Police Sued - Again - For Questionable Facial Recognition Arrest

This photo from a store security camera was used to attempt a facial recognition match. It found one. Which was, not surprisingly, the wrong person. Now he's suing the police department, the detectives that used it and the software maker.  

As the 1994 movie title reads, "This could happen to you." JL

Tim Cushing reports in Tech Dirt:

The detective secured this warrant by omitting key details, like the fact the suspect was picked out of a facial recognition database using a low quality image. It also did not note that the person who picked Williams out of Bussa's loaded lineup wasn't at the store the day the shoplifting happened. And it didn't mention bypassing of Detroit PD policies that were put in place to prevent exactly this sort of false identification. The same software and the same two detectives were involved in another false arrest -- one that occurred five months before.

On January 9, 2020, facial recognition tech finally got around to doing exactly the thing critics had been warning was inevitable: it got the wrong person arrested.

Robert Williams was arrested by Detroit police officers in the driveway of his home. He was accused of shoplifting watches from a store on October 2, 2018. The store (Shinola) had given Detroit investigators a copy of its surveillance tape, which apparently was of little interest to the Detroit PD until it had some facial recognition software to run it through.

This was the dark, grainy image the Detroit PD felt was capable of returning a quality match:

That picture is included in Williams' lawsuit [PDF] against the Detroit Police Department. Even in the best case scenario, this picture should not have been uploaded to run a search against. It's low quality, poorly-lit, and barely shows any distinguishing facial features.

What makes it worse is that all facial recognition AI -- across the board -- performs more poorly when attempting to identify minorities. That's the conclusion reached by an NIST study of 189 different algorithms. It's not just some software. It's all of it.

The Detroit PD chose to run with that photo. Then it decided the search results it had were close enough to probable cause to effect an arrest, even though the software used stated clearly search results should not be used this way. The search was performed by the Michigan State Police from the grainy image submitted by the Detroit PD. A report was returned but investigators were cautioned against trying to turn this into probable cause:

The following statement appeared prominently on the Investigative Lead Report, in the form shown: “THIS DOCUMENT IS NOT A POSITIVE IDENTIFICATION. IT IS AN INVESTIGATIVE LEAD ONLY AND IS NOT PROBABLE CAUSE TO ARREST. FURTHER INVESTIGATION IS NEEDED TO DEVELOP PROBABLE CAUSE TO ARREST.” The phrase “INVESTIGATIVE LEAD ONLY” was highlighted in red ink.

The report was also light on any other details that might have indicated Robert Williams was actually the shoplifter in question.

The Investigative Lead Report contains neither the “score” generated by the facial recognition system representing the level of confidence that Mr. Williams’s photo matched the probe image, nor the other possible matches that, upon information and belief, should have been returned by the system.

The Detroit Police Department did not attempt to ascertain the “score” generated by the facial recognition search nor request the other possible matches to the probe photo.

Two months after the PD obtained these search results, the investigation was turned over to another detective, Donald Bussa. At the point he assumed control of the investigation, Bussa was supposed to be operating under the PD's new facial recognition policy that acknowledged the limitation of the tech and stated search results would need to be peer reviewed to ascertain their accuracy. This didn't happen.

Defendant Bussa, however, ignored the new policy. Even though the facial recognition search “identifying” Mr. Williams as the shoplifter was generated by a woefully substandard probe image and had never been peer reviewed by DPD officers, as required by the new policy, Defendant Bussa decided to rely on the lead anyway.

Bussa assembled a "six pack" of suspect photos that contained an image taken from Williams' expired drivers license. (Before this investigation took place, Williams had secured a new license and an updated photo.) He tried to speak to the staff at Shinola but management refused to cooperate, stating that it was not interested in having its employees appear in court. Unable to speak to the sole eyewitness who had actually conversed with the shoplifting suspect, Detective Bussa decided to bypass Shinola completely.

Defendant Bussa then arranged to conduct a six-pack photo identification with Katherine Johnston. Ms. Johnston, then employed by Mackinac Partners, was contracted by Shinola for loss prevention services.

Defendant Bussa had no legitimate basis whatsoever for asking Ms. Johnston to participate in an identification procedure. Ms. Johnston was not an eyewitness. Ms. Johnston was not in the Shinola store at the time of the incident and has never seen Mr. Williams or the alleged shoplifter in person. Indeed, Ms. Johnston’s sole relation to the incident was that she had watched the same low-quality surveillance video that Detective Bussa possessed.

Bussa sent Detective Steve Posey out with the loaded "six pack" to pretty much guarantee Williams was selected as the prime suspect.

The photo array was not a blind procedure—Posey knew that Mr. Williams was the suspect. Indeed, Posey’s sheet was nearly identical to that given to Ms. Johnston, except that Mr. Williams’s name was printed in red while all other names were printed in black.

Ms. Johnston identified Mr. Williams’s expired license photo as matching the person she had seen in the grainy surveillance footage, and answered the question “Where do you recognize him from?” with “10/2/18 shoplifting at Shinola’s Canfield store.”

With that, Bussa went out to get an arrest warrant and the rest is facial recognition history. He secured this warrant by omitting some key details, like the fact the suspect was picked out of a facial recognition database using a low quality image. It also did not note that the person who picked Williams out of Bussa's loaded lineup wasn't even at the store the day the shoplifting happened. And it didn't mention Bussa's bypassing of Detroit PD policies that were put in place to prevent exactly this sort of false identification.

The lawsuit also points out that the same software and the same two detectives were involved in another false arrest -- one that occurred five months before the PD arrested Williams. Detective Bussa and Detective Posey used unvetted search results to arrest Michael Oliver for an assault he didn't commit. Even if the facial recognition software had done its job accurately (which it didn't), the tech would not have noticed something far more obvious: the suspect's arms (as captured in the phone recording) were unmarked. Michael Oliver's arms are covered with numerous tattoos.

Williams alleges a long list of violations he's hoping to hold Detective Bussa (and his supervisor) accountable for. It's going to be pretty difficult for the detective to argue he operated in good faith in the Williams arrest. After all, he'd already followed the same broken process to falsely arrest someone else months earlier. Then he ignored the PD's policy on facial recognition tech. He also ignored the big, bold warning printed across the search results he obtained from the State Police. And none of this information -- which would have undercut his probable cause assertions -- made its way into his warrant request.

Any reasonable officer would know a lot of what Detective Bussa did was wrong. But Bussa would know this more than even the most reasonable of officers because it wasn't the first time he'd screwed up.