Benefits of Facial Recognition in Law Enforcement: What Went Wrong

A grandmother from Tennessee spent six months behind bars in North Dakota. Not because of a confession, not because of physical evidence, not because of a single corroborating witness. Because an AI flagged her face in grainy bank surveillance footage, and then a detective looked at her driver's license and her social media and said, "yep, that's her."

That's it. That was the case. Six months of her life, gone.

13ABC reports that North Dakota police had been investigating a string of fraud cases between April and May 2025 involving a suspect who used a fake U.S. Army military ID card to withdraw thousands of dollars from a bank. Investigators ran their surveillance footage through facial recognition software, got a match, and then, here's the part that should make every working detective wince, a detective reviewed the woman's driver's license and social media to "confirm" she was the suspect. And that was apparently enough to arrest a woman who lives in another state and put her in jail for half a year.

According to a GoFundMe created by a West Fargo man to support her, she is described plainly as an "innocent grandmother jailed 6 months by AI error." The campaign went viral. The outrage was immediate. But the conversation that followed missed the point almost entirely.

Stop Blaming AI in Facial Recognition Wrongful Arrests

Everyone wants to blame the algorithm. It's a clean villain, faceless, corporate, vaguely sinister. But that framing lets the actual decision-makers completely off the hook. An algorithm didn't sign the arrest warrant. A human did. This article is part of a series, start with Stress Test Facial Comparison Method Against Deepf.

Here's what actually happened, methodologically: investigators used facial comparison as a conclusion rather than a starting point. The AI surfaced a candidate. A detective then looked for reasons to confirm that candidate rather than reasons to challenge it. That's not how evidence works. That's not even how good investigative instinct works. It's confirmation bias with a badge on it.

Cognitive science has a name for this: a confirmation bias cascade. Once an investigator receives an initial match, subsequent review of supporting materials, a license photo, a social media profile, is no longer independent verification. It becomes rationalization. The detective reviewing that driver's license wasn't running a second check; he was running a contaminated second opinion. The conclusion had already been reached. He was just building the file around it.

That number, six months, isn't a minor administrative gap. It's a systemic failure of professional standards. Imagine if fewer than 30% of agencies had documented procedures for handling DNA evidence, or chain of custody for physical forensics. There would be congressional hearings. There would be careers ended. But facial comparison technology, deployed at scale, still largely operates on vibes and departmental discretion.

Facial Recognition Accuracy vs. Investigator Responsibility

Let's talk about the footage itself for a second, because this is where the whole thing starts unraveling before a single algorithm even runs.

Research from MIT Media Lab and the National Institute of Standards and Technology (NIST) has documented that facial recognition algorithms show measurably higher error rates on women, darker-skinned individuals, and elderly subjects, with false positive rates in some demographic groups exceeding ten times those of the baseline population. Now layer on top of that: grainy CCTV footage. Low resolution. Poor lighting. Potentially shot from an unfavorable angle. Every single one of those variables compounds the others simultaneously.

Running a facial comparison on footage like that and then treating the output as operationally meaningful is like running a DNA test on a degraded sample, getting an inconclusive result, and writing it up as a match. The tool can only be as good as the input. Image quality isn't a technical footnote, it's the threshold question that determines whether the whole exercise is even worth doing. Previously in this series: Multimodal Biometrics Face Fingerprint Voice Defea.

"Police ran the surveillance footage through facial recognition software and determined she was the suspect. She said a detective then reviewed her driver's license and social media and confirmed she was indeed the suspect." As reported by 13ABC / KVLY, citing GoFundMe campaign documentation

Read that again. The detective "confirmed" the match by looking at the same woman's own photos. That's circular. You can't verify a facial comparison by running another facial comparison using materials you already believe belong to the suspect. That's not corroboration, that's echo.

Facial Recognition: Best Practices for Law Enforcement

Here's the thing: the technology itself has genuine value. It has solved cases that would otherwise go cold for decades. The cold case DNA work happening in parallel, like the Nebraska TV reports on two New York cold cases from the 1970s solved through forensic genealogy, is a useful parallel. DNA evidence is admissible and trusted not because DNA is magic, but because its methodology is documented, challenged, peer-reviewed, and hedged with quality standards. Every result comes with error rates. Every sample has chain of custody. No DA puts a DNA result in front of a jury without an expert who can explain exactly what the number means and what it doesn't.

Facial comparison deserves the same professional rigor. That means documented image quality assessment before any comparison is run. It means confidence score minimums, and transparency about what those scores actually represent statistically. It means mandatory independent corroboration: geographic evidence, behavioral evidence, financial records, witness accounts. Something that exists entirely outside the facial match chain.

Anyone who wants to understand what facial recognition software actually can and can't do under real-world conditions will find the gap between vendor marketing and operational reality considerably wider than most agencies assume.

The Investigators Who Will Survive This Wave of Scrutiny

Courts are paying attention. Defense attorneys are getting smarter. Civil rights organizations are documenting every case. The investigators who put their names on warrants built primarily on facial comparison hits, especially from degraded footage, especially without independent corroboration, are walking into a very bad few years.

