating reasonable suspicion, but the arresting officer has to form an independent belief. Consequently, if a senior officer trained in scanning technology identifies a suspect, a junior officer may only lawfully act if they are of the same independent belief in terms of reasonable suspicion.Footnote 94 Case law thus underscores that reasonable suspicion is an evidence-based and individual assessment, not a blanket or automated determination.Footnote 95 Given the already low threshold for reasonable suspicion,Footnote 96 over-reliance on machine findings risks undermining the independent focus on the individual that the law requires. A ‘hit’ on an AWD system may effectively become the reason for a stop and search, absent any other individual grounds for believing that a person is carrying a hidden weapon, and despite factors that suggest the contrary, for example age or disability. Therefore, in order to mitigate the risk of automation bias or overreliance on AWD output, police officers must as a minimum be required to articulate and record their reasons for suspicion, including how they have considered the AWD output in light of other consistent or contradictory information. Such an assessment requires a systematic process beyond a simple self-assessment, perhaps through an external check conducted by a senior officer.
(d) Does AWD constitute a ‘search’?
If an AI alert on its own is a shaky foundation for ‘reasonable grounds’, the question then arises as to whether a scan that reveals concealed objects is functionally equivalent to a physical search, and the significance of this classification for legal safeguards. As a starting point, it is important to first determine the meaning of a search under existing legal frameworks. PACE does not provide a single, explicit statutory definition of ‘search’ within its text. Instead, PACE sets out the circumstances, powers, and procedures under which police may conduct searches of persons, vehicles, and premises.Footnote 97 Similarly, Code A does not define a ‘search’. However, it provides contextual information from which to draw an inference of the nature of searches when officers are exercising their stop and search powers. All stops and searches must be carried out with ‘consideration and respect’Footnote 98 and with every effort to minimise embarrassment, observing the limitations on the power to require removal of clothing. The duration and extent of any search must be kept to a minimum and be directly proportional to the suspicion.Footnote 99 While officers can place hands inside pockets and feel around collars or socks, any more thorough search requiring the removal of other clothing must be conducted out of public view, such as in a police van or station, and by an officer of the same sex.Footnote 100
The most invasive searches involving the exposure of intimate parts are subject to even stricter protocols and are forbidden from being a routine extension of a less thorough search.Footnote 101 In all of these, what is clear is that a search must be carried out in accordance with the law; however, case law suggests that the use of AWD systems by police may interfere with this obligation. The concern arises because the deployment of AWD systems could, in effect, constitute a search of individuals before officers have formed a reasonable suspicion.
In Gillan and Quinton v the United Kingdom,Footnote 102 the European Court of Human Rights (ECtHR) rejected the Government’s argument in relation to suspicionless search under terrorism powers that a ‘superficial’ search of outer clothing, pockets, shoes and hair did not engage Article 8 of the Convention. The Court considered that powers that required a person to submit to a ‘detailed search of his person, his clothing and his personal belongings’ was a clear interference.Footnote 103
Although AWD does not involve the physical removal of clothing, its capacity to detect objects ‘hidden under clothing’Footnote 104 or in bags makes an analogy to the searches described in Gillan apt. Although the exact data analysis methodologies of AWD often remain opaque both to the user and to the individual being scanned, sensors may, as already noted,Footnote 105 indicate shapes of potential weapons against body shape in addition to movement and behavioural analysis, with scanning occurring in public. Several AWD systems inherently interact with the body itself, processing data derived from the electromagnetic or thermal interaction with the body surface. Whether or not the system produces a visible body image, the underlying process involves sensing of the body. In Roberts v Commissioner of Police of the Metropolis & Others Footnote 106 concerning stop and search for concealed knives, the Court warned that: ‘[a]ny random “suspicionless” power of stop and search carries with it the risk that it will be used in an arbitrary or discriminatory manner in individual cases’. Yet, it also accepted that the deterrent effect of unpredictability could serve the legitimate aim of preventing serious violence.Footnote 107 This ambivalence suggests that, while AWD might be defended as being deployed for a legitimate aim, it does not resolve the question of whether the technology itself performs a ‘search’ before any physical stop takes place. Moreover, the ECtHR in Gillan acknowledged that:
[T]he discomfort of having personal information [including body image] exposed to public view, might even in certain cases compound the seriousness of the interference [with the right to privacy] because of an element of humiliation and embarrassment.Footnote 108
Here, the court also rejected comparisons with airport or building entry searches, because a person ‘may be seen as consenting to such a search by choosing to travel. They know that they and their bags are liable to be searched before boarding the aeroplane and have a freedom of choice, since they can leave personal items behind and walk away without being subjected to a search’.Footnote 109 On this reasoning, a search implies the physical process (ie involving direct interaction with a person) following the formation of suspicion. However, AWD differs from conventional stop and search in its lack of physical contact, and the deployment potentially before any individualised suspicion and normally without awareness of the individual concerned. If the technology can perceive shapes and materials beneath clothing and prompt police intervention, it raises the same legal and ethical issues that PACE was designed to regulate. The capacity to ‘see’ under clothing without consent, at a population level, effectively performs an informational pat-down that is functionally equivalent to a physical search.
