Interoperable and standardized algorithmic images: The domestic war on drugs and mugshots within facial recognition technologies

Abstract

Beginning in the 1990s, the National Institute of Standards and Technology (NIST) leveraged the 1980s’ American War on Drugs to improve and expand facial recognition technology (FRT) infrastructure, including the domestic building of FRTs reliant on mugshots. When examining mugshot databases gathered by the NIST, such as the Multiple Encounters Dataset (MEDS) I and II (2010) and Special Database 18 Mugshot Identification Database (SD-18) (2016), it is clear that the same gendered and racialized dynamics present in policing practices related to the War on Drugs is reflected in the mugshot databases that continue to use for FRT research and evaluation into the contemporary moment. This paper details the SD-18 and MEDS databases, as well as the MORPH database, showcasing how their representational, technical and political protocols operate. The desires for frictionless interoperability built into the images’ technical protocols supersede concerns for eugenic political and representational protocols, resulting in a current moment where the deployment of mugshot datasets cannot be contained to their original intended use with FRTs, but leak into other forms of algorithmic governance as well as into algorithmic image-making and visual culture, including generative artificial intelligence systems such as DALL-E.

Introduction

The 1997 National Institute for Standards and Technology (NIST) report ‘Best Practice Recommendation for the Capture of Mugshots’, built from the 1995 Mugshot and Facial Image Workshop that was held in Gaithersburg, MD, lays out an interlocking set of protocols ‘intended as a means of establishing or improving interoperability between mugshot systems’ (1997: 1). Such recommendations not only include best practices for capturing lens-based images such as pose, depth of field, centering and lighting but also include digitally oriented protocols such as the ideal aspect ratio, minimum number of pixels, pixel aspect ratio and file format. In the same year, the NIST also released ‘Data Format for the Interchange of Fingerprint, Facial & SMT Information’ that similarly concerns itself for the best practices for the preservation and circulation of digitized biometrics, including mugshots (1997). This second report, written in conversation with the Gaithersburg workshop, was similarly put in place to ensure the standardization of digital mugshots such that they could be stored, transferred between other systems and adapted into future systems with frictionless ease. Taking the two documents and workshop together, the most effective mugshot-enabled facial recognition technology (FRT) was seen, as it still is in 2024, as one that simultaneously concerned itself with both traditional image-making (pose, lighting etc) and digital protocols (storage, image resolution, file format), wherein mugshots became both image and data and, in doing so, became interoperable between different FRTs.

Over the same time period, the NIST was in the process of building the Face Recognition Technology database (FERET; 1993–96), the first large-scale facial database that was used to improve the state-of-the-art of FRTs, which was then utilized in developing the Face Recognition Vendor Tests (FRVTs; 2000-) that would become the American governmental benchmarking infrastructure for FRTs for the decades to come. Established in 1901, the NIST has long been involved with the development and standard setting of an incredible array of technologies within America, including the world's first automated fingerprint-matching algorithm in the early 1960s (NIST, 2023). Importantly, as I wrote in the article ‘Interdiction, the 1980s War on Drugs, and Building Future Infrastructure for Facial Recognition Technologies’, the FERET database and FRVTs were constructed in reaction to the American War on Drugs, and involved the Defence Counterdrug Development Program, the Office of National Drug Control Policy (formed in 1988), and the Counterdrug Technology Assessment Center (formed in 1990) (Tucker, 2022). While FERET deployed portraits of volunteers gathered at George Mason University and the Army Research Laboratory in Adelphi, Maryland, the resultant FRVTs have made consistent use of mugshots over the past decades into the contemporary moment: as one recent example, the 2020 FRVT that evaluated different FRTs’ identification capabilities utilized four datasets, with the primary dataset being composed of law enforcement mugshots images ‘comprised of 26.6 million reasonably well-controlled live portrait photos of 12.3 million individuals’ (Grother et al., 2020: 2;6). In the interest of furthering improvements in FRTs, two NIST mugshots databases have been made publicly available: the Multiple Encounters Dataset (MEDS) I and II and Special Database 18 Mugshot Identification Database (SD-18), both of which contain facial data likely gathered during the 1980s and 1990s.

