A camera records a crime.
The image is not particularly good.
The person's face is visible, but only for a moment. The lighting is poor. The angle is imperfect.
Years ago, an investigator might have circulated the photograph among officers or shown it to witnesses.
Today, there is another possibility.
The image can be submitted to facial-recognition software.
The computer compares characteristics of the face against photographs in a database and returns possible matches.
One of those matches can give investigators a name.
That can be an extraordinarily useful lead when police have little else to work with.
But there is a problem hidden inside the word possible.
A computer-generated match is not necessarily an identification.
And when investigators treat an algorithmic suggestion as though it were proof, the consequences can fall on someone who had nothing to do with the crime.
Robert Williams Was at Work When the Crime Happened
One of the most widely documented examples began with the theft of expensive watches from a Shinola store in Detroit.
Security cameras recorded the incident.
Detroit police obtained an image of the suspected thief and used facial-recognition technology in an effort to identify him.
The system produced a possible match: Robert Williams.
Williams was not the man in the surveillance footage.
But the algorithmic result became part of an investigation that eventually led police to his home.
In January 2020, officers arrested Williams in front of his wife and children.
He was detained for approximately 30 hours before being released.
The case against him was eventually dismissed.
Williams later sued, and in 2024 the litigation resulted in a settlement requiring Detroit police to adopt safeguards governing the use of facial-recognition technology. The American Civil Liberties Union, which represented Williams, said the agreement prohibited police from making arrests based solely on facial-recognition results and required additional investigative steps.
The case demonstrated something that would become increasingly difficult to ignore:
An algorithm can be wrong.
More importantly, people can make the algorithm's mistake much more consequential.
Facial Recognition Does Not Necessarily Say “This Is the Person”
The public description of facial recognition can make the technology sound more definitive than it is.
A photograph goes into the computer.
A name comes out.
But many law-enforcement facial-recognition systems are better understood as tools for generating investigative leads.
The system compares an unknown face with images in a database and identifies candidates that appear sufficiently similar under the system's mathematical analysis.
The result can help an investigator determine where to look next.
It does not necessarily establish that the candidate and the person in the original image are the same individual.
That distinction is critical.
A lead tells investigators where to investigate. Evidence determines whether the lead is correct.
The Original Image Can Determine How Much the Algorithm Has to Work With
Facial recognition cannot recover information that the camera never captured.
Imagine a perfect driver's-license photograph.
The person faces forward.
The image is sharp.
Lighting is even.
The face occupies much of the frame.
Now compare that with a convenience-store surveillance image.
The person is twenty feet from the camera.
The head is turned.
A hat covers part of the forehead.
The image is compressed.
Motion creates blur.
Those are very different inputs.
An algorithm may still produce candidates from a poor image.
But the existence of a candidate does not magically improve the quality of the original photograph.
Investigators still need to consider what the source image actually shows.
A Similar Face Can Become an Investigative Lead
Human faces share characteristics.
Two unrelated people can look remarkably similar.
Family members can look even more alike.
An algorithm converts facial characteristics into mathematical information and compares those characteristics with other images.
Depending on the system, investigators may receive one candidate or a group of possible candidates.
That can be useful.
Suppose police have no idea who appears in a robbery video.
A facial-recognition search returns several possibilities.
Investigators examine the first candidate and discover that he was in another state at the time.
They eliminate him.
The second candidate does not resemble the person when better photographs are examined.
They eliminate her.
The third candidate owns the vehicle seen leaving the crime scene and has additional connections to the investigation.
Now investigators have something worth pursuing.
In that example, the algorithm did what an investigative tool should do.
It generated possibilities.
The investigators tested them.
The Danger Begins When the Search Result Becomes the Conclusion
Now imagine a different process.
The computer returns a candidate.
Investigators see the name and begin interpreting everything else through that assumption.
A photograph of the candidate is shown to a witness.
The witness makes an identification.
Police seek an arrest warrant.
The suspect is arrested.
At each stage, the original algorithmic suggestion can gain apparent credibility.
The computer result influences the investigation.
The investigation produces additional evidence.
That additional evidence then appears to confirm the computer result.
If the original match was wrong, investigators can unintentionally build a case around the wrong person.