An analogy can be drawn with the debate around whether the ‘sniff’ of a drug detection dog should be defined as a search.Footnote 110 Marks highlights an inconsistency between attempting to deny that a sniff is a search (because there is no physical touching) while simultaneously attempting to link the individual to the odour to provide legal grounds for a subsequent physical search.Footnote 111 She argues against the approach that would view an indication from a sniffer dog as always constituting reasonable grounds as failing to take into account concerns over a dog’s accuracy and discernment.Footnote 112 It could be said to be similarly inconsistent to deny that the ‘digital sniff’ of a AWD system is a search, yet attempt to rely on the probabilistic output to justify physical intrusion.
Furthermore, AWD could be regarded as a continuous, non-consensual form of suspicionless search and should therefore be regulated as such, rather than merely as an intelligence-gathering tool. It is pertinent to distinguish between suspicion-based and suspicionless searches. In contrast to suspicion-based searches, section 60 of the CJPOA 1994Footnote 113 provides a power for police officers to stop and search individuals without needing reasonable grounds for suspicion (ie ‘suspiciousless search’). The primary purpose of section 60 is to prevent serious violence and the widespread carrying of weapons, particularly when normal stop and search powers based on reasonable suspicion would not be sufficient.Footnote 114 Although section 11 of the Public Order Act 2023 also introduced powers to stop and search without suspicion, section 60 of the CJPOA 1994 directly refers to stop and search powers relating to offensive weapons and dangerous instruments. Given the stated goal of AWD – to ‘detect concealed weapons before a violent incident takes place or recover weapons used in a recent violent incident’Footnote 115 – it appears conceptually aligned with section 60’s anticipatory and preventative aims, and arguably should be subject to similar explicit authorisation and safeguards. For instance, the authorisation given has a time limit – initially lasting up to 24 hours but can be extended up to 48 hours.Footnote 116 However, a scenario where AWD scanners are permanently deployed by police officers, say via mobile devices, would far exceed the contemplated use of section 60. This would effectively create a de facto continuous section 60 regime, which then redefines the interaction from ‘stop and search’ to ‘search and stop’, where the search (by the AWD) occurs first, and the human intervention (stop) follows an alert. As discussed below,Footnote 117 this transformation has profound implications for civil liberties and the rule of law.