As Brian Jordan Jefferson outlines in Digitize and Punish, many of the digital policing technologies that are ubiquitous in 2024 ‘were initially rolled out in the late 1980s, in the thick of the War on Drugs’ (2020: 2). Within digital policing technologies, including biometric systems like FRTs, ‘digital databases, not detention centers, jails, or prisons, are the leading edge of criminal justice in the United States’ where ‘racial identities are not anachronisms but logical elements’ (ibid.: 1;6). The use of mugshots as interoperable image-data objects as found in the 1990s examples in the opening paragraph of this article are a crucial historical point in the development of FRTs and are established from materials built in the shadow of, if not in direct relation to, the American War on Drugs.

Such data is laced with gendered and racialized dynamics that carry the international and domestic effects of the War on Drugs forward into the contemporary moment. The targeting of racialized and gendered populations by the War on Drugs parallels the well-established critiques of algorithmic governance as leveraging the affective bodies of such populations for the improvement and deployment of biometrics such as FRTs, and for reaffirming precarity and instability within (to name only four areas): the education, health care, banking and real estate systems (Eubanks, 2018; O’Neill, 2016; Noble, 2018; Broussard, 2023; Amaro, 2022; Benjamin, Race After Technology, 2019; Chun, 2021). In such instances, Simone Browne centres ‘digital epidermalization’, where ‘the exercise of power cast by the disembodied gaze of certain surveillance technologies…[is] employed to do the work of alienating the subject by producing a truth about the racial body and one's identity (or identities) despite the subject's claims’ (2015: 110); within algorithmic governance, such truths label certain populations, via their facial data, as financial risks, as bad employees, as criminal or as any of a number/combination of politically volatile categories. Within the contemporary moment of algorithmic governance and surveillance, understanding facial data, in particular mugshots, means grappling with the face as a long-time site of political power, wherein facial data must be treated differently than other forms of data that are not burdened with the face's long history of being used as a marker of racial and gendered identity that is then used to harmfully sort and hierarchize individuals and populations.

Such dynamics are far from new, as evidenced, for example, by Browne's work tracing biometric and modern surveillant logics of Black people to nineteenth-century lantern laws, branding and slave ships (ibid.). Knowing this, Valdivia and Tazzioli stress the urgent need to actively historicize datafication such that it is not considered only within processes of digitization, and so that ‘critiques of racial categories in algorithmic fairness [move] forward by considering datafication as an historical mechanism that also reproduces racial profiling’ (840); this also includes, as Shoshana Magnet writes extensively on in When Biometrics Fail, biometrics that are imbedded within prison systems, with the prison industrial complex offering up its prisoners as new products for biometric companies (2011: 62–63). For mugshots specifically, it is important to wrestle with such images and image-making given their deep roots in securitization and law enforcement that goes back to the nineteenth century through figures such as Francis Galton (1869; 1910), Alphonse Bertillon (1885; 1908) and Cesare Lombroso (1876). For Lombroso, the measuring of the body and facial characteristics in particular was a central part to his racist atavist arguments denigrating of non-white populations. For Galton, his Composite Profiles of criminalized faces were core to his racist beliefs, wherein the deployment of facial data was in service of larger eugenic forces he associated with the health and security of the nation state. From the nineteenth century onwards, these harmful logics were meant to be applied as a mass technology and be interoperable between systems of governance: advocates of Bertillion's signaletics argued that while signaletics was originally built around criminality and proto-mugshots, ‘in order for society to reap [the system's] full benefit, every human being should be partially signalized’ (1885: viii); the authors go on to argue that not only could such a system take the place of passports and other identifying documents but also it would be incorporated into documents such as life insurance policies and permits.

A century later, with the War on Drugs as an important landmark in gathering such facial data and establishing key protocols within FRTs, the past calls for signaletic interoperability, haunted by eugenics, echo loudly in a present moment where facial data slips across the many contexts that FRTs are deployed. As Ruha Benjamin argues, ‘From credit-scoring algorithms to workplace monitoring systems, novel techniques and devices are shown to routinely build upon and deepen inequality. Racist and classist forms of social control, in this sense, are not limited to obvious forms of incarceration and punishment; rather, they entail what sociologist Carla Shedd calls a “carceral continuum” that scales over prison walls’. (‘Introduction’ 2019: 2). In the contemporary moment, interoperability ensures that policing strategies can combine mugshots with expanded access and deploy facial data across a wide variety of fields, many of which were never intended for law enforcement purposes: for example, in the wake of protests over the murder of George Floyd, the Minneapolis Police Department created fake and/or covert social media accounts to illegally surveil citizens, part of which was compiling datasets that could then be deployed in on-the-street policing practices (Ryan-Mosley and Richards, 2024).