This is sometimes described more broadly as automation bias—the tendency to place excessive confidence in results produced by automated systems.
The computer did not make the arrest.
People did.
But the computer may have influenced where they looked.
A Witness Can Be Influenced by the Candidate Police Select
Eyewitness identification creates a particularly important problem.
Suppose a witness saw a robber for ten seconds.
Days later, police use facial recognition and obtain a possible match.
Investigators then place that person's photograph in a lineup.
The witness selects the person.
At first glance, that appears to be two pieces of evidence:
The algorithm identified the person.
The witness identified the person.
But those pieces of evidence are not completely independent.
The algorithm influenced which photograph police placed before the witness.
If the algorithm selected the wrong person, the witness has now been asked to choose from a process already shaped by that error.
That does not mean the witness identification is automatically invalid.
It means investigators and courts need to understand how the identification process developed.
An Arrest Warrant Does Not Make the Algorithm Correct
When police believe they have probable cause, they may seek an arrest warrant from a judge.
A judge's approval is an important constitutional safeguard.
But the judge evaluates the information presented in support of the warrant.
If investigators overstate the significance of a facial-recognition result, omit important limitations or present derivative evidence without explaining how it originated, the technological uncertainty can become difficult to see.
That is why transparency about the role of automated tools matters.
A statement that “the suspect was identified” can sound very different from:
“Facial-recognition software generated this person as a possible candidate, after which investigators conducted the following independent investigation.”
The details tell the real story.
Robert Williams Was Not the Only Person
Other documented wrongful-arrest cases have also involved facial-recognition technology.
In New Jersey, Nijeer Parks was arrested after facial-recognition technology contributed to police identifying him as a suspect in crimes committed in Woodbridge.
Parks denied being the person involved and later sued after the charges were dismissed.
His case became another prominent example in the national debate over law-enforcement use of facial recognition.
In Detroit, Porcha Woodruff was arrested in 2023 while eight months pregnant after facial-recognition technology contributed to an investigation that incorrectly identified her as a suspect in a robbery and carjacking case. The charges were later dismissed.
These incidents are significant not because every facial-recognition search produces a wrongful arrest.
They demonstrate what can happen when an incorrect candidate survives the investigative process instead of being eliminated by it.
The Technology Does Not Perform Equally in Every Situation
Facial-recognition accuracy depends on many variables.
Image quality matters.
Camera angle matters.
Lighting matters.
The algorithm matters.
The database matters.
The characteristics of the people being compared can matter.
Research by the National Institute of Standards and Technology has documented demographic differentials in the performance of many facial-recognition algorithms, although the magnitude and direction of those differences vary among algorithms and applications.
Technology also changes rapidly.
A finding about one algorithm at one point in time should not automatically be treated as a permanent conclusion about every facial-recognition system.
But neither should improved technology erase the need to verify individual results.
Even a highly accurate system can produce a false match.
When millions of comparisons are possible, rare errors can still involve real people.
Accuracy Rates Can Be Misleading Without Context
Imagine a hypothetical system that performs correctly 99.9 percent of the time under a particular test.
That sounds nearly perfect.
But the practical meaning depends on what is being measured.
Is the system verifying whether two controlled photographs depict the same person?
Is it searching one poor surveillance image against ten million photographs?
What threshold is being used?
How many candidates are returned?
What is the quality of the source image?
How often does the correct person even exist in the database?
Different tasks produce different kinds of errors.
A single accuracy percentage cannot answer all of those questions.
This is why real-world use matters as much as laboratory performance.
The Database Changes the Meaning of the Search
Facial recognition needs photographs to compare.
Those images can come from different sources depending on the system and jurisdiction.
A database might contain booking photographs.
Another might include driver's-license or identification photographs where legally permitted.
A private system might use a completely different collection of images.
The composition of the database affects who can appear as a candidate.
If a person's photograph is not in the searchable collection, the system cannot return that exact photograph as a match.
If millions of people are included, the system has millions of opportunities to find faces sharing characteristics with the unknown image.
The database therefore matters almost as much as the algorithm.
Facial Recognition Can Also Exclude People
The technology is not useful only for finding suspects.
Suppose an investigator initially suspects someone because of a witness statement.