3. Equality, bias, and procedural fairness
(a) Algorithmic bias impact
AWD is likely to raise significant equality concerns, particularly regarding its disproportionate impact on marginalised groups.Footnote 118 Other surveillance technologies, such as facial recognition, have already brought these issues to the fore. These risks are largely linked to algorithmic bias, which can arise from biased training data, leading to disproportionate and negative effects on already marginalised groups.Footnote 119 As Broussard explains, ‘We don’t live in a perfect world, so the data that we’re using to train AI systems is data from our imperfect world … then the AI perpetuates these biases’.Footnote 120 A prime example is facial recognition technology (FRT), which has been shown to be less accurate on people with darker skin.Footnote 121 As digital technology evolved, many ‘inherited configurations [were] optimized for light skin’, with the result that cameras today often still fail to capture the features of darker-skinned individuals clearly, even under good lightning conditions.Footnote 122 Similarly, in the context of AWD, devices used, for instance, by persons with disabilities could be erroneously assessed as a weapon,Footnote 123 as has been the case with traditional sensor-based systems such as those in airports. The risk of such errors, as previously noted, is compounded where AWD systems incorporate gait or behavioural analysis. This is because bias can occur when the dataset used to train the algorithm does not represent the population it intends to serve or contains unintentional bias which could lead to inaccurate or unfair results.Footnote 124 Persons with physical disabilities or different body types may exhibit movement patterns that diverge from the norm in the system’s training data. This weakness is worsened by AI’s limited capacity for contextual understanding and the possibility that repeated errors may be reinforced through machine learning.Footnote 125 If an AWD system generates an alert, and that alert is the sole basis for an officer’s suspicion leading to a physical search, it directly conflicts with the principle of individualised suspicion and raises significant human rights concerns regarding non-discrimination and the presumption of innocence.Footnote 126
The Equality Act 2010 prohibits both direct and indirect discrimination in the exercise of public functions, including policing, and imposes a PSED.Footnote 127 This duty requires public authorities to have due regard to the need to eliminate discrimination, advance equality of opportunity, and foster good relations between persons of different characteristics. This duty is non-delegable and must be ‘exercised in substance, with rigour, and with an open mind’.Footnote 128
The court in Bridges noted that the PSED is a duty of ‘process and not outcome’ but is important for ensuring accountability.Footnote 129 It requires a public body to give thought to the impact of a policy and its potential disproportionate effect on a certain community, and thus the risk of indirect discrimination. The public body must take ‘reasonable steps to make enquiries about what may not yet be known to a public authority about the potential impact of a proposed decision or policy on people with the relevant characteristics’.Footnote 130 In the Bridges case, it was found that the police had not done everything reasonable to fulfil the PSED – they had not satisfied themselves through independent verification that the live facial recognition tool did not have an unacceptable bias on the grounds of sex or race, neither had they obtained access to the training datasets to assess their demographic composition.Footnote 131 The ‘human failsafe’Footnote 132 argument – that a human must decide whether to act on the facial recognition match – was not sufficient to discharge the PSED, as the police had not satisfied themselves as to the possible bias within the tool they were using. Furthermore, ‘human beings can also make mistakes’.Footnote 133
The court emphasised that automated facial recognition is a ‘novel and controversial technology’ and so everything reasonable should be done to assure against racial or gender bias.Footnote 134 AWD is similarly novel and controversial and raises parallel concerns about bias and indirect discrimination. These systems, particularly those reliant on pattern recognition algorithms, may inadvertently reproduce or exacerbate discrimination through technical bias. As previously highlighted, individuals with medically necessary itemsFootnote 135 such as insulin pumps, prosthetic limbs, or mobility aids may face disproportionate and erroneous flagging by AWD systems due in particular to imbalance in subjects and objects represented in training datasets. If such outcomes result in unnecessary scrutiny, interventions, and potentially humiliating experiences during searches,Footnote 136 this could amount to indirect discrimination under section 19 of the Equality ActFootnote 137 and to a breach of Article 14 of the Convention due to discrimination in the enjoyment of a Convention right. The discriminatory consequences, while perhaps unintended, could reflect structural flaws in the technology that affect particular groups more severely than others. The outcome of the judgment in Bridges highlights the importance given to the duty on public bodies to take all reasonable steps to investigate and avoid bias in their novel technologies. Withholding relevant details due to reasons of commercial confidentiality, while ‘understandable’, will not allow a public body to discharge its duty.Footnote 138
(b) Procedural fairness and transparency