Utilizing a close analysis of the aforementioned MEDS I and II (2010) and SD-18 (2016) mugshot databases produced and circulated by the NIST, alongside the MORPH database (2006), this paper details the specific logics and images within the SD-18, MEDS and MORPH facial databases and how their representational, technical and political protocols operate in order to understand how mugshots specifically have contributed to the history of automated FRTs within the twenty-first century. However, as will be seen in the examination of the SD-18, MEDS and MORPH mugshot datasets, such circulations of harmful data are too often subsumed within the concern for technical protocols, wherein the desires for frictionless interoperability are prioritized over recognizing and grappling with dangerous political and representational protocols. As the conclusion will detail, these interoperable images do not remain contained to mugshot datasets, but leak into other forms of image-making and visual culture, including generative artificial intelligence (AI) systems such as DALL-E.

Face recognition is a face problem, not just a vision problem

It is perhaps tempting to label the image-making within FRTs’ data-image objects as algorithmically generated and invisible to human vision which, therefore, produces an object that is outside of representational systems. Such thinking is inline with Vilém Flusser's writing on technical images, images that are particles of information formed by calculations (2011). Harun Farocki names similar types of images ‘operational images’, images that are part of an operation driven by computer vision systems, such as a drone strike (2004); Jussi Parrika further defines such images as those that ‘[concern] operations, a key term that ties to infrastructures, logistics, and all manner of actions that function to sustain, mobilize, analyze, and synthesize the thing we have grown to call “images.”’ (2023: viii). Along with the same logics, Trevor Paglen describes invisible images as image-making that is centred upon machine-to-machine exchanges of vision, such as those within AI machine learning systems, that have little to nothing to do with human vision (2016). More recently, Andrew Dewdney, in Forget Photography, defines ‘networked images’ as non-representational images that are entirely embedded within digital circulation and computational logics that bear little to no resemblance to forms of image-making based in photography (2021).

To be clear, FRTs do involve images and image-making that is similar to networked, invisible, operational and/or technical images. Importantly, however, before a facial image becomes one of the forms of image-making named in the prior paragraph, it is an image of the face. Facial recognition remains a unique problem compared to other generalized acts of computer vision and object recognition in that the face has a long harmful history as the site of authority, racism and domination, in particular within photographic practices, that must be acknowledged when thinking through the problematics of FRTs.

Specific to this paper, the face within a mugshot database is an example of Browne's digital epidermalization, wherein the mugshot is an especially acute example of a site of political categorization and hierarchization based on imposed logics of difference and criminality. Such thinking is further supported by work such as Jonathan Finn's Capturing the Criminal Image (2009), and Katherine Biber's Captive Images (2006), wherein mugshots are examples of the types of images that are ways of making visible the types of faces deemed to be criminal while also producing discourses that construct the very fact of their criminality. As importantly, McKay and Lee provide an excellent literature review of policing image-making, in the end arguing that ‘photo-technology has had a multiplying effect for Black and brown bodies – already marked as flawed, the mugshot intensifies their presumed criminality’ (2020: 435).

Image making that targets the face, whether invisible or visible, networked or non-networked, is necessarily bound up in specific histories, actors and examples of power that must be addressed differently that image making involving other acts of computer vision. Speaking specifically to mugshots, such images are cross-cut with vectors of socio-technical logics (such as policing practices, neighbourhood compositions, access to health care and education) that are surfaced within the image itself, then compounded when collected into mass databases, then again when trained with deep learning techniques and deployed within AI-enabled technologies like FRTs. As such, and as this practice continues into 2024, FRTs that utilize mugshots produce a visuality wherein the vigilance guiding its observation is always going to disproportionally target vulnerable populations, over-producing facial data of those populations, such as mugshots, then reincorporating such materials and images back into the improvement and evaluating of the technology, cyclically producing and reinforcing that vulnerable population.

Examinations of specific datasets produced in relation to the American War on Drugs show how the mugshot, throughout its historical and contemporary use in FRTs, has been designed to be interoperable and interchangeable so as to enable the fastest and most efficient forms of the technology, propagating widely and quickly, so that the digital epidermalization captured within the mugshots produced during the War on Drugs becomes inscribed into future uses of FRTs, haunting and resonating outside the mugshot and into the daily life of citizens.

