AI Post 5 (AI Ethics): The Man Arrested While Driving Home: A True Story of AI Gone Wrong

An algorithm said he was guilty. He spent 30 hours in jail. He was innocent.

January 2020. Detroit, Michigan.

Robert Williams is driving home from work on a normal Thursday afternoon. He pulls into his driveway. Before he can even step out of the car, police vehicles surround him. He was arrested. Handcuffed and taken away.

The charge: Stealing expensive watches from a Shinola store.

The evidence: An AI facial recognition system matched his face to grainy surveillance footage.

The problem: Robert Williams had never been to that store. He didn’t steal anything. He was completely innocent.

But he still spent 30 hours in a jail cell.

This is a true story. And it’s happened to at least a dozen other people since.

Let’s talk about what went wrong — and what it means for all of us.

What Actually Happened

Let me walk you through the timeline.

The Crime (2018)

Someone walked into a Shinola store in Detroit and stole several expensive watches.

The store had surveillance cameras. But the footage was grainy. Low quality. Poor lighting. The suspect wore a hat that partially covered his face.

Police had video of the crime. But no clear suspect. No leads.

The case went cold.

The AI “Solution” (2019)

Detroit Police decided to use technology to crack the case.

They took a still image from the blurry surveillance video. They sent it to the Michigan State Police, who ran it through their facial recognition system.

The system searched a database of millions of faces — mostly driver’s license photos.

It returned dozens of possible matches. People whose faces looked somewhat similar to the blurry image.

From those dozens of possibilities, Detroit Police picked one: Robert Williams.

Why him? His photo came up from an expired driver’s license.

Interesting detail: His current, up-to-date driver’s license photo didn’t even match. Only the old, expired one did.

The Investigation

Here’s what Detroit Police did next:

They showed witnesses a photo lineup. Six faces. Five were random fillers. One was Robert Williams — the face the AI had selected.

The witnesses looked at the lineup. They picked Williams. “That’s him.”

Sounds like confirmation, right?

Except there’s a problem. When you show witnesses a lineup where one person was specifically chosen by a computer algorithm and the others are just random fillers who look less like the suspect, people naturally gravitate toward the AI-selected face.

It’s called “lineup bias.” The AI’s choice influences the human decision.

Police got their witness identification. They applied for an arrest warrant. Probably some traditional detective work was missing. They trusted the AI.

The Aftermath

Prosecutors eventually dropped the charges.

Williams was released after 30 hours in custody.

But the story doesn’t end there.

“It doesn’t just affect the person who is arrested,” Williams said in an interview. “I have a whole family, and they were also affected by this.”

Before this happened, Williams would have said he supported the use of facial recognition technology.

After? “They have a long way to go because there are so many ways that it could go wrong.”

Why This Happened: The Four Critical Failures

Let’s break down exactly what went wrong. Because this wasn’t one mistake. It was a cascade of failures.

Failure #1: Garbage In, Garbage Out

Remember from Post 2: AI is only as good as the data you feed it.

The “data” in this case was:

• A grainy, low-quality surveillance image
• Poor lighting
• Subject wearing a hat, face partially obscured
• Converted to a still frame from video

This is terrible input for facial recognition.

The AI system even returned dozens of possible matches — a sign that the confidence was low, the image was unclear, the results were questionable.

Failure #2: AI Bias

Here’s something critical: Facial recognition technology is significantly less accurate for people of color.

Federal testing by the National Institute of Standards and Technology in 2019 found that facial recognition systems were up to 100 times more likely to misidentify Black and Asian faces compared to white faces.

Why?

Because the training data used to teach these systems was historically skewed toward white faces. When you train an AI predominantly on one type of face, it gets really good at recognizing that type — and worse at recognizing others. Robert Williams is Black.

The technology that identified him was statistically more likely to get his face wrong than if he had been white. This isn’t a bug. It’s a well-documented feature of how these systems work when trained on biased data.

Failure #3: Over-Reliance on AI

Detroit Police treated the facial recognition match as if it were proof.

But here’s what facial recognition systems are designed to do: provide investigative leads. Narrow down possibilities. Give detectives a starting point for further investigation. They are not designed to be the sole basis for an arrest.

Failure #4: The System Amplified the Error

Once the AI made its initial mistake, every step afterward made it worse.