Beyond equality issues, AWD-led detection raises questions about procedural fairness and compliance with police duties to provide information during stop and search encounters. Under PACE Code A,Footnote 139 before conducting a search, officers are required to inform individuals of their identity, the legal power being exercised, the purpose of the search, and the grounds on which it is based. These safeguards are crucial for transparency – they allow the individual to understand why they are being stopped and to later challenge the stop if it appears unlawful.Footnote 140 However, when an AI system provides the basis for a stop, meeting these obligations becomes problematic if the system has not been built to provide the type of information and explanations that the law requires. If the only ground is an algorithmic output that the officer cannot further elucidate, the explanation will be vague. ‘The system flagged an anomaly’ is not the kind of plain, concrete reason that Code A contemplates, suggesting rather a machine-generated ‘hunch’. This risks undermining the procedural justice of the encounter, potentially therefore adversely affecting perception of the police as legitimate and trustworthy. Furthermore, the reversal of ‘stop and search’ to ‘search and stop’ brought about by AWD could mean that duties to provide information fall by the wayside – indeed, if a decision is made not to engage or stop following a scan by AWD (or this was practically not possible), an individual may never know that a ‘search’ has taken place, even though a record may be kept of the search and the conclusions drawn about the individual, perhaps combined with a facial image. The collection, retention and use of such nominal records and conclusions will raise issues about the criteria being used for such collection, and the safeguards in place for retention and eventual deletion.Footnote 141
4. Surveillance, privacy, and public consent
Having examined suspicion and search definitions (Section 2) and issues of equality, bias, and discrimination (Section 3), we now turn to the broader regulatory context of surveillance and privacy – namely, the implications of AWD within existing surveillance law, and human rights requirements, particularly Article 8 of the Convention.
(a) Privacy concerns
The right to respect for private life under Article 8 of the Convention protects individuals from unlawful or arbitrary interferences with their personal life. To represent a justified interference with Article 8, three criteria must apply: the interference must be in accordance with the law; pursue a legitimate aim; and be necessary and proportionate in a democratic society.Footnote 142 Mass surveillance, even without explicit facial recognition, is considered to infringe on privacy if it is not necessary for reaching a legitimate aim and proportionate to the end sought.Footnote 143 In a stop and search context, the ECtHR has held that there is a zone of interaction between a person and others, even in a public context, which may fall within the scope of ‘private life’.Footnote 144 Relatedly, it has been affirmed that powers granting an individual the right to stop and search a person ‘anywhere and at any time’ and to subject them to a detailed search of person, clothing and belongings constituted interference with the right to respect for private life.Footnote 145 Indeed, there is a growing body of case law from the ECtHR pertaining to privacy rights in connection to law enforcement activity and mass surveillance efforts.Footnote 146 Although the court has not yet directly ruled on AWD systems, its reasoning in other contexts is highly instructive. It has held that the right to private life is a broad and evolving concept, ‘not susceptible to exhaustive definition’.Footnote 147 In Sciacca v Italy,Footnote 148 the ECtHR found a violation of Article 8 owing to the publication of a photograph taken by public authorities of the applicant and distributed without adequate legal basis.Footnote 149 Although the case concerned image dissemination rather than scanning, the reasoning underscores the Court’s recognition of the individual’s right over their own imageFootnote 150 and the notion that bodily appearance and its public exposure may fall within the scope of private life. AWD systems such as Evolv Express and Cambridge Terahertz actively scan individuals for concealed objects before any human officer intervenes. This means that a search of a person’s concealed belongings is effectively being performed by a machine, without individualised suspicion, simply by their presence in a public space. Since their operation involves a form of automated observation that may blur the boundary between lawful monitoring and intrusive surveillance, it becomes pertinent to address the further question of how such technologies should be classified within the existing regulatory framework.
(b) Overt or covert surveillance
The potential for AWD systems to operate covertly – for example where individuals are unaware of scanners or through the use of drones – raises questions about their regulation under the Regulation of Investigatory Powers Act 2000 (RIPA 2000). This concern is particularly relevant where the technology functions similarly to FRT, allowing individuals to be scanned without their knowledge. Covert deployment by police forces would not be unprecedented. The UK Home Office has previously sought proposals for technologies capable of detecting knives in ‘unobtrusive and potentially covert’ ways,Footnote 151 and recent developments in drone-based knife detectionFootnote 152 suggests an increasing movement towards discreet, even hidden, automated surveillance.
