First American Wrongfully Arrested by Flawed Facial Recognition Technology

In 2020, Detroit resident Robert Julian-Borchak Williams became the first known American wrongfully arrested due to a mistaken facial recognition match. Police officers showed up at his home, handcuffed him in front of his wife and young daughters, and held him for more than 30 hours. The entire case rested on a single low-quality surveillance image that the Detroit Police Department’s facial recognition system had flagged as a possible match for Williams. Later analysis revealed the software had made a clear error, matching a heavier, lighter-skinned man to Williams, who is Black. The incident, covered extensively by Wired, exposed serious weaknesses in one of the oldest police facial recognition tools still in regular use inside the United States.

The technology at the center of the arrest traces its roots to the early 2000s when Detroit police began experimenting with facial recognition software. Over time the department adopted a system supplied by DataWorks Plus, a South Carolina-based company that integrates multiple biometric algorithms into a single platform. Unlike newer cloud-based services that dominate headlines today, DataWorks Plus offers an on-premise solution that lets agencies maintain their own databases and control updates. Many midsize and larger departments across Michigan and neighboring states still rely on versions of this platform because it connects directly to their existing record management systems and mugshot repositories. The software scans thousands of stored arrest photos and driver’s license images, generating similarity scores when officers submit probe images captured from street cameras or cellphone videos.

Accuracy problems surface quickly when the source image is blurry, taken at an odd angle, or poorly lit. Research consistently shows that error rates climb dramatically for darker skin tones, a pattern documented across multiple independent studies. In Williams’s case, the original surveillance still came from a security camera mounted high above a store entrance. The man in the image wore a surgical mask and a baseball cap, conditions that strip away many of the facial landmarks algorithms depend upon. Despite these obvious limitations, an analyst inside the Detroit Police Department ran the image through the DataWorks system and received a candidate list that placed Williams near the top. Officers treated the result as probable cause rather than an investigative lead, a practice that civil rights groups have repeatedly criticized.

The arrest set in motion a series of events that would reveal deeper procedural gaps. After Williams spent a night in a holding cell, detectives finally allowed him to see the incriminating photo. He immediately pointed out that the man in the picture had a different complexion, body type, and facial hair pattern. Only then did investigators begin to question the match. A second, more careful review confirmed the mistake. Williams was released, but the episode left his family shaken and his reputation damaged in the eyes of neighbors who had watched the arrest unfold on their street. The Detroit Police Department later acknowledged that its facial recognition workflow lacked adequate human oversight and that officers had received minimal training on the technology’s limitations.

This case did not occur in isolation. Across the country, other documented misidentifications have involved similar legacy systems. In 2019, the American Civil Liberties Union examined facial recognition deployments in multiple jurisdictions and found recurring patterns of false positives, especially when the reference databases contain outdated or low-resolution booking photos. The National Institute of Standards and Technology has published repeated evaluations showing that many older algorithms exhibit higher false positive rates for women, elderly adults, and people with darker skin. These technical shortcomings become especially dangerous when police treat the software’s output as definitive rather than suggestive.

Detroit’s experience highlights a broader tension between technological capability and responsible implementation. The DataWorks Plus platform itself is not uniquely flawed; several competing systems display comparable error profiles under challenging conditions. The problem lies in how departments integrate the technology into daily operations. During the Williams investigation, the facial recognition unit operated with little supervision from command staff. Analysts sometimes ran searches without documenting the image quality or the confidence score returned by the software. Defense attorneys later argued that such omissions prevented judges from properly evaluating the reliability of the evidence. In the wake of the lawsuit Williams filed against the city, Detroit announced policy changes that now require supervisory approval before any arrest based solely or primarily on a facial recognition hit.

Legal experts point out that the Williams case tested the outer boundaries of probable cause standards. The Fourth Amendment protects citizens from unreasonable searches and seizures, yet courts have historically granted law enforcement significant discretion when evaluating technological aids. If a fingerprint match or DNA hit can support an arrest, some prosecutors have argued that a facial recognition match should receive similar weight. Critics counter that fingerprints and DNA undergo rigorous validation protocols, whereas many facial recognition systems still operate as black boxes with proprietary algorithms shielded from outside scrutiny. The Williams litigation, which settled out of court, did not produce a definitive appellate ruling on these constitutional questions, leaving room for continued debate in other jurisdictions.