The American war on drugs and mugshots

Within the United States, Isacson contends, the late-1970s and 1980s saw a turn away from concerns related to Communism and the Cold War towards the securitization of America against drug trafficking; the ‘health’ of the nation depended on not only containing the threat of drugs to individual user's lives but also protecting citizens from the knock-on effects of crime and violence related to drug use (2005). Though it began with Richard Nixon during his presidency in 1971, the Ronald Regan presidency leaned heavily on this rhetoric to greatly expand America's War on Drugs in the 1980s, granting the American military the authority to combat drug trafficking, with a particular focus given to Central and South American countries. Recalling the introduction to this article, looking at the construction of FRTs during the 1990s showcases the political, representational and technical protocols working symbiotically across military, corporate, governmental and university infrastructures to produce mugshots that could be incorporated as interoperable image-data objects into FRTs. The political protocols must be understood as an essential part of the development of FRTs, in particular as they relate to border securitization, law enforcement and criminality. Initially, FRTs were seen as a border security tactic with extremely high potential because the technology was capable of decentralized and automated vision, or ‘interdiction’, a vision that was able to identify and sort individuals and populations without having to stop them; technologies like FRTs were then tasked with interdiction, or the non-intrusive ‘sorting of legitimate traffic from that which might be illegal’ (American Joint Chief of Staffs, 1998: 28). The 1980s and 1990s War on Drugs, building from prior decades of rhetoric centred on the governmental War on Crime, leveraged facial databases of border-crossers and immigration applications, instigating the infrastructure that would be essential to the mass explosion of the technology post-9/11 and the rise of FRTs to the point of contemporary near-ubiquity.

However, such discussion frames the War on Drugs as largely an international conflict, with FRTs tasked with deploying a complex of visuality at international borders (namely the Mexican-American border), and, as such, does not address the very real effects that the War on Drugs had within the United States and the subsequent production of mugshots as part of domestic law enforcement practices. Knowing this, within America, expanded forms of control, which included expanded police presences in neighbourhoods populated by BIPOC communities and mandatory sentences for drug possession, fit within long-standing rhetoric that have associated racialized bodies with illicit activity and the need for their reform for the safety of society at large. The War on Drugs became a way, as Paul Manning states, to ‘legitimate the deepening strategies of formal regulation and surveillance’ (2015: 83). This includes the 1986 Anti-Drug Abuse Act which re-introduced mandatory minimum sentences for drug crimes such as possession and where a gun was present; it also expanded the abilities to more greatly punish low- and mid-level drug dealers with drug conspiracies (Corva, 2008: 180). As Don Stemen details, this is directly tied to a 227% rise in people arrested for drug law violations between 1970 and 1989, with arrests for cocaine or heroin rising an astounding 463% over that same time period (2017: 397–398); in terms of the resulting mugshots from such arrests, it must also be made clear that having one's mugshot taken only points to an arrest, not a conviction, and so the faces within a mugshot are criminalized without necessarily being proven guilty of a crime. Yet, such effects were not equally spread across racialized and gendered demographics: Michael Tonry, in ‘Race and the War on Drugs’ makes clear, backed by a wealth of compelling statistics, that ‘The War on Drugs foreseeably and unnecessarily blighted the lives of hundreds of thousands of young, disadvantaged Americans, especially black Americans, and undermined decades of effort to improve the life chances of members of the urban black underclass’. (1994: 27). The focus on street-level drug dealing alongside enforcement of mandatory minimum sentences for drug charges in the last decades of the twentieth century lead to a reality where ‘racial disparities in imprisonment began to rise in the 1960s and reached all-time highs in the 1980s and early 1990s’. (National Research Council, 2014: 94); more specific to the War on Drugs, drug-related arrest rates were six times higher for black individuals than white individuals (Blumstein and Wallman, 2006: C-3).

As established in the introduction to this article, the NIST and the related government agencies in the 1990s then utilized the discourses justifying the War on Drugs, despite their biases, to create and improve the state of FRTs. The production and reinforcement of young black and Latino men as a vulnerable population was an undeniable result of the War on Drugs, in turn creating bureaucratic materials, such as mugshots gathered into facial databases, that could be appropriated into technologies such as FRTs.

Further, speaking generally about the introduction of increasing computational automation into policing practices, Ian Alan Paul argues, ‘The automation afforded by electronic machines also opened the way for an interoperability to be established through the use of protocols – shared digital languages which facilitate the automated communication between computers – which made it possible for heterogeneous forms of surveillance and control to be executed interchangeably and concurrently’ (author's italics, 2022: 11). Recalling that interoperability was flagged specifically as a goal within ‘Best Practice Recommendation for the Capture of Mugshots’ outlined in the opening paragraph of this paper, the standardizing of mugshots into interoperable image-data policing materials was essential to the War on Drugs and the decades following 9/11. Beginning in the 1990s, mugshots are images that were/are meant to be frictionless digital objects, collected into larger and larger databases that enable interchangeable and concurrent algorithmic policing practices and beyond.