RIPA 2000 governs covert and directed surveillance – defined as surveillance conducted so that subjects remain unaware, such as following someone from a distance.Footnote 153 AWD systems are often described as non-intrusive, and assertions are made that these systems do not use facial recognition or track personal identifiers,Footnote 154 and that millimetre wave imaging only produces a generic body-shaped outline, avoiding exposure of intimate details or the need for clothing removal.Footnote 155 From this perspective, the initial operation of an AWD system might be viewed as a general, automated observation comparable to overt CCTV, which does not usually require a directed surveillance authorisation under RIPA 2000 unless it becomes covert or targeted.Footnote 156 However, once an AWD system detects a potential weapon and transmits a real-time alertFootnote 157 identifying a specific location and visual of an individual, the process arguably transitions from general observation to the contemplation of an individualised stop and search. Generalised data analysis by public authorities, such as mapping crime hotspots, is not usually treated as directed surveillance because its focus on individuals is considered ‘sufficiently cursory’; yet once analysis begins to concentrate on specific persons or groups, authorisation becomes appropriate.Footnote 158 The ECtHR has affirmed that stop and search powers, even under anti-terrorism legislation, constitute a clear interference with the right to respect for private life.Footnote 159 Such intrusion concerns not only the collection of data but also the experience of personal exposure, discomfort, and potential humiliation, even where the initial processing is automated.Footnote 160
The government’s 2021 Surveillance Camera Code of Practice reaffirms support for overt surveillance in public spaces where it pursues a legitimate aim, meets a pressing social need, and remains proportionate. It emphasises that ‘… it is the way in which technology is used that is potentially intrusive rather than the technology itself’.Footnote 161 Once reasonable steps have been taken to inform the public that surveillance is in operation, such activity ordinarily does not require directed surveillance authorisation.Footnote 162 By contrast, covert surveillance, defined as being carried out in a manner calculated to ensure subjects are unaware it is taking place, requires specific authorisation where it is ‘directed’ or ‘intrusive’ and likely to result in obtaining private information about a person.Footnote 163 Accordingly, AWD systems may fall below the threshold for directed surveillance at the point of initial scanning, if no individual is targeted or identified and no personal data are stored. The threshold is arguably crossed only when a potential threat is detected, and human intervention is contemplated. However, this would be an unsatisfactory state of affairs, leaving the public unclear about how the deployment of AWD would be authorised and overseen.
The Bridges v South Wales Police case is also instructive. The Court of Appeal found that automated facial recognition, by capturing and processing images of members of the public (the majority of no interest to the police) to extract sensitive biometric data in an automated way, was a novel technology, and not equivalent to CCTV, as the police had argued. The legal framework currently in place was insufficient for the purposes of Article 8(2) and thus the use of automated facial recognition by the police was not ‘in accordance with the law’.Footnote 164 The deficiencies of the law related to the questions of where the technology could be used and who could be placed on the watchlist, both of which left too much discretion to the police.Footnote 165 Although AWD systems do not rely on biometric identification, they are novel forms of technology performing an automated, individualised ‘informational pat-down’ from a distance, intrusive (in the general sense) in their own way and with the clear potential to be operated covertly and in a directed manner. Reading across from the Bridges decision, in the absence of statutory law or clear policies in this area, the use of AWD systems would also raise issues regarding the ‘where’ and ‘who’ questions (in terms of who would be targeted by such systems and the limits of their use). Furthermore, the judgment in Peck v UK Footnote 166 further illustrates that the right to privacy extends to a person’s bodily image as a whole, not merely to facial or distinguishing features. The UK GDPR definition of personal data – ‘any information relation to an identified or identifiable natural person’Footnote 167 – supports such an interpretation, even where identification is not actively pursued.Footnote 168 Moreover, as Peck demonstrates, the detection of a weapon does not invariably signify criminal intent. Law enforcement authorities must consider the relevant context, such as self-harm, as was the case in Peck. Likewise, the reasoning in Malone v United Kingdom,Footnote 169 where the absence of clarity regarding surveillance powers was itself found to breach Article 8 of the Convention, underscores the continuing importance of legal precision in the governance of AWD systems.
