Resource guide

How AI Predictive Policing Creates Synthetic Suspicion

How a data flag becomes a police stop, with cases from the American Civil Liberties Union (ACLU) and 404 Media reporting.

Last updated November 20, 2026 3343-word guide By Jonathan West

Synthetic suspicion is the term Tech Policy Press uses when artificial intelligence (AI) creates the justification for police actions, a pattern that 404 Media reports in Border Patrol predictive policing units.

Reported cases include stops tied to license plate driving data, an address tool used by U.S. Immigration and Customs Enforcement (ICE), and facial recognition arrests.

How AI Predictive Policing Flags Turn Into Stops

The reports describe four steps, from the data a system reads to the action an officer takes.

  1. Data source. Driving data for some flags comes from automated license plate readers. The American Civil Liberties Union (ACLU) and Tech Policy Press both describe that driving data as the input for travel-pattern flags. 404 Media names financial activity as another input for Border Patrol predictive units. Surveillance images are an input in the ACLU's account of a facial recognition arrest.
  2. AI flag or score. According to Tech Policy Press, the Associated Press reported on a Border Patrol program in which AI designates vehicles as "suspicious" based on driving patterns. Enhanced Leads Identification & Targeting for Enforcement (ELITE) is a tool that Palantir developed for ICE. It provides an "Address Confidence" score for a person's current address. In Tech Policy Press's account, a facial recognition match can come with an AI confidence score.
  3. Officer acts. The reports describe flags that lead officers to stop drivers. 404 Media reports that local police pull over people "not suspected of any specific crime." In the case of Kimberlee Williams, the ACLU says detectives relied on a facial recognition lead and their own visual comparison of photos.
  4. Stated reason. The reason an officer gives can differ from the reason the reports describe. Knowles told the ACLU that the officer said she had been in the left lane too long. The ACLU says the real reason was a travel-pattern decision made by a surveillance system.

Jake Laperruque is a fellow at Tech Policy Press and Deputy Director of the Security and Surveillance Project at the Center For Democracy & Technology (CDT). On November 18, 2026, he published a perspective titled "'Synthetic Suspicion' and the Perils of AI-Prompted Police Actions." It is the second piece in a Tech Policy Press series on the perils of AI in policing. He warns that "a major risk is that the technology will replace evidence, investigative techniques, and human review for police activities." He writes that drivers learned, through inquiries and court proceedings, that an AI flag was the true reason for some stops. He describes those stops as parallel construction, which is his characterization.

Reported Cases of Stops Tied to Driving Patterns

Tech Policy Press says Jay Stanley's ACLU piece lists stops based on AI scans of driving patterns in California, Michigan, Texas and Wisconsin. The ACLU text does not label each case as AI-flagged. Each case below is reported or alleged, and the named source is given for each.

Grace Knowles, from northern Virginia, was on a road trip through Oregon, Texas and Louisiana in September 2025. A police officer pulled her over on her way to Florida. Jay Stanley of the ACLU Speech, Privacy, and Technology Project published the account on October 26, 2026. Knowles told the ACLU that the officer said she had been in the left lane too long. She told the ACLU that the officer claimed to be part of the Federal Bureau of Investigation (FBI) and said: "I just had to be sure because I'm not a regular cop, I'm part of the FBI, and when I scanned your plate I could see you had suspicious travel patterns." Knowles told the ACLU this is how she thinks the officer worded it. Knowles said the officer showed her software with a heat map of the United States and a photo of her license plate.

The ACLU writes that the real reason for the stop was that a surveillance system had access to automated license plate reader data on where Knowles had driven. The system, according to the ACLU, decided her travel pattern was "suspicious." The ACLU also says Knowles may have been stopped under the Border Patrol program.

The Associated Press reported in November 2025 that Border Patrol monitors drivers nationwide for "suspicious" travel patterns, according to the ACLU. The ACLU describes a driver for a small furniture-moving company who received a pretext stop. A local police officer made the stop at the direction of Border Patrol about 100 miles from the Mexican border, after the driver had come from South Carolina. According to the ACLU, the search of his truck found nothing illegal. Police arrested him on suspicion of money laundering because he carried thousands of dollars of his own cash, and police seized the cash.

The ACLU defines a pretextual stop as one in which someone is pulled over for a claimed, often invented minor infraction, as cover for the real reason. It says police call these "whisper stops" or "intel stops." The ACLU says it has fought to ban pretext stops for years.