Three examples of mugshot databases within the development of FRTs

Mugshots databases are primarily kept from public access, but the NIST, operating under mandates of technological benchmarking and improvements for decades, has made a number of such databases available under the rationales of furthering the advancement of the technology within its FRVT infrastructure. The SD-18 is one such example which, according to its internal documentation, ‘is being distributed for use in developing and testing of mugshot identification systems’ (NIST, 2016). Of the 1573 people in the dataset, only 78 are women; as Keegan also notes, 175 of the photos are of minors as young as 12 years old (ibid.: para. 6). While the database was released to the public in 2016, an exact date of its establishment is unclear: Richard Press, a spokesperson for the NIST, in an interview with Jon Keegan, states that ‘they were shared with NIST in the 1990s by the FBI to ‘support research in mugshot identification’. Press said that when the FBI provided the collection of images to NIST, it confirmed that each of the individuals in the photographs was deceased and that they each had “criminal records.”’ (2023: para. 18). The time period laid out by Press overlaps with this article's arguments around the height of the American War on Drugs and also demonstrates the long life of such image-making, where mugshots, so long as they are structured to be interoperable and frictionless, can be relevant data resources for decades.

The SD-18 database is stark both in representation and in data practices: there are front and profile photos, some over-exposed and washed out, while other images, in particular those capturing faces that present as darker-skinned men, carry facial features that are indistinct and blurry; some faces are bruised and/or are covered by bandages, signalling violence. Yet, any sense of the bodily history and/or violence attached to the person in the mugshot is completely unaddressed. Instead, each pair of faces is annotated with only the three categories of ‘Gender’, ‘Age’ and ‘Position’. The photos are not accompanied by a data category for ‘Race’ but such a category will become a standard protocol in later mugshots databases. However, within SD-18, despite the presence of race at the photographic site of the ‘criminal’ face, its absence in the technological workings of the database effectively ignores the power dynamics of a racialized face and its representations, in turn refusing to acknowledge the realities of the socio-political relationships between race and criminality in the dataset.

By contrast, the technical details of the images’ transformation into the digital is far more robust, describing the scanning tools used, the process for converting image files into a standardized file format, the software used for compressing and extracting the images once they are made digital, as well as describing the nesting directory where the digital files sit (Watson and Flanagan, 2016). This fixation on technical protocols is very common within facial databases, and, in that, SD-18 strips away any consideration of the people, their affect, their lived experience, their life, from the image. Such thinking extends to mugshot databases in general, where those captured in photos are rendered as facial and demographic data; there is no sense of the other vectors of policing and other forms of power that may have led to the face being included in a mugshot database. Instead, the uni-directional vision within such databases flattens the faces into a basic representation and, in the case of SD-18, three data points, only to then transform those data-objects into standardized, digital materials that can be bundled and circulated with ease. The focus on digital transportability and interoperability is clear and takes precedent over any other concern.

The NIST's MEDS and MEDS II databases, like the SD-18 database ‘is a test corpus organized from an extract of submissions of deceased persons with prior multiple encounters’ (NIST, 2010: para. 1). Unlike SD-18, the photos are primarily in colour and include expanded data annotations such as eye colour, height, weight and race. Such annotations increase the legibility of the face, while also aiding in the expansion of the technical protocols to become increasingly effective at its acts based in predictive logics based on demographic information. As further proof of this, the NIST flags the images immediately as multi-functional and adaptable: ‘MEDS is provided to assist the FBI and partner organizations refine tools, techniques and procedures for face recognition as it supports Next Generation Identification, forensic comparison, training, and analysis, and face image conformance and inter-agency exchange standards’ (ibid.: para. 1). Kate Crawford, in writing about MEDS in her Atlas of AI, notes that the NIST, ‘in collaboration with the Intelligence Advanced Research Projects Activity, has run competitions with these mugshots in which researchers compete to see whose algorithm is the fastest and most accurate’ (2021: 92). By design, the images can shift between different organizations and purposes with ease, flexibly adapting to any number of facial recognition operations as if neutral data, resulting in a contemporary moment, Crawford argues, where such mugshots represent ‘a shift from image to infrastructure’ (author's italics, ibid.: 93).

