By: Kyle Kennedy

As technology has continued to evolve, law enforcement has begun to employ technological advances like facial recognition in their everyday pursuits. This has included issuing arrest warrants exclusively based on facial recognition matches.

Randall Reid was on his way to Thanksgiving dinner at his mother’s house in Georgia when he was arrested and jailed for stealing $10,000 of Chanel and Louis Vuitton handbags from a New Orleans suburb. The only problem was, Reid had never even been to Louisiana. This didn’t stop Reid from spending nearly a week in prison because a facial recognition tool had matched him to surveillance footage of the suspect in the case. Reid was arrested and held despite being about 40 pounds lighter than the suspect on the tape. 

This is not the first case of facial recognition technology leading to a wrongful arrest. Nijeer Parks spent 10 days in prison after an incorrect facial match and began to feel pressure to accept a plea deal because of his prior criminal record. Faced with the pressure of battling the court system and often the weight of prior charges, “[d]efense attorneys and legal experts say some people wrongly accused by facial recognition agree to plea deals”. Robert Williams has had multiple strokes since he was released from a 30 hour stint in jail after he was incorrectly matched to video of a suspect robbing a watch store. These are far from the only instances of inaccurate facial recognition leading to wrongful arrest. These arrests raise two important questions: can a facial recognition match be the basis of probable cause to substantiate a warrant? And if so, should it be? 

Facial recognition is broadly used to accomplish two tasks: verification and identification. Verification, also called one-to-one matching, is used to confirm a person’s identity like when logging into their smartphone or a banking app. Identification, or one-to-many matching, is when software compares an unknown face to a large database of known faces. Identification can be used on cooperative subjects who consent to having their faces scanned or uncooperative subjects whose faces are not scanned. The accuracy of these identification measures is much lower for uncooperative subjects whose facial images were captured in the real world. One algorithm had a 0.1% error rate when matching to high-quality mugshots, but this rate climbed to 9.3% when matching to images captured ‘in the wild’. The accuracy rates of facial recognition technology also vary by vendor; one top algorithm achieved an identification accuracy of 87% at a sporting venue while the accuracy rate of another vendor’s software was a dismal 40%. In addition to variation by vendor, even the most accurate algorithms tended to have “higher false positive rates in women, African-Americans, and particularly African-American women.” 

Probable cause differs for arrest warrants and search warrants. For an arrest warrant, probable cause is interpreted according to a flexible reasonableness standard based on a totality of the circumstances. For search warrants, there is probable cause when there is a fair probability there is evidence of a crime in the place to be searched. Based on the issues with accuracy and the frequency of inaccurate facial matches leading to wrongful arrests, a one-to-many match acquired through use of a facial-recognition algorithm cannot substantiate probable cause for a warrant without further evidence. This is especially true where the algorithm is used on images or video captured from uncooperative individuals ‘in the wild’, which would be the case when officers are trying to find unknown suspects with surveillance footage. A facial match with corroborating evidence linking a location to the individual matched might be able to substantiate a search warrant. On its face this might be less concerning due to the lower risk of wrongful arrests, but wrongfully or erroneously substantiated search warrants lead to many issues of their own.

The greatest danger posed by facial recognition inaccuracies are false positives because they lead to false accusations against otherwise innocent individuals. In response, some scholars have attempted to impose confidence thresholds on facial recognition algorithms to reduce the rate of false positives. One study found that a set of algorithms failed to return a match 4.7% of the time when no threshold was imposed, but that the rate jumped to 35% when a 99% threshold was imposed.  This means that 30% of the time the algorithm identified an individual, it was at a confidence level of below 99%. Many law enforcement departments that use facial recognition technology do not impose confidence thresholds on potential matches. While these confidence thresholds reduce the risk of false accusations, they still leave the door open for problematic uses of technology by law enforcement. Confidence thresholds are a band-aid solution which are difficult to enforce externally and allow the continued use of invasive surveillance technology.

Overall, one-to-many matches of uncooperative faces acquired through facial recognition algorithms are neither accurate nor reliable enough to be the sole basis of probable cause substantiating a warrant. These algorithms are susceptible to high error rates which vary unpredictably based on many factors, including the race and gender of the target individual. Additionally, the use of facial matching as the sole basis for arrest warrants has led to many wrongful arrests with long-term consequences for arrestees, like those mentioned above. The surrounding evidence of inaccuracy and history of wrongful arrests stemming from inaccurate facial matches make it clear that facial recognition matching shouldn’t be the legal basis for probable cause absent corroborating evidence. The remaining questions are to what degree a facial recognition match must be corroborated to substantiate a search warrant, whether police should be liable for harms stemming from inaccurate facial matches, and if so, what duty of care the police have in determining whether a given facial recognition match is accurate. Beyond further inquiry, the need for regulation surrounding the use of facial recognition and other surveillance technologies by law enforcement has never been more apparent.