According to the ACLU's account of the Schott case, a Texas jury recently sided with Alek Schott, who sued his local police department after a pretext stop. The ACLU says the officer claimed Schott had drifted over his lane line. According to the ACLU, dashcam footage showed that claim was bogus, and the officer later admitted it. The ACLU says a federal fusion center alerted local police that Schott, who sells oilfield supplies, was suspicious. The reason given was that he drove from Houston to Carrizo Springs and back the next day. The Institute for Justice, which represented Schott, noted that the same county subjected thousands of other drivers to the same routine, the ACLU says. The ACLU's account does not say whether software flagged him.

In March 2026, according to the ACLU, The Markup reported that U.S. Customs and Border Protection (CBP) officers at the border interrogated a grandmother in California about the time she spends at local casinos. The ACLU says the questioning was "apparently based on her driving patterns." She was told, "It just seems suspicious."

In April 2025, according to the ACLU, a man was targeted for a stop because he drove back and forth between Wisconsin and Michigan. Local police noted that Michigan was "a known source State for Marijuana as it is legal there."

The ACLU first noticed Flock Safety touting AI in July 2025 that could alert law enforcement to drivers whose travel patterns were "suspicious." According to the ACLU, Flock Safety said it could track associations by analyzing which vehicles were frequently seen with other vehicles. The ACLU describes this as a leap to "generating suspicion where none previously existed." For background on the camera system, read what Flock cameras record.

The ACLU writes that a year later Wired obtained files describing Flock's next-generation AI system. Some police already used the system. It has capability for open-ended AI queries of license plate data and other data, including arrest records, case files and commercial data broker databases.

Joseph Cox wrote a September 8, 2026 article for 404 Media. Tech Policy Press says Cox detailed how the Department of Homeland Security (DHS) runs "Predictive Intelligence Targeting Teams" that flag individuals for stops and searches based on their "financial activity patterns." The visible lead says Border Patrol runs predictive policing units that analyze financial activity and other data. It says the units feed that intelligence to local police. The rest of the 404 Media article is behind a paywall, so only its opening paragraph is used here.

ELITE Address Confidence Scores at ICE

404 Media reported on January 15, 2026, that Palantir is working on a tool for ICE. The visible lead says the tool populates a map with potential deportation targets and brings up a dossier on each person. It also provides a "confidence score" on the person's current address. The lead says ICE uses the tool to find locations where many people it might detain could be based.

Tech Policy Press says ELITE uses "advanced analytics" and commercial and government data sources. It creates dossiers of potential deportation targets and builds maps of their locations with an accompanying "Address Confidence" score. The 404 Media lead says the tool receives people's addresses from the Department of Health and Human Services (HHS), among a range of other sources, according to internal ICE material.

Tech Policy Press also relays testimony from an Oregon court document from 2025, published by 404 Media. In it, a deportation officer with ICE's Fugitive Operations Unit described using the system to select raid locations. He is quoted: "It tells you how many people are living in this area and what's the likelihood of them actually being there…" The officer testified that a raid under court review was based on an "ELITE lead." He said he did not know what data or evidence the system's recommendations were generated from. The testimony has not been independently verified.

The cited reports do not describe how the address confidence score is calculated. They do not state an error rate for it.

Wrongful Arrests from Facial Recognition Matches

The ACLU published an April 14, 2026, article by Lauren Yu and Nathan Freed Wessler of the ACLU Speech, Privacy, and Technology Project. It describes Kimberlee Williams, an ACLU client, as "the fourteenth person known to be wrongfully arrested due to the technology's failures." That count and characterization belong to the ACLU. For a separate look at ICE facial recognition systems, see that explainer.

According to the ACLU, Williams, a grandmother living in Oklahoma, was detained on June 23, 2021, at a base entry checkpoint in Lawton, Oklahoma. Security found outstanding Maryland arrest warrants. The warrants sought her arrest over fraudulent over-the-counter cash withdrawals at bank branches in Maryland in December 2019 and January 2020. The ACLU says she was in Oklahoma at the time.

As the ACLU tells it, a bank financial crimes investigator sent a surveillance image of the suspect to Crimedex, a national listserv of police and private investigators. Someone on the email list ran the image through facial recognition technology. That person sent back Williams's name and photo as a purported match. The ACLU says the bank investigator gave police no information on who ran the search or how.

