top of page

How Facial Recognition Investigative Training Can Help Avoid Wrongful Arrests

By Clearview AI BLOG

This century has seen many significant advances in criminal forensic technology. For example, forensic DNA analysis developed significantly over the last twenty years. Prior to the 1990s, criminal investigators did not have DNA as an investigative tool to assist them in verifying or refuting the veracity of a person’s involvement in a crime. Police were dependent on limited information such as eyewitness testimonies, mugshot books, anonymous tips, and informants as their primary investigative resources to solve crimes. With DNA technologies now available, numerous cold cases were solved, innocent people exonerated, and victims received justice. One of many examples is the 2019 arrest of a suspect in the 1972 murder of Jody Loomis in 2019. The Snohomish County Sheriff said in a statement, “After more than 46 years of searching for her killer, we finally have some answers for Jody’s family.” [1]

Now, in 2022, cutting edge facial recognition technology (FRT), like Clearview AI’s technology, has the potential to be a game-changer for solving crimes and exonerating the innocent. Despite FRT’s high level of accuracy and proven effectiveness, its detractors continue to make aggressive attempts to ban its use by law enforcement. These bans are motivated by unjustified fears and factually incorrect narratives.

Simply put: banning FRT is the wrong approach.

Many of today’s alleged crimes have digital image evidence associated with them. Surveillance footage of a suspect, a video image of an act of human trafficking, or a child sexual abuse victim are commonplace in criminal investigations. A 2017 study showed that CCTV proved useful to police in 65% of criminal investigations. [2] Using evidentiary images in investigations, and high-performing facial recognition technology with already available databases (i.e., criminal records, motor vehicles, and public images located on the internet), law enforcement has access to incredible investigative leads.

High-performing FRT, designed around machine-learning and artificial intelligence, has extinguished the concern about demographic discrimination or bias in its performance. According to testimony by the Director of the National Institute of Standards and Technology (NIST)’s Information Technology Laboratory, high performing algorithmic FRT has undetectable demographic bias in the technology.

Unfortunately, some policy makers seek to limit or outright ban FRT as an investigative tool, citing common myths about the technology and outdated incomplete information. A lack of understanding about FRT fosters an environment where policymakers craft laws that ultimately harm victims rather than protect them by eliminating a valuable crime-fighting tool.

The limitations on eyewitness accounts are well-known and prevalent. Prior to using DNA as evidence, many wrongful convictions were linked to misidentifications by an eyewitness. [3] Not to say that all human eyewitness accounts are inaccurate, but the human memory is flawed in its identification abilities. If you were to place five images of a child you know on a table with a thousand images of other children, you have a high chance that you will accurately identify the child you know. However, humans unfortunately are inconsistent. Since it is the brain that “sees” rather than the eyes, humans are influenced by many other factors such as familiarity, emotion, stress, and interaction context. Today, eyewitness testimony still remains valuable, but it is subject to great scrutiny. Thus, every effort should be made to confirm the veracity of an eyewitness identification which high-performing FRT can do exactly that.

High-performing FRT provides the consistency that is lacking in eyewitness identifications. Factors that may negatively influence eyewitness consistency are minimized from FRT processing. High-performing FRT can apply that consistency across billions of images in seconds – Clearview AI’s algorithm attained 99.85% overall accuracy in testing by NIST. [4] While the use of FRT in lead generation is typically at the front end of an investigation, in the appropriate circumstances it can also be used to aid in the confirmation process of other witness identifications.

When DNA testing came to the forefront twenty years ago, it was controversial. People were unsure of its reliability and afraid of potential government abuse. Despite such skepticism, DNA testing was not banned. The vast benefits of DNA testing outweighed the hypothetical risks. To ensure proper use of DNA testing, policies were formulated and working groups within the criminal justice and scientific communities were formed. As confidence in DNA testing grew, it became the main tool to rectify former wrongful convictions.

Like DNA testing, FRT can act as today’s groundbreaking identification tool. Prohibiting or severely restricting the use of FRT by law enforcement to prevent wrongful arrests is self-defeating, because these bans actually increase the likelihood of wrongful arrests and convictions. While the conclusivity of FRT results is not the same as DNA, the use of FRT for identification purposes in investigations can benefit from similar prudent measures taken to introduce DNA testing as a criminal investigative technology. Facial recognition technology has an additional favorable quality compared to DNA testing–any layman can easily understand and form his or her own opinion about the accuracy of a facial recognition search. DNA matches must be determined by an expert using laboratory methodology not easily understood by non-experts–like lawyers, judges, investigators, and indeed the citizens who sit on juries. This increases the likelihood that mistakes in DNA matching will go unnoticed and unquestioned, or that juries can be misled by expert testimony that overstates its case. In contrast, anyone can look at two images of human faces and form an accurate impression about whether or not they depict the same person–meaning that investigating officers, prosecuting and defense attorneys, and jurors all can readily detect any faulty use of facial recognition search technology, substantially reducing the likelihood of wrongful arrests and convictions as a result of faulty facial recognition searches, compared to many other investigative technologies and methods. Faulty ideology does not have to be chosen over actual human victims. Crime victims, public safety, and protection of our civil liberties will all benefit from the use of FRT.

There are thousands of U.S. cases where proper use of FRT over the last decade has aided law enforcement in identifying and rescuing a victim, identifying a suspect, and helping identify someone in a medical event or mental health crisis. FRT is reliable technology that can, with appropriate use policies and procedures, save lives and help prevent wrongful arrests. In the U.S. today there are three well-publicized cases that FRT critics have attributed to a failure of the technology, which is simply erroneous. With better police procedures and adoption of advanced FRT, as opposed to less, you will see how such measures would have assisted in each legal investigation below.