The ones who will be fine are the ones who can show their work. Image quality assessment, documented. Confidence score, recorded with its statistical meaning explained. Corroborating evidence, independent and specific. A clear evidentiary chain demonstrating that the match was a starting point, and that everything after it was built on its own foundation. Up next: Face Images Personal Data Gdpr Pseudonymisation.

That's not an unreasonable bar. That's just methodology. That's just the job.

Look, nobody's saying this technology should be shelved. That would be a waste of a genuinely powerful investigative tool, and it would leave real victims without recourse. But right now, fewer than 30% of agencies using this technology have auditable protocols in place. That means the majority of facial comparison work happening in American law enforcement today operates without documented standards, without quality thresholds, and without mandatory corroboration requirements.

A Tennessee grandmother spent six months in a North Dakota jail because a fraud suspect used a fake military ID at a bank. She had nothing to do with it. She is home now, presumably, while somewhere a detective who "confirmed" her identity using her own social media profile is still carrying a badge.

So here's the question every working investigator should be sitting with right now: If a facial match contributed to an arrest on one of your cases, what minimum safeguards, image quality, confidence threshold, corroborating evidence, would you insist on before you put your name on that warrant? And if your honest answer is "whatever the software spits out plus a quick Google," you might want to think harder about that before the next case lands on your desk.

Recognition Accuracy Depends on Input Quality, Not Just the Algorithm

One of the real benefits of facial recognition in law enforcement is speed: a facial recognition system can scan thousands of images and surface a candidate in seconds, something no human reviewer could do at that scale. But recognition accuracy is only as good as the photo or video frame going in. When police feed a facial recognition tool a blurry, poorly lit still from bank surveillance, the recognition output should be treated as a weak lead, not a strong one. Agencies that document image quality before running facial recognition get far more reliable identification results than agencies that don't.

Facial Identification Works Best as a Starting Point, Not a Verdict

Facial identification is meant to narrow a pool of potential suspects, not to name a single person with certainty. When facial recognition technology returns a candidate, that identification still needs independent evidence, location data, financial records, or witness statements, before an arrest is justified. Treating facial identification as the finish line instead of the starting line is exactly how this Tennessee case went wrong. Departments that train officers on this distinction see far fewer wrongful arrests tied to facial recognition technology.

Digital Forensics Adds the Corroboration Facial Recognition Alone Cannot Provide

Digital forensics, phone records, transaction logs, IP addresses, device metadata, gives investigators something a facial recognition match cannot: evidence that exists independently of a person's face. When digital forensics and facial recognition point to the same suspect, confidence in an arrest rises considerably. When only facial recognition points to a suspect, as happened here, digital forensics should have been the very next step before any arrest warrant was signed.

Public Safety Depends on Getting Facial Recognition Deployment Right

Public safety is the whole justification for using facial recognition in law enforcement in the first place, the technology helps identify suspects in fraud, violent crime, and missing persons cases faster than manual review ever could. But public safety is also harmed every time a wrongful arrest happens, because it erodes public trust in the very tool that's supposed to protect communities. Agencies that want to keep using facial recognition technology to serve public safety need documented standards that prevent cases like this grandmother's from repeating.

Operational Outcomes Improve When Facial Recognition Has Guardrails

Operational outcomes for facial recognition programs, meaning arrests that hold up, convictions that stick, and public confidence that doesn't erode, depend directly on how disciplined an agency is about using the technology. Departments with confidence-score thresholds, mandatory corroboration rules, and image-quality checks report better operational outcomes than those relying on a facial recognition hit alone. The West Fargo case is a clear example of poor operational outcomes stemming from a shortcut methodology rather than a bad algorithm.

How Facial Recognition Can Help Law Enforcement Generate Leads Responsibly

Used correctly, facial recognition can help law enforcement generate leads in fraud rings, cold cases, and missing persons investigations where no other starting point exists. The technology's real value is in narrowing an impossibly large pool of possibilities down to a manageable handful of names for detectives to investigate further. The mistake isn't using facial recognition to generate leads, it's stopping there instead of treating the output as the first step in a longer, corroborated investigation.

Facial recognition technology, video analytics, and other biometric tools are becoming standard equipment in modern policing, and for good reason: they help identify individuals faster than any manual process could manage. FRT vendors market their products on match speed and claimed accuracy rates, but neither speed nor a headline accuracy number tells an officer whether a specific match, on a specific piece of grainy footage, is trustworthy enough to build an arrest around. That gap between marketed capability and real-world reliability is exactly where cases like this Tennessee grandmother's fall through.

Learn from this case, and the lesson isn't that facial recognition should be abandoned, it's that identity verification built on a single biometric match is inherently fragile. Identity is complex, and a face alone, especially on degraded footage, is a thin thread to hang an arrest on. The agencies that will still be using facial recognition technology confidently five years from now are the ones building identification workflows around corroboration today, not the ones hoping the next AI match happens to be right.