In the ACLU's account, the detectives relied on the facial recognition lead and their own visual comparison of the photos. They did not try to establish whether Williams could have been in Maryland, and they did not investigate other leads. According to the ACLU, Williams spent six months in jail in total before the last charges were dropped.

The ACLU and the ACLU of Maryland sent letters to three Maryland police departments on Williams's behalf. The ACLU alleges that police relied on a false facial recognition result. It also alleges that police concealed that reliance from the court when applying for an arrest warrant.

Beyond Williams, the ACLU lists at least thirteen other people. It says each is publicly known to have been wrongfully arrested by police in the United States because of erroneous facial recognition results. Those arrests are dated from February 2019 to August 2025. The ACLU says wrongful arrests after police moved from facial recognition results straight to photo lineups affected at least seven people.

On accuracy, the ACLU writes that facial recognition "often produces false matches." It says multiple studies have shown higher false match rates for people of color, women, younger people and the elderly. It also says most known wrongful arrests have been of Black people.

Jake Laperruque writes that facial recognition "has served as the primary or sole basis for many arrests." He writes that at least three separate cases were reported in the year before November 18, 2026. In each, someone spent months in jail because of an erroneous facial recognition match.

House Homeland Security Committee Ranking Member Bennie Thompson (Democrat, Mississippi) made a statement in 2025 that Tech Policy Press relays. He said ICE officials have described an apparent Mobile Fortify biometric match as a "definitive" determination of a person's status. According to that statement, an ICE officer may ignore evidence of American citizenship, including a birth certificate, if the app says the person is an alien.

Laperruque also cites a Chicago Police Department training PowerPoint obtained through a public disclosure request. It advised officers: "Do not let the software become exculpatory evidence."

Reported Systems Compared

The table records what each named source reports. A cell reads "Not stated in the ACLU account" or "Not stated in the source" when the cited report does not give that detail.

System or caseData it usedWhat the AI output wasReported effectSource and date
Border Patrol travel-pattern programAutomated license plate reader driving dataVehicles designated "suspicious" based on driving patterns and flagged to be pulled overPretext stop. Driver arrested on suspicion of money laundering, per the ACLUAssociated Press, November 2025, as relayed by the ACLU (October 26, 2026) and Tech Policy Press (November 18, 2026)
Knowles stopAutomated license plate reader data on where her vehicle had driven, per the ACLUSurveillance system decided her travel pattern was "suspicious," per the ACLU. Officer said her plate showed suspicious travel patterns, per KnowlesTraffic stop. Officer said she had been in the left lane too long, per KnowlesACLU, October 26, 2026
Schott stopNot stated in the ACLU accountNot stated in the ACLU accountStop. A federal fusion center alert called him suspicious, per the ACLU. A Texas jury sided with himACLU, October 26, 2026
Border Patrol predictive unitsFinancial activity and other data, per the 404 Media leadIntelligence fed to local police, per the 404 Media leadLocal police pull over people "not suspected of any specific crime," per the 404 Media lead404 Media, September 8, 2026
ELITECommercial and government data sources, per Tech Policy Press. Addresses from HHS among other sources, per the 404 Media lead"Address Confidence" score on a person's current addressA raid under court review was based on an "ELITE lead," per Tech Policy Press citing an Oregon court document404 Media, January 15, 2026. Tech Policy Press, November 18, 2026
Kimberlee Williams arrestSurveillance image sent to the Crimedex email list, per the ACLUFacial recognition match returning a name and photo as a purported match, per the ACLU. Software not statedArrest. Six months in jail. Last charges dropped, per the ACLUACLU, April 14, 2026

Why Reliability Is Hard to Judge

Jake Laperruque argues it is "incredibly challenging to gauge how much weight to give AI-generated recommendations or evidence." He contrasts an eyewitness who identifies a perpetrator from 10 feet away in daylight with one who recalls a face from a distance on a rainy night.

He asks how police or a judge should assess a face match with an AI confidence score of 97, compared with scores of 92.5 or 88.25. He also asks how to weigh an ELITE confidence score or an AI evaluation of how suspicious someone's driving patterns were.

Laperruque defines automation bias as "the tendency of people to place trust in automated and computer-generated recommendations, even when those systems are flawed or contradicted by independent evidence." He argues that AI-generated suspicion will tend to receive excess deference.