Detroit PD, 2019

Overview (based on open-source information): In October 2018, an african american male committed a theft at a Shinola store in Detroit, Michigan. A third party loss-prevention service employee (who did not physically witness the crime) pulled the store security video, captured still images of the suspect from the tape, and sent them to Detroit Police Department (DPD). DPD sent the grainy suspect images to the Michigan State Police (MSP) lab for analysis. An FRT query was conducted by MSP against the state’s criminal record and driver license databases and a candidate list was returned to DPD. The FRT query report generated 200 potential matches to DPD investigators. Williams was one of them based on his driver license database image and the DPD investigators focused on him. DPD put Williams’ in a “6 pack” photospread (which many professional law enforcement agencies abandoned years ago due to issues with witness presentation). DPD then showed the photospread to the loss-prevention service employee (again, not an actual witness to the crime) and received an alleged positive identification on Williams.

MSP was limited to searching only government databases. Would it have been helpful if DPD investigators had access to query a database of billions of public internet images to offer additional leads? It is possible they would have received candidates with higher confidence rankings. It is possible they would have received images of candidates posting themselves wearing the same or similar clothing depicted in the crime scene image. It is possible they would have received a candidate image of the true suspect. It is possible they would have been able to eliminate Williams from suspicion early in the investigation. The fact that FRT was used in this case is not the problem, but rather, the limited dataset and poor police procedures were at fault.


Detroit PD, 2019

Overview (based on open source information): During a disturbance at a local high school, a teacher was attempting to capture (from his vehicle) the incident on his smartphone. A black male walked up to the window of the vehicle and took the phone from the teacher. The teacher was able to access the video and gave Detroit police a copy. DPD queried the suspect image in their FRT system, querying government databases and Oliver was provided as a candidate. The FRT system does not compare other physical features – that is the job of investigators. Obviously, the investigators failed. Absent the poor police work here, there would still have been other, potentially more viable leads to pursue if DPD expanded their search beyond the limited government databases. A query of a larger image dataset such as public internet images contained in Clearview AI’s technology platform would have likely returned a candidate image that had more a likeness to the suspect, and more importantly, not returned any images of Oliver). If you run the image of Oliver in the Clearview AI technology platform, the search results do not return any images with a likeness of the suspect. Better data, along with better police procedures, could have prevented Oliver from being included in the investigation.


Woodbridge, NJ, 2019

Overview (based on open source information): Woodbridge Police Department (WPD) responded to a shoplifting incident. The suspect provided a fake Tennessee driver’s license (DL) with the suspect’s picture on it. When WPD attempted to arrest him, he fled, damaging a police car and almost hitting an officer. WPD showed the fake driver’s license to witnesses who confirmed the image was the suspect. WPD requested assistance from a regional agency who conducted an FRT query against government databases. The query resulted in Nijeer Parks being included as a candidate mostly likely from an image in the criminal record database. Based solely on the witness identification of the driver’s license and the FRT result, WPD issued an arrest warrant for Parks. After learning that he was wanted, Parks turned himself in to clear his name. Despite his claims of innocence and other exclusionary investigative clues, the WPD still arrested him. Parks was able to produce solid evidence that he was thirty miles away from the crime scene and was released from custody. As with the other two cases, the police work involved was poor at best. Had the police been able to run additional FRT searches against a broader dataset, it is likely those leads would have taken investigators in a different direction preventing Parks’ wrongful arrest. When tested with the Clearview AI technology platform, the query of the suspect image provided multiple images of alternative candidates with higher degrees of match precision.


The three cases above highlight several points:

  1. Placing the blame on FRT as the reason for the wrongful arrests was incorrect.

  2. The use of FRT in an investigation is a starting point providing a lead to follow in the investigation process. FRT leads require police to follow basic investigative steps to verify or refute the results and are no different than a phone tip, a crime-stoppers tip, an eyewitness statement, or an informant tip. The tip/lead is not the problem in these cases. The investigative process broke down in all three cases and exclusionary clues were abundant and ignored.

  3. There is not an extensive record of law enforcement use of FRT leading to wrongful arrests. Two of the three cases above involved the same law enforcement agency in the same year. This indicates that the faulty police investigative procedures used were the cause for the wrongful arrests.

  4. Increasing access and the breadth of a dataset, not limiting it, would have likely led investigators in a more effective direction, and potentially prevent the wrongful arrests of Oliver and Parks. It can also help identify the actual perpetrators in each case. Simply having more data is not always helpful, but having more useful data is. Utilizing public internet images in conjunction with images from government databases optimizes the lead generation process and provides investigators more viable lead direction to mitigate wrongful arrests.

  5. None of these 3 publicly-known wrongful arrests are due to the use of Clearview AI’s technology, but it used older unreliable algorithms and bad investigative techniques. To this date, we do not know of an instance where Clearview AI's technology has resulted in a wrongful arrest.

  6. Using Clearview AI’s bias-free algorithm and worldwide dataset with proper facial recognition training would have prevented these wrongful arrests from happening.


[1] Source: C-SPAN NIST Congressional Testimony Feb2020, Starts at the 33:04 mark
[3] Source: Innocence Project
[4] Source: NIST 1:N FRVT Verification, November 2021


bottom of page