SARAH ST.VINCENT: Parallel construction is one of the most frightening civil liberties issues that you may never have heard of. For the purposes of this discussion, I’m defining parallel construction as something that a government body does deliberately to create an alternative explanation for how it found a piece of evidence. The government did Thing A, but it doesn’t want the court, or the defense attorney, for example, to know that it did Thing A, so it goes back and does Thing B.
Sarah St.Vincent
So how might this work? The Intercept recently had an article about immigration enforcement in Texas, and how state troopers are starting to find people through traffic stops. And I thought, my gosh, when somebody gets pulled over for having windows that are tinted too dark, and then lo and behold! They have an immigration issue! That is kind of too happy for law enforcement to be a coincidence.
So, hypothetically, say the government has some very large database of phone records—as we know they have had, thanks to the Snowden revelations—and suppose they use those phone records to look at someone who they know is undocumented. They could then ask: “Who was that person in touch with? Who else might be undocumented?” Because they know that this might raise constitutional issues, they might ask a local officer to pull over the car that belongs to a person that interests them. And they don’t always tell the officer what they’re looking for. So, the officer follows the car, waits for it to drift over a line, or to not signal 20 feet before a turn. They pull the person over and then say, “Hey, now that I checked your license, can I ask you a couple of questions? Where do you live? How long have you lived there? What’s your nationality?” And then as far as you would know, as the defendant, that’s how your case started—with the traffic stop. You wouldn’t find out about this huge underlying program that might potentially render the rest of the case “fruit of the poisonous tree”—in other words, something that shouldn’t be used in court because it was illegally obtained.
Similarly, suppose the government secretly applies a new facial recognition technology to a video from a store security camera to identify a robbery suspect. It identifies a suspect, and then it sends a human informant to their house, or to a street corner, to strike up a conversation and start asking questions. And then suppose according to the records that get filed in court, that case started with a human informant. You don’t find out that the government may have actually used this facial recognition technology that may be biased, that may be inaccurate, or that may draw on sets of data that are otherwise illegal or problematic. You don’t know, because the government has engaged in parallel construction. It recreated the evidentiary trail.
So, what’s the problem? First, this shields government conduct from constitutional scrutiny—you’ll never have a chance to have a court rule on whether that phone records program was constitutional, or whether that facial recognition technology is constitutional.
Second, in U.S. court cases, if evidence was obtained illegally, courts will normally not allow the prosecution to introduce it into evidence. This is called the exclusionary rule. And the point of it is to deter law enforcement misconduct. If you remove this incentive for officials to obey the law, it could lead to coercion, perjury, and other problems. Think about the traffic stop scenario, where you have an officer who has been told to find a reason to pull this car over and then find a way to make an arrest or search the car. If you’re the officer, you may feel pressured to do those things even where the Constitution and court rulings would suggest that you’re crossing a line. You might illegally prolong a stop. You might illegally coerce someone to consent to the search. Officers might feel pressured to falsify reports or lie about what they did.
It also creates Brady problems. Brady v. Maryland is a Supreme Court case that says the government must turn over any evidence it has that may be favorable to the defense. But if you as the defense don’t know about this big collection of data, you may not be able to get at your Brady information, because again, you don’t know that information even exists.
What does the government seek to conceal through parallel construction? Well, potentially anything—and I mean, really, anything. It could be a wiretap with a warrant. It could be a human source, or it could be a gigantic NSA or DEA program.
There was a 2015 case in the Southern District of Ohio that I don’t think got enough attention—a case of someone who supposedly had attempted an attack motivated by support for ISIS. And the FBI said, well, we found their Twitter feed. Of all the Twitter feeds in the world, you found that one? House Speaker John Boehner came out and trumpeted that FISA was involved, and said this is why it was so important to renew FISA Section 702, a major intelligence surveillance law. But that was not anywhere in the indictment, or any of the court records we’ve seen, and the prosecution claimed that this came from publicly available information.
If parallel construction is legal, does the Bill of Rights still have any meaning? If the government can do things that are constitutionally questionable, or that it should know are unlawful, and then simply never tell anyone, do those rights still have any meaning? I think this is very, very frightening.
CLARE GARVIE: I want you to imagine for a second that you’re home at night. It’s 8:00 p.m., and the police knock on your door and say there’s been a robbery in the neighborhood. They say, “Don’t worry, we think we have the person who did it in custody. However, we want you to come down to the station and stand in a lineup.”
Clare Garvie
I think a lot of us would say no. You might be thinking, I must look a lot like this guy. What if the witness points to me instead of the real suspect?
Well, the reality is, thanks to facial recognition, at least 53 percent of all American adults are now in what we call a “perpetual lineup.” This is not because they have had prior interaction with law enforcement, but because they have a driver’s license. The use of face recognition technology by law enforcement is far more pervasive and far more advanced than most people realize.
As a preliminary, I want to briefly go through how face recognition is used by police today. The first use is what we call “stop and identify.” A handful of agencies across the country have face recognition applications on their phones, meaning that officers in the field, after they’ve stopped somebody, can actually look up that person’s identity with a face recognition application simply by taking their picture. This process takes about three seconds. Second, there’s “arrest and identify”—upon arrest in most jurisdictions that have a face recognition system, the person arrested, regardless of whether they’re charged or later convicted, will have their mugshot taken, and that photo will be searched against the existing database and enrolled into a face recognition database for future searches. The third is “investigate and identify.” Let’s say officers have a surveillance video of a bank robbery taking place, or a cell phone video of a theft. They can take that photo, and if there’s a good enough face in any of those stills, they can run that through a face recognition system and search for that individual against whatever database they have.