He argues that vendors of AI policing technology have a motive to configure systems to generate "suspicion" more aggressively and more often. His reason is that police want flexibility to conduct stops. He also writes that departments could configure AI to provide more responsive results. This is his argument.

Laperruque writes that in 2019 the Washington County Sheriff's Office in Oregon complained that its facial recognition system often returned no results. According to The Washington Post, as Tech Policy Press describes it, the department dropped its required search confidence threshold and returned five results every time.

He also writes that the FBI testified before a House Oversight and Reform Committee in 2019. According to that testimony, the FBI configured its facial recognition systems to return 2 to 50 potential matches with every scan.

Laperruque writes that many department policies and state laws prohibit relying on a facial recognition match as the sole basis for an arrest. He writes that matches are repeatedly treated as all but definitive. In one 2018 robbery case, he writes, a New York Police Department detective texted an eyewitness a picture of the match and asked "Is this the guy?"

Safeguards Proposed by Named Authors

Laperruque proposes the following measures.

  1. Independent vetting and testing of systems before acquisition.
  2. Development of best standards and practices for system settings.
  3. Isolation of AI systems from other parts of investigative work, similar to forensic labs.
  4. Guidelines that assign analog levels of evidentiary weight to AI outputs. He cites the Miami police department, which recommends treating a match as akin to an anonymous tip.
  5. Rigorous training for personnel who use AI.
  6. Strict limits on when AI tools can be deployed.

Laperruque writes that these measures "are far from sufficient" and might prove insufficient. He concludes: "Until we've developed effective measures and reviewed their efficacy, the best course of action is to simply press pause. Synthetic suspicion should have no role in policing."

The ACLU argues that more people will face police encounters as these algorithms "inevitably sweep in far more innocent people than actual wrongdoers." It calls for an end to license plate surveillance, location tracking and mass surveillance.

The ACLU says more than 20 cities and other jurisdictions have banned police from using facial recognition technology. The ACLU says that under a settlement in Robert Williams's wrongful arrest case, Detroit police no longer permit officers to request arrest warrants based only on a photo lineup combined with a facial recognition-based lead. The ACLU says Indiana has enacted a similar protection into state law.

None of the cited reports says whether any agency or vendor has adopted these measures.

What the Reported Cases Show

Based on the cases the ACLU, Tech Policy Press and 404 Media report, an AI flag or match can start a police stop. In the reported cases, the reason given to the driver can differ from the reason the reports describe. Tech Policy Press argues that AI scores are hard to weigh, and that people tend to defer to automated recommendations. Whether a particular stop or arrest was lawful or accurate is a question for courts, and the cited reports do not settle it.

The reported cases involve drivers whose travel was logged and people matched by facial recognition. The picture would change if courts, agencies or vendors published error rates or rules that limit AI flags as grounds for police action. The cited reports contain no agency or vendor response, and no published error rate for travel-pattern flags.

Where to Check Local Surveillance Tools

In a November 18, 2026 post by Beryl Lipton, the Electronic Frontier Foundation (EFF) wrote: "Do you know what surveillance tools are being used by your local police departments? Check out the Atlas of Surveillance to learn more about the tools deployed in your city, and sign up to volunteer if you want to push for more transparency across the country."

Look up your city in the Atlas of Surveillance named in the EFF post.

Sources

Frequently asked questions

▸ What is predictive policing?
404 Media uses the term for Border Patrol units that analyze financial activity and other data. The other cited reports describe AI flags without using that term, and none gives a general definition.
▸ How is AI used in predictive policing?
The reported uses are driving-pattern flags built on license plate reader data, ELITE address confidence scores used by ICE, and facial recognition matches. The cited reports describe these uses without measuring their results.
▸ Is predictive policing effective?
The reports cited here do not measure whether these programs reduce crime or improve stops, so no effectiveness figure is available from them. The named authors argue about reliability and safeguards, as covered above.
▸ What is synthetic suspicion?
Tech Policy Press, in a perspective by Jake Laperruque published November 18, 2026, uses the term for a situation in which AI creates the justification for police actions.
▸ Can police stop a driver because of an AI flag?
The reports describe stops that followed AI flags, according to the ACLU and other named sources. Whether such stops are lawful is a question for courts, and the cited reports do not settle it.