Case Study 4.2: The Mugshot Goes Digital — Facial Recognition and the Legacy of Bertillon

Overview

This case study examines the use of facial recognition technology in policing and public space monitoring as the direct descendant of Alphonse Bertillon's nineteenth-century identification photography system. It traces the continuity between the mugshot's social sorting logic and the algorithmic classification systems used in contemporary law enforcement, border control, and public space surveillance.

Estimated Reading and Analysis Time: 60–75 minutes

Background: From Bertillon to FaceFirst

In 1882, Alphonse Bertillon persuaded the Paris police prefecture to let him photograph arrested criminals using a standardized format: frontal and profile views, taken under consistent lighting, with neutral expression. Combined with eleven body measurements, the resulting card enabled police to match a currently-arrested person against their prior arrest records — identifying recidivists who might use aliases.

The mugshot's design was intentionally anti-individual: it stripped away the subject's self-presentation, imposed a standardized format, and produced an image designed for institutional comparison rather than personal expression. The mugshot's subject was not a person being portrayed — they were an identification datum being captured.

More than 140 years later, the mugshot's descendant — the facial recognition algorithm — performs the same identification function at a scale Bertillon could not have imagined: comparing a face against a database of millions, in milliseconds, without requiring the subject's cooperation or even awareness.

How Facial Recognition Works

Modern facial recognition systems operate in a three-stage process:

Detection: The algorithm identifies faces in an image or video stream, locating and isolating each face for processing.

Feature extraction: The algorithm maps the geometry of each face — the distances between specific facial landmarks (eyes, nose, mouth, jawline) — and produces a numerical representation called a "faceprint" or facial feature vector.

Matching: The faceprint is compared against a database of stored faceprints, and candidates whose stored faceprints are within a defined similarity threshold are returned as potential matches, ranked by confidence score.

The accuracy of the system depends on the quality of the enrollment image (the stored faceprint), the quality of the probe image (the new face being identified), and the algorithm's ability to handle variations in lighting, angle, age, and expression.

Law Enforcement Applications

The FBI's Facial Analysis, Comparison, and Evaluation (FACE) Services

The FBI's FACE Services unit provides facial recognition search capabilities to federal, state, and local law enforcement agencies. Its databases include the Next Generation Identification (NGI) System's Interstate Photo System (IPS), which contains photographs from criminal arrest records across the country, as well as external databases including state driver's license photos, passport photos, and visa application photos.

The access to driver's license and passport photos is significant: these databases contain images of millions of law-abiding citizens who have never been arrested. The surveillance function of the mugshot has expanded — through function creep — from the identification of criminal recidivists to the identification of any person in a photograph, regardless of whether they have any criminal history.

A Government Accountability Office (GAO) report in 2019 found that federal agencies were using facial recognition from 18 different external databases, many of which had not been assessed for the accuracy, privacy implications, or civil liberties consequences of their inclusion.

Detroit and the Robert Williams Case

In January 2020, Robert Williams — a Black man in Detroit — was arrested at his home in front of his family for shoplifting theft he did not commit. The case against him was based entirely on a facial recognition match that identified Williams as a suspect. The match was wrong.

An FBI analyst had submitted a still image from grainy surveillance footage to Michigan State Police's facial recognition system. The system returned a match to Williams' driver's license photo. A detective confirmed the match — despite Williams' driver's license photo being substantially clearer and higher-quality than the surveillance image, and despite the visual similarity being, by most accounts, quite limited.

Williams was arrested, held for 30 hours, and presented with his own driver's license photo in an interview room. He pointed to the photo, then to his face, and said: "I hope you don't think all Black men look alike." The detective acknowledged that the case rested on the facial recognition match, and that Williams should "be thankful" the analyst had done further checking. Williams was eventually released; the charges were dropped in 2021. The Detroit Police Department suspended its use of facial recognition for investigative leads.

The Williams case is not isolated. The MIT Media Lab's AI researcher Joy Buolamwini, in research published in 2018, demonstrated that leading commercial facial recognition systems had error rates of up to 34.7% for darker-skinned women, compared to 0.8% for lighter-skinned men. The accuracy disparity was dramatic and had been known to vendors but not disclosed to law enforcement clients.

Surveillance Cameras and Real-Time Identification

Beyond the use of facial recognition to identify suspects from still images, several law enforcement agencies and private actors have deployed real-time facial recognition systems that match faces captured by surveillance cameras against databases of identified individuals.

New York City: The NYPD deployed facial recognition beginning in 2011. A 2019 investigation by the Financial Times found that the NYPD had used facial recognition in thousands of cases, with matching sometimes performed by detective units without formal policy guidelines.

London: Metropolitan Police conducted several "live facial recognition" trials beginning in 2019, deploying surveillance cameras in public areas and matching faces in real time against a watchlist of individuals with outstanding warrants. Civil liberties organizations including Liberty challenged the deployments as unlawful mass surveillance.

Clearview AI: A commercial facial recognition company, Clearview AI scraped approximately 10 billion facial images from social media platforms without the subjects' consent, trained a facial recognition model on these images, and sold search access to law enforcement agencies and private companies. The Clearview model enables the identification of individuals not from criminal databases but from any photograph posted to the public internet — including images posted by subjects' friends, at events the subject attended, or in contexts entirely unrelated to law enforcement.