However, that infrastructure and its desires for technical protocols must not be allowed to overshadow the images and the individuals within: as Os Keyes describes looking at the MEDS databases and their documentation, ‘Where I saw a screaming middle-aged man, Andrew P. Founds, Nick Orlans, Genevieve Whiddon (of the Mitre Corporation), and Craig Watson (of NIST) saw an example of the sort of image that requires manual editing to be useful in facial-recognition testing’ (2019: para. 5). The MEDS databases, and their accompanying violence and social-political logics, the individual images and the images massed together as a dataset, do nothing to acknowledge their own infrastructures or systems of production; nor do they acknowledge the affect and narrative of the individuals captured within those mugshots. Further, when mugshot databases are skewed by phenomena such as the American War on Drugs and racialized policing practices, the resulting enhanced complex of vision will continue to perpetuate the very logics that ensured the over-representation of racialized populations in the database to begin with.

While SD-18 ignored race, and its impacts, the categorization of race as data in the MEDS databases grants FRTs that use the database the abilities to better target and sort specific faces based on how their race is defined in the dataset despite its obvious bias. This is incredibly problematic: as Keyes establishes, ‘though black people make up just shy of 13 percent of the U.S. population, they make up nearly 50 percent of the MEDS photographs’ (ibid.: para. 2). This disproportionate gaze, and its activation of digital epidermalization, is made even more problematic by the fact that the photos within the database capture deceased people who cannot advocate for themselves is essential to the datasets’ construction and circulation. This is acceptable only if that data is erroneously viewed as neutral and/or of subjects who have given up their citizenship rights: the rationales for including these photos dovetail with logics of securitization which are amplified by further rationales that argue that those captured within carceral apparatuses do not have the same rights as those labelled as law-abiding citizens. When developing FRTs, the mugshots in the MEDS databases are then appealing in two key ways: first, tying into technical protocols, the controlled environments provide stable forms of the face with little variation in pose, light, etc; second, uniting with political protocols, those captured are viewed as having forfeited their rights as citizens, alive or dead, and so the use and re-use is untroubled by questions of privacy and personal data sovereignty.

The MEDS databases also deliberately involves the same person captured over different periods of time and can therefore be used to solve the technical problem of tracking identity, via the face, over time as a person ages. While the databases could be used to train and test, in general, how to identity and verify specific faces, the desire to track faces over time acknowledges the volatile indeterminacy of the body as it ages; yet, the databases, and documentation attached, treats this as a computational problem in search of the correct series of technical protocols.

In this way, MEDS and SD-18 are similar to the MORPH database which gathered longitudinal data by way of mining public records between 26 October 1962 and 7 April 1998 (Ricanek Jr. and Tesafaye, 2006: 2). What the MORPH researchers do not overtly state, but is centrally present in their documentation, is that uncritically mixed in with their data are a large number of mugshots, many of which also appear in the SD-18 database. While the first paper detailing the construction of the database was published in 2006, work with the dataset continued into at least 2017, where Morgan Ferguson gave a presentation on the second generation of the dataset, Morph-II, and resulting research based on the dataset (2017).

The MORPH database houses 55,134 images, 46,645 of which (84.6%) are ‘African’ and 36, 832 of which (66.8%) are ‘African Male’; the median age is 33 years old, with 39, 151 (71%) of the subjects being under the age of 40 (University of North Carolina at Wilmington, n.d.). Nothing in the documentation complicates these alarming statistics. Further, returning to the prior discussion of the domestic impacts of the War on Drugs, the MORPH database houses a disproportionate amount of young, black male faces compared to their inclusion in the population as a whole. Acknowledging that there is no way to prove the exact percentages of arrests related to the War on Drugs without examining specific criminal records and also repeating that being arrested does not mean a person has been convicted of a crime, the broad statistics in a database like MORPH do demonstrate potential connections between the vigilance and suspicion that constructs the domestic drug criminal by way of over-policing practices and therefore mugshots, and the consequent over-inclusion of such faces in FRT databases and their subsequent deployment within FRTs.

These three databases largely ignore the impact of representational protocols based in difference that are central to digital epidermalization while also demonstrating the erasing of socio-political factors that go into the production of criminalized bureaucratic material, including in reaction to the American War on Drugs. As Jefferson reminds ‘Attention must be directed towards the actual conditions in which [policing data and information] is produced and put into action’ that then structure the criminal justice apparatus. (11) The authors of The Prison House of the Circuit echo this, insisting that the observer effects of algorithmic policing must be acknowledged and centralized within such applications of power: the observation itself is biased towards certain predictions which in turn reflect the resulting behaviour of that observation (Packer et al., 2022: 115).

