Police in several American cases fed a photo of a suspect, often a grainy still from a security camera, into face-recognition software. The software returned one lookalike. Police then showed a witness a lineup of six photos that included that lookalike, and the witness picked him.
The people arrested say that picking proves nothing. The computer had already chosen the face the witness was shown, so the witness may only have agreed with the computer. Police say a search result is a tip, not proof, and that a lineup is a reasonable next step. The dispute is over whether a lineup built around the computer's pick counts as a real check.
A man who says he had never been to the city
Robert Dillon lives in Fort Myers, Florida. He says he had never been to Jacksonville Beach, more than 300 miles away. On November 2, 2023, a man there tried to lure a child into a McDonald's bathroom 12.
According to Dillon's federal lawsuit, a Jacksonville Beach officer sent surveillance images to the Jacksonville Sheriff's Office. An analyst there ran them through a Florida face-search system and relayed that they showed a "93% match" to Dillon 2. The lawsuit says that number misleads. A similarity score is not a probability, it argues, and 93% does not mean a 93% chance of the right man 2.
Next came a lineup, on January 10, 2024. The officer asked for a six-photo array with Dillon as the person of interest. It was shown to a restaurant employee who, the lawsuit says, had not seen the suspect's encounter with the child. She picked Dillon, writing that she was "100% percent sure" 2.
The lawsuit's logic runs like this. A lineup is a lookalike contest: the suspect's photo sits among fillers chosen to resemble him. When a computer supplies the suspect, it has supplied the face that most resembles the culprit, and the fillers resemble that face, not the culprit. On that account, the witness's choice can echo the computer's selection rather than independently verify it 2.
The case then went quiet for months. The lawsuit says the record has no entries between January 19 and July 31, 2024. Dillon was arrested at home on August 26. The warrant application, the lawsuit says, left out that a license-plate-reader search had found neither of his vehicles in the county on the days around the incident. It also left out that the witness had not seen the encounter, and called the lineup "sequential" although she was shown all six photos at once after seeing them one at a time. The State Attorney's Office dropped all charges on October 24, 2024, 69 days after the arrest. Dillon says it took nearly a year to clear the arrest from his record 21.
Some officials have criticized the practice. The lawsuit quotes Jacksonville's sheriff: "If you came to me with a facial recognition hit and that was your probable cause, I would probably kick you out of my office because that's not how it works." The Pinellas County Sheriff's Office, which runs the database, calls the training claim "patently false," says its training is clear that results are never matches, and says the Jacksonville Beach officer who treated a photo's appearance as probable cause bears the responsibility. Jacksonville Beach police and the sheriff's office told ABC News they could not comment 21. The case is pending and nothing in it has been decided.
A woman jailed for six months, with no lineup at all
Not every case involves a lineup. Kimberlee Williams was riding along on her daughter's DoorDash delivery to a military base in Lawton, Oklahoma, on June 23, 2021, when a security check found a Maryland arrest warrant for her. She says she had never been to Maryland. Prosecutors dropped the last charges that December, after she had spent about six months in jail 34.
The ACLU's account of how it began runs through an email list. A bank's fraud investigator posted a surveillance photo of someone making fraudulent withdrawals to CrimeDex, a listserv used by police and private investigators. Someone on it ran a face search and sent back Williams's name. The investigator then passed the lead to detectives in three Maryland counties 34.
The ACLU says detectives did little more than compare her old arrest photos with the bank's stills and judge them alike. Its letter to Anne Arundel County police argues that such a check can only confirm that the computer did its job, which is to find someone who looks like the suspect. It says nobody checked where Williams was 53.
The police departments have said little. Montgomery County declined to comment, citing an active investigation. Anne Arundel County said it checks outside tips before seeking charges. Prince George's County did not respond to the reporter's questions 4. The ACLU filed administrative complaints on April 14, 2026, asking for apologies and policy changes. No court has ruled.
The Detroit case a judge chose not to decide
In Detroit, a lineup did follow. In February 2023, an investigator searched a gas-station camera still of a woman wanted in a carjacking. The search produced 73 candidates. He narrowed them to one lead, Porcha Woodruff, and two colleagues reviewed that choice. The carjacking victim then picked her photo from a six-photo lineup. The lineup used an eight-year-old mug shot of Woodruff 6.
She was arrested on February 16, 2023, while eight months pregnant. At booking, the officer who had sought the warrant realized Woodruff, visibly pregnant, could not be the woman on the video. Woodruff was released that evening on a $100,000 personal bond, and the charges were dismissed on March 7 6.
She sued, and a ruling might have settled the lineup question. It did not. In August 2025 a federal judge granted the officer summary judgment, finding that Woodruff had not shown the warrant lacked probable cause. The ACLU, as a friend of the court, had argued that a face-search lead plus a lineup cannot support an arrest. Woodruff told the court she was not adopting that argument, so the judge said she could not consider it. She called the amici's points about the technology's limits compelling. She added that the officer would likely have qualified immunity anyway, because no case put before the court had given her fair warning 6.