A careful comparison with high-quality video may demonstrate obvious differences.
Other digital evidence may place the person elsewhere.
Automated tools can sometimes help investigators prioritize or eliminate possibilities.
The important point is that an algorithmic result should be treated according to what it actually establishes.
Technology can assist an investigation without being allowed to dictate its conclusion.
Humans Can Make the Same Kind of Mistake
It would be wrong to frame the issue as accurate humans versus inaccurate computers.
Human identification has its own long history of error.
Witnesses can misremember faces.
Investigators can focus on the wrong suspect.
People can be influenced by expectations.
Photographs can be misleading.
Two strangers can look alike.
Facial recognition was developed partly because computers can perform comparisons at a scale humans cannot.
The relevant question is not whether humans or algorithms are perfect.
Neither is.
The question is how to build an investigative process that recognizes the limitations of both.
Independent Corroboration Can Stop a Bad Match From Becoming a Bad Arrest
Return to the original example.
An algorithm identifies a possible suspect in a robbery.
What happens next determines how dangerous the error is.
Investigators can ask where the person was when the crime occurred.
They can compare height, age and other observable characteristics.
They can examine vehicles.
They can seek video from other locations.
They can interview witnesses without improperly suggesting the answer.
They can investigate whether the person has any actual connection to the event.
They can look for evidence inconsistent with the algorithmic lead.
If the candidate was hundreds of miles away, that should matter.
If the candidate has no connection to the vehicle, location or people involved, that should matter.
If better images show obvious differences, that should matter.
Independent investigation turns an algorithmic suggestion into something that can be tested.
The Algorithm May Be Invisible to the Person Being Accused
There is another complication.
A person arrested after a technology-assisted investigation may initially have no idea that an algorithm played any role.
From the defendant's perspective, police simply arrived with an accusation.
But understanding how the investigation began can be important.
Was facial recognition used?
Which image was submitted?
Which system performed the search?
How many candidates were returned?
What did investigators know before the search?
What did they learn independently afterward?
Was a witness shown a photograph because the algorithm selected that person?
Those questions can matter when attorneys and courts evaluate how evidence was developed.
Artificial Intelligence Will Make This Problem Much Larger Than Faces
Facial recognition is only one example of algorithm-assisted decision-making.
Automated systems can help analyze license plates, fingerprints, financial transactions, communications, documents and enormous collections of digital information.
Artificial intelligence can identify patterns that would take humans far longer to find.
That creates extraordinary investigative potential.
It also creates a recurring problem.
When software produces an answer, people naturally want to know whether the answer is correct.
But another question can be just as important:
What exactly did the software determine?
A risk score is not a fact.
A similarity score is not an eyewitness.
A statistical correlation is not necessarily causation.
A generated candidate is not necessarily the perpetrator.
Understanding the output is essential before deciding how much weight to give it.
An Algorithm Cannot Be Cross-Examined Like an Eyewitness
Traditional evidence often comes with a human source.
A witness takes the stand.
Attorneys can ask what the witness saw.
Where was the witness standing?
How long did the observation last?
Was it dark?
How certain is the witness?
Software creates a different challenge.
The meaningful questions may concern training data, error rates, thresholds, source-image quality, database composition and how investigators interpreted the result.
Some systems are proprietary.
Their internal operation may not be obvious to the people affected by their output.
That can make transparency and documentation particularly important when automated analysis contributes to consequential government decisions.
The Best Safeguard May Be Treating the Computer as the Beginning
Facial recognition can solve a problem that once seemed almost impossible.
Investigators can take an unknown face from a video and rapidly compare it with enormous collections of photographs.
That ability can generate leads in cases that otherwise might go nowhere.
But the same scale that makes the technology powerful also makes mistakes consequential.
A false candidate is not merely a wrong answer on a screen.
It is a real person.
That person may be investigated.
Questioned.
Placed in a lineup.
Searched.
Arrested.
Or forced to prove that the computer pointed investigators toward the wrong life.
The central lesson from documented wrongful-identification cases is therefore not that algorithms have no place in investigations.
It is that an algorithmic match should not be allowed to become a substitute for investigation.
The computer can say:
Look here.
It is still the job of investigators to determine whether looking there reveals evidence—or reveals a mistake