The problems with Clearview AI’s facial recognition system, particularly in the hands of police, are myriad and serious. That the technology exists as it does at all raises significant ethical concerns, and how it has been used to feed people into the criminal justice system raises significant due process ones as well. But an article in the New York Times the other day might seem to suggest that it perhaps also has a cuddly side, one that might actually help criminal defendants, instead of just hurting them.
But don’t be fooled – there is nothing benign about the facial recognition technology pushed by Clearview AI, and even this story ultimately provides no defense for it. It was not the hero here, because the problem it supposedly “solved” was not the problem that actually needed solving.
In the article, Kashmir Hill told the story of how Clearview AI’s facial recognition system apparently helped exonerate someone criminally charged with causing the single-car accident he had been in, which had killed the person he was with. The survivor defendant insisted he hadn’t been the one driving, but he was charged anyway. Proving his innocence was going to require finding the Good Samaritan witness who had pulled him from the burning car and could verify which seat he had been pulled from.
In November 2019, almost three years after the car accident, Mr. Conlyn was charged with vehicular homicide. Prosecutors said that Mr. Conlyn had been the one recklessly driving Mr. Hassut’s Mustang that night and that he was responsible for his friend’s death.
Mr. Conlyn vehemently denied this, but his version of events was hard to corroborate without the man who had pulled him from the passenger seat of the burning car. The police had talked to the good Samaritan that night, and recorded the conversation on their body cameras.
“The driver got ejected out of one of the windows. He’s in the bushes,” said the man, who had tattoos on his left arm and wore an orange tank top with “Event Security” emblazoned on it. “I just pulled out the passenger. He’s over there. His name is Andrew.”
The police did not ask for the man’s name or contact information. The good Samaritan and his girlfriend, who was with him that night, drove off in a black pickup truck.
Yet how could the defendant track down the witness? He had no idea himself who had helped him, and the police had not bothered to fully document what they observed at the scene. The only identifying data was the footage the police bodycams captured while they had been talking to the witness. Which led the defense to wonder, what if there was some sort of way to identify who was pictured in the footage. So defense counsel wrote to Clearview AI and asked for access to its facial recognition system to see if it could identify the person in the picture. And Clearview said yes, apparently smelling a PR opportunity by asking that that the defense “talk to the news media about it if the search worked.” And it turns out that it did work: the identity of the witness was readily found, the witness then located, and, with their testimony, the charges were consequently dropped.
The company would like the takeaway from this particular happy ending to be that Clearview AI’s facial recognition system might at least be a double-edge sword, offering some good and important benefits to the accused that might somehow counterbalance the tremendous threat it poses to all putative defendants (aka everyone), especially insofar how facial recognition technology in the hands of police tends to lead to people finding themselves in the crosshairs of the criminal justice system in the first place, often unaware and even by mistake.
Civil liberty advocates believe Clearview’s expansive database of photos violates privacy, because the images, though public on the web, were collected without people’s consent. The tool can unearth photos that people did not post themselves and may not even realize are online. Critics say it puts millions of law-abiding people in a perpetual lineup for law enforcement, which is particularly concerning given broader concerns about the accuracy of automated facial recognition.
But the story actually supports no such conclusion: Clearview AI was no way the solution to the problem presented here, because the problem here was not that the defense couldn’t find its witness. The problem was that there was obviously reasonable doubt as to his guilt, which prosecutors chose to ignore in deciding to charge him anyway. Which then meant that instead of the prosecution having the burden to prove his guilt, the defendant now had the burden to prove his innocence, which is how he found himself needing Clearview at all.
But he never should have had that need because that is not how things are supposed to work in our criminal justice system, where the accused are supposed to be presumed innocent and it is the prosecution’s job to prove beyond a reasonable doubt that they are not. True, to simply bring charges the prosecution may have needed to meet a lesser standard than reasonable doubt, but the problem here was that the prosecution gladly chose to pick this fight even though it knew that it should ultimately not be able to win the war.
Because think about how much evidence the prosecution already knew about that cast doubt on the suspicion the defendant had been driving. As the article lists, there was the defendant’s own denial, plus forensic evidence that was at best inconclusive to support that he had been the driver (and then there was also the fact that the passenger side door had been blocked by a tree, which would have meant that anyone in the car would need to leave by the driver’s side, even if they hadn’t been driving).
The body-camera footage did not seem to hold much weight with the prosecution.
“There was contradicting evidence,” said Samantha Syoen, the communications director for the state attorney’s office.
Witnesses who had arrived late to the scene saw Mr. Conlyn pulled out of the driver’s side of the car. Mr. Conlyn said, and police body camera footage appeared to confirm, that the passenger’s side door had been against a tree, which was why he’d had to be rescued from the other side. The police found his blood on the passenger’s side of the car but also on the driver’s side airbag.
An accident reconstruction expert hired by the prosecution said the injuries to the right side of Mr. Conlyn’s body could have come from the center console, not the passenger door. After Mr. Hassut’s father sued Mr. Conlyn in civil court in 2019, for the wrongful death of his son, Mr. Conlyn’s insurance agency settled the suit, unable to prove that Mr. Conlyn had not been driving the car.
But even if none of that contrary evidence had been compelling, THERE WAS ALSO A WITNESS! That the prosecution knew about! Because the police had spoken to him! AND THAT’S WHY THERE WAS EVEN A PICTURE TAKEN ON THE BODY CAM!
There should have been no need for the defense to ID the witness, because the mere fact that there was a witness, whose contemporaneous statement had cast doubt on the prosecution’s theory, should have been enough reasonable doubt to put an end to the prosecution. That the prosecution nevertheless continued, in spite of this contrary evidence, is the true problem that this story reveals. And it is the problem that Clearview AI in no way solves. The failure here is much more systemic, that overzealous prosecutors can go after defendants with such weak hands, and force defendants to have to do what may be impossible to prove their innocence (especially thanks to inexcusably poor documentation practices by investigating police).
That something like Clearview AI may make a defendant’s task slightly less impossible does not solve the ultimate problem, nor does it redeem technology as troubling as Clearview AI just because in this particular case it may have helped even the odds. The issue is that the odds were ever so uneven in the first place. And Clearview AI ultimately just helps make sure they will stay uneven by so cavalierly dismissing the significant privacy rights that should be protecting citizens from exactly this sort of overzealous policing.
The article quotes the NACDL’s Jumana Musa accurately noting that offering defense counsel access to Clearview isn’t going to solve the problems with it:
Jumana Musa is the director of the Fourth Amendment Center at the National Association of Criminal Defense Lawyers, where she works to keep defense lawyers informed about the newest surveillance tools used by law enforcement.
“[Giving defendants access to the system] is not going to wipe away the ethical concerns about the way in which they went about building this tool. [I]t’s not going to do is make us feel comfortable with the secrecy around how this tool works.”
As she continued, “You don’t address issues in a broken criminal legal system by layering technology on them.” Which is exactly the issue. Technology is not the solution to all our failings. Sometimes the thing we need to do is just fail less, but that requires recognizing what has actually gone wrong and what therefore needs addressing.
So, no, Clearview AI is not the solution here because the problem is not that not enough people don’t have access to facial recognition systems. Rather, the problem is that our system of justice railroads people even in the face of reasonable doubt. That it does is the problem we should be fixing, but it is not one that something so invasive like Clearview AI could ever do for us because it is not like its own existential problems can somehow cure or cancel out the current constitutional infirmities of our criminal justice system. Instead each will only make the other worse.