The effects of algorithmic policing are not lost as the representations become invisible-, networked- or operative images within the construction and deployment of a FRT; instead, their effects are multiplied by the ease in which the interchangeable data-objects are able to be digitally circulated and re-used, re-appropriated and renewed into the current moment. As one final example of such reappropriating into further versions of FRTs, at the outset of the Covid-19 pandemic, the NIST utilized the MEDS databases in their proposed methodology for solving for masked faces under observation by a FRT: the initial report computationally adds masks to the mugshots as an example of how such a methodology can be executed (Ngan, Grother and Hanaoka, 2022: 4).

Conclusion – Generative AI and interoperable and standardized infrastructural materials

The three discussed databases are evidence of the standardizing and desires for interoperability within the long history of computational policing practices: a mugshot needs to not only capture the face of criminality but do so in a way that, when folded into increasingly larger digital databases, can be operationalized faster and in more frictionless manners than the generation of infrastructure before. These dual forms of knowledge produced by the criminal justice apparatus, of the criminal's face and how to digitally circulate that criminal's face, are essential to algorithmic policing and governance.

Yet, recalling Ruha Benjamin's quote from the introduction to this article, these forms of knowledge are not limited to those labelled as criminal, and the interoperability within FRT mugshot databases allows such materials to slip into civilian life. Subsuming digital epidermalization and carceral logics into databases that are then algorithmically processed in ways that help to standardize the general state-of-the-art of FRTs, as the NIST and FRVTs do, ensure that those logics become embedded within FRTs included in test-taking environments, job applications and other probabilistic evaluations of everyday citizens. Expanding further, Paul, in ‘Are Prisons Computers?’ (2022), argues that the logics within carceral practices that separate and provide commensurability are a form of the digital and that the two, as structuring elements of the modern society, work symbiotically. This is evidenced by the way that materials gathered by the War on Drugs linger and expand over decades, providing protocols and methodologies for improving and deploying technologies like FRTs. This broadens the complex of vision within a FRT to the whole of citizenry, wherein the generalization of ‘policing across all of space should not be read as a claim that everyone is equally policed, but rather that the hierarchies that structure all social life and which are maintained by police necessitate the policing of all of society’ (para. 62). This dynamic is clearly present in the contemporary moment, where Clearview AI asserts that their database, compiled by scraping social media and other civilian data sources such that they’ve constructed a database of over 50 billion faces, has been accessed by American police forces over one million times (Clayton and Derico, 2023). The conflation of ‘criminal’ and ‘civilian’ in this instance is then further expanded not only by the gaze of police dashboard cameras and police body cameras that are enhanced by FRTs but also by observation assemblages like CCTV systems that cover large areas of public and private space with unblinking vision.

The mugshots gathered during the War on Drugs continue to be used in algorithmic governance today. Within this moment, not only to images become interoperable but also the AI models that are built from such images slip between different algorithmic acts: a FRT does not necessarily need access to facial data; it needs access to a AI model already trained on facial data, with many such models freely available and circulating online. In this, the potential carceral logics and digital epidermalization within mugshots are further engrained within the AI models then further still when those models are deployed. The accompanying racializing is not limited to mugshots: as the report from the November 2023 Washington Post report details, Generative AI systems like Midjourney produce an array of portraits of non-white people given the prompt ‘A person at social services’ and images of white men when prompted with the phrase ‘a productive person’ (Tiku, Schaul, and Chen).

The long tail of mugshots does not stop at the materials themselves, but extends to the visualities, digital epidermalization and representational protocols established by mugshots in general, to the point where such logics are being replicated by Generative AI systems such as DALL-E (Open AI, 2023). Prior to the technology's public release in late-2021, I prompted beta-versions of DALL-E with the phrases ‘A citizen's face’, ‘A migrant's face’ and ‘A mugshot’. I do not reproduce such images, as I feel that doing so recirculates the harmful logics within them, so I will describe them. ‘A citizen's face’ produced primarily light-skinned male-presenting faces; ‘A migrant's face’ generated mostly brown- and dark-skinned children, their large eyes staring pleading directly to the audience; finally, ‘A mugshot’ produced dark-skinned, male-presenting faces, scowling, many with tattoos on their necks, looking remarkably similar to many images within the MEDS database. This makes complete sense when one understands that generative AI systems, like all AI systems including those embedded within FRTs, depends entirely on the data they have been trained on. Prompting such systems with ‘a mugshot’, reproduced those images labelled as mugshots in its training datasets, in turn visualizing the standardized and interoperable face of the criminal.