Public reaction to the story revealed sharp divisions. Supporters of expanded police technology argue that facial recognition can accelerate investigations, prevent mistaken identity in other contexts, and bring resolution to cold cases. They note that the same Detroit system has helped identify suspects in serious violent crimes where traditional methods stalled. Opponents emphasize the human cost of errors, particularly in communities that already experience strained relations with law enforcement. For many residents of color, the Williams incident reinforced long-standing fears that emerging tools will amplify existing biases rather than correct them.

In response to growing pressure, some agencies have introduced layered review processes. A typical modern workflow might require an image to pass an initial quality threshold before being submitted to the algorithm. Analysts then compare the candidate list against additional contextual evidence such as vehicle descriptions, clothing, or geolocation data. Only after this multi-step verification can the technology support an arrest warrant. Training programs now stress that the software produces a list of possible matches rather than a positive identification. Even so, resource constraints in smaller departments often mean these best practices remain aspirational rather than routine.

The DataWorks Plus software continues to receive periodic updates, but its core architecture reflects design decisions made nearly two decades ago. Newer competitors offer cloud-based machine learning models that retrain on massive datasets and claim improved accuracy across demographic groups. Yet many police departments hesitate to abandon on-premise solutions because of data sovereignty concerns and long-term contracts. Migrating to newer platforms would require retraining officers, validating new databases, and navigating procurement rules that can stretch for months. As a result, older systems like the one used in Detroit persist even as scientific understanding of their weaknesses grows.

Civil liberties organizations have called for outright moratoriums until independent audits can certify that error rates fall below defined thresholds. Several cities, including San Francisco and Boston, have enacted temporary bans on police use of facial recognition in public spaces. Other municipalities have opted for strict regulations instead of prohibition, mandating warrants, transparency reports, and regular bias testing. Michigan itself has considered legislation that would limit the technology’s role in establishing probable cause, though progress has been slow. The Williams case remains a reference point in these legislative discussions, frequently cited as an example of what can go wrong when safeguards are weak.

Beyond immediate policy questions, the incident raises philosophical issues about trust in automated decision-making. When a machine flags an innocent person and officers act on that flag without sufficient corroboration, the line between technological assistance and technological substitution becomes blurred. Defense attorneys now routinely file pretrial motions challenging any evidence derived from facial recognition, demanding discovery of the algorithm’s error rates, the composition of the reference database, and the qualifications of the analyst who ran the search. Some courts have begun to treat such requests more seriously, recognizing that defendants cannot meaningfully confront the evidence against them if the underlying methodology remains opaque.

For Robert Julian-Borchak Williams, the aftermath extended well beyond his release from custody. He described lingering anxiety, difficulty trusting strangers, and concern that the erroneous record might resurface in future background checks. His attorneys used the case to push for greater accountability, arguing that police departments must bear responsibility for harms caused by tools they choose to deploy. The city of Detroit ultimately paid a settlement and agreed to revise its facial recognition policies, yet the broader national conversation continues. Similar wrongful arrest cases have since surfaced in other states, each adding weight to arguments that legacy systems require urgent modernization or replacement.

Technology companies, for their part, maintain that responsibility rests with the end user. DataWorks Plus has stated that its software provides a ranked list of candidates and that final judgment must always rest with trained human reviewers. The company points to built-in confidence metrics and the ability to adjust sensitivity thresholds. Still, internal police documents obtained through public records requests sometimes show analysts treating high similarity scores as near-certain identifications, especially when under pressure to clear cases. Bridging the gap between vendor disclaimers and actual street-level practice remains an ongoing challenge for oversight bodies.

Looking forward, experts anticipate that facial recognition will remain a fixture in American policing even as the specific tools evolve. The question is whether departments will learn from high-profile failures like the Williams arrest and build stronger guardrails, or whether momentum toward wider adoption will outpace meaningful reform. Independent testing labs, academic researchers, and community advisory boards all have roles to play in shaping standards that protect both public safety and individual rights. Until clearer national guidelines emerge, local decisions about training, database quality, and acceptable use will continue to determine whether this technology serves justice or undermines it.

The Detroit case stands as a cautionary example rather than an anomaly. It demonstrates how an old system, deployed without sufficient skepticism, can produce life-altering consequences for the wrong person in a matter of hours. As more agencies integrate biometric tools into their investigative kits, the lessons from Robert Julian-Borchak Williams’s experience deserve close attention. Proper implementation demands rigorous image standards, continuous bias audits, comprehensive officer education, and a cultural shift that treats algorithmic suggestions as starting points rather than endpoints. Only through these measures can law enforcement hope to harness the potential of facial recognition without repeating the painful mistakes of the past.