The above dilemmas – the boundaries of directed surveillance; whether AWD is a search or can provide grounds for a search – suggest that we should urgently revisit our understanding of key legal terms and principles within relevant legal frameworks. AWD presents a form of remote digital noticing and, arguably, searching. The necessity of physical proximity is removed, thus representing a fundamental challenge to longstanding legislative safeguards. Wilson, Clayton and Rowe point out that the act of being noticed or observed by the police (even non-digitally) can disrupt an individual’s movement through their space, tantamount to a stop – the question then becomes who (and what) is being made visible by police action.Footnote 170 Focusing too narrowly on legally defined ‘stop and search’ can mean that other significant actions are ignored.Footnote 171 Digital remote observation, scanning and searching by AWD represents a new and significant challenge to the effectiveness of traditional safeguards that are predicated on physical intrusion, with consequent risks to trust in the police, as we discuss in the next sub-section.
(c) Erosion of policing by consent
Perhaps the most fundamental implication of AWD systems is the potential for erosion of public consent as the basis for policing. The British model of policing is founded on the ‘Peelian’ principles which emphasise that the legitimacy of the police depends on the approval and cooperation of the public.Footnote 172 Traditional stop and search, for all its flaws, takes place within a visible and accountable framework, where officers must identify themselves, state their legal authority, provide reasons for the stop, with rights for individuals to receive records.Footnote 173 By contrast, remote scanning and AWD-driven surveillance may occur invisibly, with no clear initiation point and no opportunity to challenge or even recognise that a scan has occurred. The person may not know they have been flagged until they are stopped, or not at all if no action is taken (and despite scanning occurring and a record potentially being made of the encounter in both circumstances). This lack of transparency undermines the visibility and contestability of police powers, replacing human judgement with machine inference and public interaction with procedural obscurity. Worse still, if such systems are deployed in high-footfall areas (such as shopping centres, stadiums, transport hubs) without clear signage or opt-outs, the public may come to view them as a condition of access to public space. The net effect is to normalise passive surveillance and to reshape the relationship between the citizen and the state, moving towards a society where individuals are constantly analysed.
Another important consideration is how susceptible AWD would be to resistance from the public, including attempts to bypass the scanning. There are known threats to similar technologies; for example, in the case of facial recognition, there are known vulnerabilities, including through 3D masks,Footnote 174 anti-surveillance fashionFootnote 175 and of course human-avoidance tactics. Police drones have been susceptible to signal blocking, jamming and spoofing, which has allowed the technology to be controlled by non-police personnel.Footnote 176 Such vulnerabilities could also raise issues for the implementation of AWD.
Furthermore, trust in policing is becoming increasingly strained.Footnote 177 Surveys from YouGov show a decline in public confidence, with the proportion of people who have ‘not very much confidence or no confidence at all’ in the police to deal with crime rising from an average of 38% in 2020 to 53% by April 2023.Footnote 178 The Independent Office for Police Conduct has warned that misuse of stop and search damages public confidence, particularly among ethnic minority communities.Footnote 179 The deployment of automated systems that may exacerbate disproportionality and reduce transparency risks compounding this legitimacy crisis. If public trust continues to erode, the effectiveness of these technologies may be undermined, as individuals and communities become less willing to cooperate with law enforcement, potentially reducing the overall efficacy of policing efforts.
