From CCTV Still to Court-Ready Facial Comparison

Here's something that should keep any investigator up at night: the most dangerous output a facial recognition system can produce isn't a wrong answer. It's a confident-sounding wrong answer delivered to someone who doesn't know how to interrogate it.

That's exactly what happened in cases documented across multiple jurisdictions — including a deeply troubling investigation by The Wire and the Pulitzer Center, where individuals in Delhi were arrested solely on the basis of facial recognition matches without solid corroborating evidence. One man named Ali spent more than four and a half years in pre-trial incarceration before being granted bail. Four and a half years. For a match a machine made.

The problem wasn't the technology. The problem was treating the technology's output as a conclusion rather than a starting point. So let's talk about what disciplined facial comparison actually looks like — the kind that holds up under cross-examination, earns a judge's respect, and protects both the investigation and the innocent.

What the Algorithm Is Actually Telling You (And What It Isn't)

Most people imagine facial recognition as something close to magic — the AI either "recognizes" a face or it doesn't, binary and final. That mental model is wrong, and dangerously so.

What modern facial comparison algorithms actually do is convert facial geometry into a 128-dimensional numerical vector — a long string of numbers representing precise distances between landmarks like pupillary separation, nasal bridge width, jaw angle, and the distance from the outer corner of one eye to the tip of the nose. When two faces are compared, the system calculates the Euclidean distance between their respective vectors. Fall below a defined threshold? You get a match flag. Sit just above it? You get a "possible match." Sit well above it? The system rules it out.

Here's where it gets interesting. That threshold — the line between "match" and "possible match" — is a design choice, not a law of physics. Engineers set it based on acceptable false-positive and false-negative rates for their intended use case. Which means a "possible match" returned from crowd footage isn't the system saying "this is probably the person." It's the system saying "the vector distance puts this in the range where we can't confidently rule it out." That's a very different statement. This article is part of a series — start with Deepfake Detection Accuracy Gap Investigator Workf.

"An investigation by The Wire and the Pulitzer Center uncovered troubling instances where individuals were arrested solely on the basis of facial recognition — without solid corroborating evidence or credible public witness testimonies." — Astha Savyasachi, Pulitzer Center

The real kicker? Research from NIST's Face Recognition Vendor Testing program shows that the highest-risk outputs aren't strong matches or clear non-matches. They're those mid-range similarity scores — the "possible match" zone — where confirmation bias is most likely to push an untrained reviewer toward a false conclusion. You see a face that looks like the suspect, the system flags it as possible, and suddenly your brain is filling in gaps the data never actually supported.

The Image Quality Problem Is Worse Than You Think

Let's walk through a concrete scenario. A solo PI is working a fraud case. She pulls a CCTV still from a parking garage — grainy, off-angle, taken on a camera that hasn't been maintained since 2019. She runs it through a facial comparison platform and gets back three "possible matches" from a database of known associates. Now what?

Before she touches those three names, she needs to answer one question: is the source image even interpretable?

NIST research demonstrates that facial comparison accuracy degrades non-linearly with image resolution. A face captured at 24 pixels between the eyes performs dramatically worse than one captured at 90 or more pixels. That's not a gentle slide — it's a cliff. And if the PI doesn't document the source image quality as part of her evidentiary chain, the entire comparison result becomes technically uninterpretable in court. A defense attorney worth their fee will ask exactly this question on cross.

So step one, always: measure and document the inter-pupillary pixel distance of the source image. If it's below the reliable threshold, note that explicitly. The comparison can still proceed — but the confidence weight assigned to the algorithmic output must reflect the degraded input quality. That's not pessimism. That's methodology.

The Three-Step Discipline That Turns a Lead Into Evidence

Here's the framework that separates investigators who build cases from investigators who accidentally destroy them.

