Face in the Crowd — Recognition & Algorithmic Bias

Welcome, Investigators!

A camera catches a face. A computer says "92% match." Case closed? Not so fast. Today you'll open up how facial recognition actually works, audit where it gets less reliable, and decide what a match really earns you in an investigation. The answer will surprise you. Follow the evidence!

The Case

A convenience-store camera captured the face of someone the police want to identify. Investigators run the image through a facial-recognition system, which returns a ranked list of candidates with confidence scores. The top candidate scores high.

But a good investigator asks harder questions. How did the system reach that score? And is it equally reliable for everyone? Independent testing has shown that many facial-recognition systems are less accurate for some demographic groups than others — which means a confidence score can mean different things for different faces. Your job: trace the pipeline that produced the match, run an audit of its accuracy across groups, and reach a defensible judgment about what this match should — and should not — be used for.

Learning Objectives

By the end of this investigation you will be able to:

1. Describe the four stages of a facial-recognition pipeline in order.
2. Interpret a match confidence score and a candidate ranking.
3. Analyze audit data showing accuracy differences across demographic groups.
4. Evaluate whether a face match meets the bar for evidence versus a lead.

Quick Facts

Lab type	💻 Virtual
Group size	2–3 investigators
Time	45–55 minutes
Cost	$0 — computer-based
Ties to	Ch 16 — Facial Recognition Overview, Face Detection Algorithms, Facial Landmark Detection, Algorithmic Bias, Facial Recognition Admissibility, CCTV Surveillance Analysis

Materials

Per group:

• None — a laptop and a browser.
• The instructor's provided sample probe image, candidate gallery, and the audit data table.
• Assumes access to one computer per group.

Handle This Evidence — and This Topic — With Care

• A facial-recognition match is an investigative lead, not proof of identity. Several people have been wrongly arrested after an over-trusted match — real cases with real harm. Treat a match as a hypothesis to verify.
• The accuracy differences you'll examine are a documented, technical fact about certain systems — not a statement about any group of people. Discuss the technology's limits, respectfully and factually.
• Use only the provided sample data. Running real people's faces through recognition tools without authority raises serious privacy and civil-rights concerns.

Background: How a Face Becomes a Match

Facial recognition isn't magic and it isn't a photograph comparison. It's a pipeline of four stages, and understanding them is the key to judging a result.

1. Detect — the system first finds that there is a face in the image and boxes it. Poor lighting, angle, or low-resolution CCTV can make even this first step fail.
2. Landmark — it locates key points on the face (eye corners, nose tip, mouth edges), which lets it align and normalize the face for comparison.
3. Encode — it converts the aligned face into a string of numbers, a faceprint or embedding, that summarizes the face's measurable features.
4. Match — it compares that faceprint against a gallery and returns the closest candidates, each with a confidence score and a rank.

Here's the part that matters most for justice. The U.S. National Institute of Standards and Technology (NIST) and other researchers have tested many systems and found that accuracy is not uniform — false-match rates can be substantially higher for some demographic groups (by skin tone, sex, and age) than others. A "92% match" is not an equally reliable number for every face. That is algorithmic bias: a system performing unevenly across groups, usually because of what data it was trained on. It's the central reason courts treat face matches cautiously and why a match should launch an investigation, never end one.

Explore: The Facial-Recognition Pipeline

Facial-Recognition Pipeline Interactive MicroSim

Type: microsim
sim-id: facial-recognition-pipeline
Library: p5.js
Status: Specified

Learning Objective: Trace a face through the detect → landmark → encode → match pipeline and read the resulting confidence scores, then evaluate accuracy across demographic groups (Bloom Level 5 — Evaluate).

Step the probe image through all four stages and watch a face turn into numbers, then into a ranked candidate list. Then switch to the audit view and compare the system's accuracy across groups — that comparison is the heart of this lab.

Procedure

Part 1 — Trace the pipeline.

1. Load the probe image into the pipeline MicroSim. Step through detect → landmark → encode → match, noting what each stage does.
2. Record the top candidate and its confidence score, plus the 2nd and 3rd candidates and their scores.
3. Note the gap between the top score and the next — a small gap means the system isn't very sure which candidate is right.

Part 2 — Run the audit.

1. Open the audit data. For each demographic group, record the system's reported accuracy (or false-match rate) from the table.
2. Identify the groups where the system is most and least accurate.
3. Re-examine your Part 1 match: which group does the probe fall into, and how reliable is the system for that group?

Part 3 — Reach a judgment.

1. Weigh the confidence score together with the group-specific accuracy. A high score in a low-accuracy group is weaker than it looks.
2. Decide what this match justifies: closing the case, making an arrest, or only opening a line of inquiry that needs independent corroboration.
3. Write your admissibility judgment — is this fit to present as evidence, or is it a lead to verify?

Data Collection

Item	Value / observation
Top candidate & confidence score
2nd / 3rd candidate scores
Gap between #1 and #2
Most-accurate group (audit)
Least-accurate group (audit)
Probe's group & system accuracy there

Analysis Questions

1. List the four pipeline stages in order and say, in one line each, what could go wrong at each stage with a low-quality CCTV image.
2. The top candidate scored high. Given the audit data for that group, how much should that score reassure you? Explain.
3. Define algorithmic bias in your own words, and explain one reason a system might be less accurate for some groups (hint: think about training data).
4. Should this match be used to arrest the top candidate? Argue your position using both the confidence score and the group-specific accuracy.
5. Why do courts and agencies generally treat a face match as an investigative lead rather than as identification evidence? Name one piece of independent evidence that could turn a lead into a solid case.

Deliverable

Turn in a Face-Match Assessment: the pipeline result (top candidates and scores), the audit findings (most- and least-accurate groups), and a written admissibility judgment that states plainly whether the match is fit to be treated as evidence or only as a lead — and why, citing both the score and the bias data.

What Does the Data Tell Us?

A confidence score is a number, not a verdict. The careful investigator reads it alongside how reliable the system is for that kind of face — and then treats even a strong match as a starting point to corroborate. Overtrusting a match has put innocent people in handcuffs. Precision here protects people.

Extension Challenge: The Reform Question

Since the 2016 PCAST report and NIST's demographic testing, agencies and courts have proposed rules for facial recognition — human review requirements, accuracy thresholds, disclosure to defendants, even outright bans in some cities. Research one real policy or reform, and write a short position: does it adequately address the bias problem you saw in this lab? What would you add?

Teacher Notes

Setup, timing, and grading (click to expand)

• Prep: Provide a probe image, a small candidate gallery with confidence scores (make the top-vs-second gap modest so certainty is debatable), and an audit table showing clearly uneven accuracy across groups. Frame the audit data as a property of the system, grounded in real NIST-style findings.
• Set the tone. Open by naming that this is a factual, technical topic with real civil-rights stakes; keep discussion focused on the technology's measured limits, not on people. Reference real wrongful-arrest cases if your class is ready for them.
• Differentiation: For a shorter run, provide the pipeline result and focus on the audit + judgment. For a challenge, give two probes from different groups and have students explain why identical scores don't carry identical weight.
• Assessment focus: Reward judgments that combine the score with the group-specific accuracy and that correctly land on "lead, verify further" rather than "identified." The reasoning matters more than the verdict.

Case Closed — For Now

You looked past a confident-looking number and asked the question that actually protects people: how reliable is this, and for whom? That's not just good forensics — it's the kind of careful thinking that keeps the science honest and the innocent free. Follow the evidence!