The Social Sorting Logic

Bertillon's Population

Bertillon's system was designed to identify a specific population: criminal recidivists. The system's category — "criminal" — was defined by the legal process of arrest and conviction, and the system operated within that boundary. The mugshot was, in principle, limited to people who had been processed by the criminal justice system.

Real-time facial recognition and Clearview-style systems are categorically different: they operate on an entire population, not a pre-defined suspect population. Every face captured by a surveillance camera — every pedestrian, every transit rider, every shopper — is compared against the watchlist database. The watchlist may be composed of people with criminal records, but the matching process subjects everyone in the camera's field to identification scrutiny.

This is social sorting operating at the population level: everyone is potentially identified, classified, and acted upon, with those whose names appear on watchlists treated differently from those whose names do not. The sorting is continuous, real-time, and invisible to most of those sorted.

The Discrimination Problem

The accuracy disparities documented by Buolamwini and Gebru (the "Gender Shades" study, 2018) have direct consequences for social sorting.

If a facial recognition system used in a public camera network has a 34.7% error rate for darker-skinned women and a 0.8% error rate for lighter-skinned men, then darker-skinned women walking past a camera are approximately 43 times more likely to be falsely identified as a match for a watchlist name than lighter-skinned men.

In a practical scenario: if a police watchlist contains 1,000 names, and 10,000 people walk past a camera in a day, a system with these error rates would generate approximately 3,470 false matches for darker-skinned women and approximately 80 false matches for lighter-skinned men. Each false match generates a potential police interaction — stop, questioning, possible arrest — that the individual did not merit.

The facial recognition system is not neutral. It embeds racial disparities in its training data and its accuracy characteristics, and then automates and scales those disparities in real-world enforcement decisions.

The Genealogy: Bertillon to Algorithm

The Bertillon-to-facial-recognition genealogy illuminates several key surveillance studies themes:

Continuity of social sorting logic: Bertillon's system was designed to identify a suspect population — criminals and recidivists. Contemporary facial recognition is used for the same purpose, but now operates on an entire population simultaneously, with the "suspect" designation emerging from the match rather than preceding it.

Function creep: The mugshot began as a tool for criminal recidivist identification. It expanded to all criminal arrests. Drivers' license and passport photos entered law enforcement databases without subjects' specific consent. Commercial facial recognition scraped public internet images without any consent. Each expansion of the database was presented as a routine technical decision; collectively, they transformed a criminal identification system into a universal population identification system.

Scale and accuracy: Bertillon's manual system was reviewed by human examiners who could exercise judgment about borderline matches. Algorithmic systems present matches as ranked lists with confidence scores, but frontline law enforcement personnel often lack the training to critically evaluate confidence scores or to recognize the significance of accuracy disparities across demographic groups.

Visibility asymmetry: The person whose face is captured by a public camera and matched against a database typically has no knowledge that this has occurred. The system operates entirely on the watcher's side; the watched has no access to the match results, no opportunity to correct errors before action is taken, and no formal notification that they were identified.

Discussion Questions

1. The Scope Problem: Bertillon's mugshot system was limited to people who had been arrested. Contemporary facial recognition can identify anyone in a camera's field of view. Is this scope expansion justified by the technology's capabilities, or does it represent an ethically significant change in what surveillance does? Is there a principled basis for limiting facial recognition to identified suspect populations?

2. The Accuracy Disparity: The Gender Shades research found error rates dramatically higher for darker-skinned women than for lighter-skinned men. Law enforcement agencies were using these systems despite knowing of accuracy disparities. What obligations do law enforcement agencies have before deploying facial recognition? Who should be responsible for the consequences of false matches?

3. Clearview and Public Images: Clearview AI argued that its system used only publicly posted images — photographs that subjects had made available on the internet. Does the public nature of the source images make Clearview's surveillance ethical? Is there a meaningful difference between a photograph being publicly viewable and it being incorporated into a mass identification database?

4. The Robert Williams Case: Williams was arrested based on a facial recognition match that a human detective confirmed despite the limitations of the evidence. Is the fault here primarily with the algorithm, with the detective who confirmed the match, with the department that authorized the use of facial recognition leads, or with the system as a whole? What structural changes would prevent similar cases?

5. The Genealogy and Responsibility: The chapter traces facial recognition from Bertillon's nineteenth-century mugshot system. Does this genealogy — which reveals that the technology serves a consistent social sorting function across very different technological substrates — affect our assessment of whether current facial recognition is a new problem or an old one in new form? Does it suggest that technical fixes (better accuracy) will be sufficient, or that structural changes are needed?

6. Jordan's Face: Jordan walks past surveillance cameras on the way to class. Jordan's university may use cameras in its buildings. The transit authority cameras have captured Jordan's face thousands of times. Does Jordan have any meaningful ability to refuse this surveillance? What legal protections exist, and are they adequate?

Chapter 4 | Case Study 4.2 | Part 1: Foundations | The Architecture of Surveillance