So one of the best-documented lineup cases ended without an answer to whether a lineup built around a lead can count as the next step.
What the written rules now say
Two places with written rules have drawn similar lines.
Detroit's rules come from the settlement of a different lawsuit, by Robert Williams, whom the ACLU says was also wrongly arrested after a face-search lead. The ACLU's summary of the agreement says a lineup may never be conducted based solely on a face-search lead, and a lead plus a lineup identification may never be enough to seek an arrest warrant. A detective must document independent steps and get sign-off from two supervisors. A lineup may not reuse the photo the search returned, and witnesses may not be told that a computer identified anyone. A federal court can enforce the agreement for four years 7. We read the ACLU's summary, not the agreement itself.
Maryland's statute, passed in 2024 4, says a face-search result may not be the sole basis for probable cause, which needs additional, independently obtained evidence 8. A trained person must independently verify a result before it is used for any purpose in an investigation, and police may not tell a witness beforehand that a face search picked the suspect 9. The law took effect after Williams's arrest. The Banner reports it does not specify how a match must be verified 4.
Florida's side is harder to read. Dillon's lawsuit says Jacksonville Beach's written policy calls results investigative leads but names no corroboration steps 2. A rule on paper tells you what an agency intended. It does not show what its officers did.
A witness who picks the computer's lookalike may be echoing the computer.
How common is this, and what we cannot know
Dillon's lawyers count 15 publicly known wrongful arrests in the United States after face-search results, and say more than half involved a photo array that included the false match 2. The ACLU told The Baltimore Banner that Williams was the 14th it knew of 4. Both counts come from the group that litigates these cases and include only cases that became public. Nobody has published how many searches police run, or how many lead to correct arrests, so a failure rate cannot be computed from this evidence.
Why does a lookalike turn up at all? The lawsuit argues that face searches are designed to return the closest candidates in the database even when the culprit is not in it 2. In 2019 the National Institute of Standards and Technology evaluated 189 mostly commercial algorithms on 18.27 million images of 8.49 million people. It found false-positive rates that often varied across demographic groups by factors of 10 to beyond 100 10. That test used travel, immigration and law enforcement photo sets and says nothing about the searches in these cases.
What would settle it
The question is still open in court: can a lineup built around a face-search lead supply the independent evidence a warrant needs? Woodruff's judge declined to answer it. Dillon's pending lawsuit may be where it gets decided. Three things would help the public judge the earlier cases. One is the full candidate list from Dillon's search. Another is the fillers in his array. The third is how many people the Detroit and Maryland rules have kept out of jail.
References
- Hipos, R. L., & Yu, Y.-J. (2026, June 12). Man sues law enforcement alleging AI facial recognition technology led to wrongful arrest.
- Robert Dillon (plaintiff). (2026, July 13). First Amended Complaint, Dillon v. City of Jacksonville Beach, No. 2:26-cv-1936 (M.D. Fla.). https://assets.aclu.org/live/uploads/2026/06/35-First-Amended-Complaint.pdf
- American Civil Liberties Union. (2026). Kimberlee Williams Wrongful Arrest. https://www.aclu.org/cases/kimberlee-williams-wrongful-arrest
- Conarck, B. (2026, April 22). A woman spent months in Maryland jails because of unchecked facial recognition technology. https://www.thebanner.com/community/criminal-justice/aclu-facial-recognition-maryland-police-YXUU7AXEMNFIHKHJ5FZB2RLJPA
- American Civil Liberties Union of Maryland. (2026, April 14). Letter from the ACLU of Maryland to the Anne Arundel County Police Department on behalf of Kimberlee Williams. https://assets.aclu.org/live/uploads/2026/04/2026-04-13-kimberlee_williams_aa_letter_final_Redacted-1.pdf
- United States District Court, Eastern District of Michigan. (2025, August 5). Order Granting Defendant’s Motion for Summary Judgment, Woodruff v. Oliver, No. 5:23-cv-11886 (E.D. Mich.). https://www.govinfo.gov/content/pkg/USCOURTS-mied-5_23-cv-11886/pdf/USCOURTS-mied-5_23-cv-11886-1.pdf
- American Civil Liberties Union. (2024). Summary of Detroit Facial Recognition Settlement Agreement (Williams v. City of Detroit). https://assets.aclu.org/live/uploads/2024/06/williams_settlement_one-pager_june_24-1.pdf
- Maryland General Assembly. (n.d.). Maryland Code, Criminal Procedure, section 2-502. https://mgaleg.maryland.gov/mgawebsite/Laws/StatuteText?article=gcp&enactments=false§ion=2-502
- Maryland General Assembly. (n.d.). Maryland Code, Criminal Procedure, section 2-503. https://www.mgaleg.maryland.gov/mgawebsite/Laws/StatuteText?article=gcp&enactments=false§ion=2-503
- Grother, P., et al. (2019). Face Recognition Vendor Test (FRVT), Part 3: Demographic Effects (NISTIR 8280). https://nvlpubs.nist.gov/nistpubs/ir/2019/NIST.IR.8280.pdf