When I attempted to replicate the experiment with the public version of DALL-E months later in 2022, after the technology had been made public, I was told that my prompt of ‘A mugshot’ violated its content policy and, therefore, the system refused to generate any images. When examining the content policy updated on 20 July 2022, I saw that DALL-E banned content that included materials that were based in ‘hate’, ‘violence’, ‘illegal activity’ and, incredibly broadly, the ‘political’ (Open AI, 2022); a March 2023 update to the usage policies adds further detail to the broad headlines found in the 2022 version, but again is unclear as it what exactly was causing the prompt to be banned, perhaps with the exception of prohibiting materials where there is ‘[content] that expresses, incites, or promotes hate based on identity’ (Open AI, 2023: para. 4). If this is true, then this is the system telling on itself, acknowledging, clumsily, that a mugshot is political, is potentially hateful and violent. This initial solution brings to mind a prior example, in 2015, of Google's image classifier consistently identifying Black faces as gorillas (Vincent, 2018); instead of grappling with the fundamental logics and construction of the system, the embedded racism, engineers simply erased the machines ability to identify gorillas (ibid.; Simonite, 2018). Likewise, within the above example, the complete erasure of such images, as DALL-E undertook, was not the most effective solution, but it is enlightening in that it demonstrates the spread of the racialized logics and representational protocols in the ever-expanding carceral space, in both the real and digital worlds.

As a coda to this, and as a final demonstration of the whac-a-mole-like game of content moderation around complex AI and the sustaining presence of racialized images within AI systems, I returned to DALL-E in March 2024 and promoted it again with the phrase ‘a mugshot’. Two of the images were that of a mug on a table; one was one a white woman looking into a magnifying class, and the final one was of a white man in a mugshot-like photo, where the light and pose are that of a mugshot, but his left hand appears to be holding a mug. However, the phrase ‘Mugshot of a black man’ produced four stern looking dark-skinned men, all centred in the frame and staring directly into the camera, effectively replicating the visualities of the racialized mugshot; that these images are ‘fake’ and not of a real person doesn't do away with the fact that the image-making within DALL-E is haunted by the preconditions of mugshots, such as those gathered during the War on Drugs. I received similar results for the ‘A mugshot of a black criminal’ though the face was of the same gaunt dark-skinned man across the four images. By comparison, ‘a mugshot of a white man’ and ‘a mugshot of a white criminal’ produced only two images, of eight, that is marked by the classic visuality of a mugshot, while the other three have variations in facial expression (smiling, yelling), action (pointing at the camera) and pose (at angles not directly facing the camera). Interestingly, however, the prompts ‘A black criminal’ and ‘a white criminal’ resulted in a cartoon of a cat and dog that informed me that ‘It looks like this request may not follow our content policy’ with a URL to said policy. These simple experiments in the manipulation of prompts (‘A mugshot’ versus ‘a mugshot of a black man’) speak to the ways in which we might ask such systems to visualize their biases for us, to tell on themselves and reveal the logics within their deep learning and data practices. That ‘criminal’ is a flagged term but ‘mugshot’ is not speaks to the messy, and often ignored, practice of data labelling within machine learning datasets and the deeply flawed ways in which the frontier of Generative AI is being unevenly monitored.

Although technologies like Generative AI appear to emerge and proliferate from nowhere to ubiquity, a historicizing of their image-making and data practices reveal instances, such as the War on Drugs, that continue to be deeply influential on the technology's operations. For technologies like FRTs, which have a straighter line between past and present versions, it is equally important to underline the constant desires for interlocking and interoperable digital infrastructures from an unending multitude of government and corporate stakeholders, infrastructures that are built from interchangeable data-objects, dating back in the case of mugshots within automated FRTs into at least the 1990s and the American War on Drugs, that structure algorithmic policing practices. Such images and image-making are not contained to law enforcement tactics and strategies, nor do they lose their effects or affects after decades of circulation: rather they leach into the whole of society, wherein the criminality of certain populations of faces, most often those that are racialized, circulate and perpetuate and establish themselves as ‘fact’ within algorithmic systems. The representational qualities and histories of such image-making are not lost nor subsumed within algorithmic processing, but rather deserves constant acknowledgement and urgent continued attention.