Conclusion
AWD systems represent a paradigm shift in policing – one that reverses the usual sequence of reasonable suspicion followed by search, effectively introducing a model of ‘search (surveillance) first, stop later’. This approach risks undermining fundamental safeguards that have long governed police powers and physical intrusion. Our analysis has highlighted the inherent tensions between this technology and the current legal framework in England and Wales. AWD is both a remote digital search, and its outputs used to justify a further physical search. Yet an AWD alert that is statistical rather than based on concrete facts, and generalised rather than particularised, does not provide the kind of specific, articulable facts about an individual that current legal frameworks demand for a search on reasonable grounds. It is, at best, a statistical probability of something. Accepting such alerts as sufficient grounds could dilute the reasonable suspicion standard and effectively automate police judgement. Admittedly, AI outputs may be useful as one factor in decision-making, but they cannot, by themselves, fulfil the legal standard for a lawful stop and search. Ensuring an officer’s independent assessment remains central is crucial; otherwise, the safeguard of human discretion (and accountability for it) is lost.
The analysis in this paper has shown that the introduction of these systems will engage the PSED, raising, inter alia, concerns about indirect discrimination. We noted, for example, that individuals with disabilities or those from minority ethnic communities might experience higher false positive rates. Without rigorous equality impact assessments and ongoing monitoring, the supposed neutrality of the algorithm could mask systemic biases. Furthermore, the opacity of AWD-driven suspicion has the effect of undercutting procedural fairness. PACE Code A requires officers to explain the reason for a stop and search to the individual and later in the record. Arguably, a response based on a system flag without further information or explanation provides no meaningful opportunity for the person to understand or challenge the basis of their stop. This inscrutability risks turning the encounter into what feels like an arbitrary imposition – the very scenario PACE’s safeguards were designed to avoid. From a surveillance law perspective, AWD systems clearly fall within the ambit of Article 8 of the Convention (right to respect for private life), requiring rights interference to be in accordance with the law, pursue a legitimate aim, and be necessary and proportionate in a democratic society. If deployed as currently contemplated, their use would lack a clear legal basis and robust safeguards. The current statutory regime governing covert and directed surveillance provides us with little certainty over how AWD would be defined or governed.
We predict that AWD will be incorporated and combined with other remote and digital surveillance, monitoring and analysis policing tools – body-worn video, drones, facial recognition and behavioural analysis AI, to name but a few. The Home Office’s consultation on a new legal framework for law enforcement use of biometrics, facial recognition and similar technologies recognises the need for a specific legal framework for these technologies.Footnote 180 But a significant opportunity will be missed if a new legal framework does not cover covert and overt AWD capabilities and other remote, sensing and scanning technologies, whether or not involving biometrics, with the aim of achieving a clear understanding of the probabilistic nature of these technologies and therefore consistency of regulatory approach. If AWD tools are to be used fairly and proportionately (whether on their own or in combination with other tools), certain reforms should be considered. At a minimum, consideration could be given to restricting the deployment of these systems to contexts where exceptional suspicionless search powers already apply, such as under section 60 of the CJPOA 1994, albeit still requiring enhanced safeguards. In effect, the deployment of AWD systems should be treated with the gravity of a special policing operation, not routine patrol. Conceivably, there is some potential for these systems to reduce discriminatory personal interactions as they could assist the police in upholding Peelian principles by minimising intrusive contact and ensuring objective decision-making. However, this possibility depends largely on the integrity and accuracy of the systems, the quality of oversight, and the ability of the public to understand and contest their use. Conversely, such technologies could make policing more opaque and unaccountable if their deployment is not accompanied by robust safeguards, including meaningful information rights, clear testing and audit trails, and enforceable legal standards. Deployments should require pre-authorisation by an independent regulator and the use of pre-certified technology. Diluting the informational requirements of stop and search would significantly weaken accountability mechanisms and widen the existing trust deficits between police and certain communities.
Policing by consent in the twenty-first century demands that the introduction of novel, controversial technologies be done with public input.Footnote 181 The ongoing Government consultation gives us a chance to examine the operation of AWD and similar tools, to define their functionality and understand how they observe, scan and search – absent physical proximity – and to develop a legal and regulatory regime that genuinely and robustly responds to the implications for our society. This highlights the need for further public engagement and advice on such intrusive technologies,Footnote 182 particularly among groups already disproportionately affected by stop and search practices. Without public engagement and a clear legal foundation, these technologies risk deepening existing inequalities and weakening the legitimacy of law enforcement.
Open access