Step 1: Low-Quality Still → Documented Source Assessment

Before running any comparison, the investigator logs everything about the source image: capture date and time, camera specifications if available, lighting conditions, estimated angle deviation from frontal view, and — critically — the inter-pupillary pixel distance. This isn't bureaucratic box-ticking. It's the foundation that allows every downstream finding to be defended. If the source image is compromised, that gets noted. The comparison proceeds with appropriate caveats, not inflated confidence. Previously in this series: Why Second Facial Match Result Matters More.

Understanding the technical limitations of face recognition software under real-world conditions is what separates an investigator who gets embarrassed in court from one who gets convictions.

Step 2: Structured Comparison → Feature Alignment and Dual-Method Analysis

This is where the discipline lives. The investigator doesn't just accept the algorithmic similarity score — they perform a structured manual feature alignment alongside it. That means placing the source image and the candidate image side by side, annotating corresponding landmarks (brow ridge, alar base width, philtrum length, ear morphology where visible), and documenting agreements and discrepancies independently of the score.

Why both? Because NIST's FRVT program research shows that trained forensic facial examiners catch errors that algorithms miss — and vice versa. Algorithms struggle with partial occlusion and unusual lighting angles. Human examiners struggle with systematic biases and unconscious pattern-completion. The combination — algorithmic similarity scoring plus structured manual feature alignment — achieves a documented dual-methodology that neither approach reaches alone. That's the standard that survives cross-examination.

Think of it like the GPS analogy: a GPS gives you a coordinate. A navigator checks the terrain, cross-references landmarks, and confirms the route before committing. The algorithm's output is the GPS coordinate. The structured comparison is the navigation.

Step 3: Court-Ready Report → The Documented Chain

The final output isn't a name circled on a printout. It's a structured report that contains: the source image assessment and its quality limitations, the algorithmic similarity score with the platform's stated confidence parameters, the manual feature alignment findings (agreements, discrepancies, and any features inconclusive due to image quality), and a clearly stated conclusion — not "this is the person," but "the comparison is consistent with" or "the comparison does not exclude" the candidate, with a documented rationale either way.

Brazil's Polícia Civil do Distrito Federal has demonstrated what disciplined biometric methodology looks like at scale — achieving a 99 percent positive identification rate in forensic examinations by combining fingerprint analysis, face biometrics, and advanced latent print analysis as an integrated, documented system — not as isolated algorithmic outputs. The lesson isn't that their technology is better. It's that their methodology treats each tool as one documented piece of a chain, not an oracle.

What Gets You in Trouble vs. What Gets You a Win

Look, nobody's saying this is simple. Low-quality footage is the norm, not the exception. Time pressure is real. And AI outputs feel authoritative in a way that's genuinely hard to resist — the number comes back, it looks official, and there's a name attached. That pull toward early closure is one of the most well-documented cognitive traps in investigative work. Up next: Face Match Is A Lead Not A Verdict.

But here's what the wrongful arrest cases have in common: the investigator stopped at the algorithm's output. The comparison result became the conclusion, rather than the first step in a structured process. ABC7 New York's coverage of wrongful NYPD arrests linked to facial recognition tech shows exactly this pattern — a match was surfaced, and instead of triggering a disciplined evidentiary chain, it triggered an arrest.

Meanwhile, the investigators who win — who build cases that hold up, who either definitively rule someone out or establish genuine probable cause — treat the facial comparison hit as what it actually is: a statistically significant narrowing of the candidate field, requiring structured verification before it earns any weight in the chain of evidence.

Facial recognition technology measures up to 68 distinct facial datapoints to build a comparison, as noted by The Regulatory Review. That's remarkable precision — when the source image quality supports it, and when a trained examiner documents what those 68 points actually show. Without that documentation, it's just a number.

So here's the question worth sitting with — not just for facial comparison, but for every automated tool that returns a confident-looking output: when you get a "possible match" from any technology, what's the very next manual step you take before you're willing to put your name on it?

Because the answer to that question is the difference between an investigator who uses AI as a powerful tool and one who uses it as a shortcut. Shortcuts don't survive cross-examination. Methodology does.