When 93% Isn’t Enough: How Algorithmic Certainty Corrupts Police Investigations

The Deceptive Authority of a Number

A 93% match. To a jury, to a public eager for swift justice, that number sounds reassuringly precise, almost irrefutable. It suggests a high degree of certainty, a mathematical precision that traditional, messy police work often lacks. But when Florida police arrested Robert Dillon in August 2024, alleging he tried to lure a child, that seemingly robust 93% facial recognition match was, according to his lawsuit, dead wrong. Dillon, a 52-year-old resident of Fort Myers, was hundreds of miles away, had never been to Jacksonville Beach, and license plate reader data confirmed his absence. The case isn’t just about an error-prone algorithm; it’s about how the perceived authority of a probabilistic score can warp fundamental investigative principles, turning algorithmic output into a self-fulfilling prophecy for law enforcement.

The lawsuit explicitly states, “This case is about what happens when police let an error-prone artificial intelligence system stand in for an investigation.” This claim cuts to the heart of a critical structural implication that Silicon Valley’s cheerleaders for ‘smart policing’ routinely overlook. It suggests that rather than serving as one tool among many, facial recognition technology risks becoming the primary lens through which evidence is filtered, thereby inverting the foundational presumption of innocence and nudging officers towards confirmation bias.

The Feedback Loop of Digital Forensics

The core problem isn’t the existence of facial recognition technology itself, but the context and quality of its deployment. In Dillon’s case, the ‘match’ was reportedly based on a low-quality image: a photograph of a McDonald’s computer screen displaying surveillance footage. This is not the pristine, high-resolution data set often used to train and validate these complex neural networks. It’s a fundamental misunderstanding of GIGO — garbage in, garbage out — applied to human liberty.

When a system delivers a ‘93% match’ from such degraded input, that percentage becomes less a measure of objective truth and more an indicator of the algorithm’s confidence in its best guess, given what little it had to work with. The critical step, then, should be for human investigators to apply robust skepticism. Instead, the lawsuit alleges officers actively built a case to confirm the machine’s answer, effectively sidelining or concealing exculpatory evidence like Dillon’s geographic distance and the lack of other digital footprints. This creates a dangerous feedback loop where algorithmic ‘evidence’ dictates the direction of the investigation, rather than being subjected to it.

Internationally, the debate around biometric surveillance and its deployment in law enforcement has been far more fraught than in the U.S. Privacy regulations like GDPR in Europe often place a heavier burden on authorities to justify mass surveillance technologies, particularly those that involve constant, untargeted data collection. Many European cities and nations have imposed outright bans or severe restrictions on public sector use of facial recognition, recognizing the profound civil liberties implications. This contrasts sharply with the often permissive, and certainly less restrictive, environment in many American jurisdictions, where the allure of technological efficiency often eclipses concerns for due process.

Incentives and the Erosion of Due Process

So, why does this happen? The incentive for law enforcement agencies to embrace tools like facial recognition is clear: perceived efficiency. Facing budget constraints, staffing shortages, and mounting pressure to solve crimes quickly, an AI system promising a ‘93% match’ offers a powerful shortcut. It reduces the need for painstaking traditional detective work—interviewing witnesses, canvassing neighborhoods, cross-referencing multiple data points. This perceived advantage, however, carries a hidden cost: the erosion of investigative rigor and the potential for grave miscarriages of justice. It’s far cheaper to trust an algorithm than to fund a fully staffed investigative team.

My sharpest skeptical observation is this: these technologies are often sold, and bought, under the premise of making policing ‘smarter,’ but in practice, they often make it less thoughtful and more prone to automated injustice. The narrative around artificial intelligence in policing often focuses on its potential to enhance safety, but rarely on its capacity to systematically undermine the very justice system it claims to serve. The rush to adopt new tech without corresponding investments in human oversight, critical thinking, and a clear understanding of algorithmic limitations isn’t progress; it’s a dangerous abdication of responsibility.

The Dillon case serves as a stark reminder that a ‘match’ from a machine, no matter how confident its percentage, is merely a hypothesis. The police’s role is to test that hypothesis rigorously, not to blindly validate it. The future of equitable justice demands that we push back against the seductive simplicity of algorithmic certainty and reaffirm the messy, human-centric processes that are the true bulwark against wrongful prosecution. Anything less risks transforming our justice system into a bureaucracy of automated suspicion, where a person’s digital ghost can condemn them, regardless of physical reality.