The photo lineup was tainted. When you show witnesses a lineup where one face was specifically chosen by AI (and the others are random fillers), you’re essentially asking witnesses to confirm the AI’s choice. It’s circular reasoning.

This Isn’t an Isolated Incident

Robert Williams’ case was the first publicly documented instance of a facial recognition false arrest. But it wasn’t the last.

As of 2025, there are at least 12 known cases of wrongful arrests due to facial recognition technology in the United States.

The Settlement: A Step Toward Accountability

In June 2024, the City of Detroit reached a landmark settlement with Robert Williams. Detroit agreed to pay Williams $300,000.But more importantly, they agreed to change their policies.

The New Detroit Police Facial Recognition Policy

What changed:

1. Police cannot arrest someone based solely on facial recognition results. They must have independent evidence linking the person to the crime.
2. Police cannot conduct photo lineups directly after a facial recognition match without other reliable evidence. This prevents the circular reasoning that happened in Williams’ case.
3. All officers must receive training on facial recognition technology. Including its risks, limitations, and the documented higher error rates for people of color.
4. Audit of all past cases. The city agreed to review every case since 2017 where facial recognition was used to obtain an arrest warrant.

Civil rights advocates called it the strongest police facial recognition policy in the nation.

What This Story Teaches Us

Robert Williams’ story isn’t just about one wrongful arrest.

It’s a case study in what happens when we deploy powerful AI systems without proper safeguards.

Lesson 1: AI Amplifies Existing Inequalities

Lesson 2: AI Should Assist Decisions, Not Make Them

Lesson 3: Transparency Is Essential

Lesson 4: Accountability Matters

The Bigger Picture

New technology making advances by leaps and bounds is awesome. It should be used. But -

This is about using technology responsibly.

It’s about understanding its limitations.

It’s about protecting people from its failures.

It’s about ensuring that when powerful tools are deployed, there are safeguards, oversight, and accountability.

AI can be powerful. But it’s not infallible. It’s not neutral. And it’s not fair by default.

When it fails, real people pay the price.

What’s Next

In the next post, we’ll explore the deeper systemic issues behind Robert Williams’ story:

“AI Ethics: Bias, Privacy, Accountability, and What We Can Do About It”

We’ll dive into:

• How bias gets baked into AI systems from the very beginning
• What companies and governments actually do with the data they collect
• Who should be held accountable when AI causes harm
• What “fairness” in AI really means (and why it’s complicated)
• How we as a society can push for more ethical AI

Stay curious. Stay critical. Stay engaged.

Sources and Further Reading

Primary Sources on Robert Williams Case:

• ACLU case page: Williams v. City of Detroit (filed April 2021, settled June 2024)
• Settlement agreement announced June 2024, Detroit pays $300,000 and implements new facial recognition policies
• University of Michigan Law School Civil Rights Litigation Initiative documentation

Recent Wrongful Arrest Cases:

• Angela Lipps (Tennessee/North Dakota, 2024–2025) — CNN, Fargo Police Department statements, March 2026
• Trevis Williams (New York, 2025) — ABC7 New York, August 2025
• Jason Killinger (Nevada, 2023–2025) — Reno Gazette-Journal, Futurism, IBTimes UK
• Jason Vernau (Miami, 2024) — Washington Post investigation, January 2025
• Porcha Woodruff, Michael Oliver, Nijeer Parks — ACLU documentation and news reports

Facial Recognition Accuracy and Bias Research:

• National Institute of Standards and Technology (NIST) 2019 study on demographic differentials in facial recognition accuracy
• Federal testing showing up to 100x higher false match rates for Black and Asian faces compared to white faces

Policy and Advocacy:

• ACLU report: “More than a Dozen Wrongful Arrests Due to Police Reliance on Facial Recognition Technology” (April 2026)
• Innocence Project: “AI and The Risk of Wrongful Convictions in the U.S.” (January 2025)
• Washington Post investigation: “Arrested by AI: Police ignore standards after facial recognition matches” (January 2025)

Cities that have banned police facial recognition:

• San Francisco, Boston, Portland (OR), Minneapolis, and 20+ other U.S. jurisdictions

All facts, quotes, and case details in this post are drawn from verified news reports, court documents, and official statements from law enforcement and civil rights organizations.