The fourth and most concerning application of face recognition that we’re seeing is real‐time biometric surveillance. Increasingly, law enforcement agencies are expressing interest in using face recognition at the back‐end of CCTV systems in real time to monitor who’s walking by those cameras.
The other component of face recognition that we need to talk about briefly is who’s in the database. The first and most common are mugshot databases, but as I said before, this is not necessarily limited to individuals convicted of a crime. Increasingly, however, driver’s license photo databases are optimized for face recognition searchability, and now they are open to search by law enforcement agencies.
The final form of databases is watch lists or hot lists—these are lists of individuals that a real‐time system would be looking for right now. Because of technological limitations, these are relatively small lists. But the narrowness and the targeted nature of these types of watch lists are going to disappear and the databases for real‐time systems will look like the databases for investigative searches.
Our first finding is that law enforcement face recognition technology is far more advanced and widespread than most people realize. A Government Accountability Office report early last year indicated that the FBI has access to search 16 states’ driver’s license photos for FBI criminal investigations, representing 64 million people across the country. Our report includes state and local law enforcement’s access to driver’s license photos, and that number jumps up to 26 states, 117 million people as of the launch of our report. We’re continuing to investigate, and this number just continues to grow. We’re now at 30 states and 119 million people, which is 53 percent of all American adults.
We found that six, probably seven major jurisdictions across the country have either looked into or actually purchased the ability to do real‐time face recognition surveillance: Seattle, Los Angeles, Dallas, Chicago, New York, West Virginia, and probably Detroit. Every major company that sells face recognition systems to law enforcement in the United States advertises—and heavily markets—the ability to do so.
Our second finding is that law enforcement face recognition will have a disparate impact on African Americans. This is because of three factors. The first is illustrated by a study in San Diego, which found that people of color were between 1.5 and 2.5 times more likely to be targeted than expected by their presence in the population by advanced technology, specifically license plate readers and face recognition. The second is that most face recognition systems run on mugshots, and if we look at arrests‐to‐population ratios, African Americans are arrested at far greater rates than their proportion of the population would suggest. So African Americans are overrepresented in the searches, the probe photos, and in the databases themselves.
The third prong is the search itself. Taking a step back, face recognition is not very accurate. A few years ago, the FBI ran a test on the searches that they conduct and found that they were about 86 percent accurate, meaning that in six out of seven of the searches they ran the suspect, who was indeed in the database, would show up in a list of 2 to 50 possible candidates. One out of seven searches would result in a list of completely innocent candidates, even though the suspect was in their database. But compounding this, the errors are not distributed evenly. A 2012 study coauthored by an FBI face recognition expert found that face recognition algorithms are 5 to 10 percent less accurate on African Americans, women, and young people. More recent studies continue to demonstrate that these algorithms perform differently depending on your demographics, particularly race and gender.
Our third finding is that face recognition, however pervasive or advanced it is, is not under control. We found that there are no comprehensive state or federal laws that govern the use of this technology by law enforcement. We then took a look at how the agencies themselves are choosing to regulate this and found that a very limited number of agencies had policies to begin with. Very few of those existing policies required individualized suspicion, limited the searches to certain crimes, required suspicion to begin with, or prohibited searches on First Amendment–protected activities. Many agencies told us after our Freedom of Information Act request that they didn’t have any rules governing how the technology is used—some policies went even further than that, to say that law enforcement is encouraged to use this technology “whenever practical.” That line is from the Pinellas County, Florida, sheriff’s office policy. We talked to Sheriff Bob Gualtieri there, in charge of operating the longest‐running system in the country. The system has been used for about 16 years and has never been audited, despite the fact that about 8,000 searches are run on it every year.
This leads to one last question: Is there any judicial control here? We went across the street and spoke to Bob Dillinger, the public defender for Pinellas County, and he said never in his entire time running the public defender’s office—and he’s been around for the entire duration of the face recognition program—has any case had face recognition disclosed as Brady evidence. This is something one would expect, because face recognition gives a list of possible candidates. Any candidate who is listed by the algorithm who was not the person charged would potentially be exculpatory evidence that must be turned over to the defense.
This technology may also be leading to chilling free speech. After the death of Freddie Gray in police custody, the Baltimore County Police Department reportedly used Geofeedia in conjunction with face recognition to take photographs at public protests, run them through face recognition, and identify people at protest sites while the demonstrations were going on. Law enforcement agencies themselves have said in a 2011 report that this runs the risk of chilling free speech, and yet that hasn’t stopped them from using it with no limits.
There is reason for optimism. The House oversight committee did hold a hearing after the launch of our report on the use of face recognition technology in which they put the FBI in the hot seat over its lack of transparency and oversight. We’ve seen action in Vermont, Maryland, and New York, including an introduction of a comprehensive bill in Maryland.
But so that we don’t end on too optimistic of a note, I want to make two quick points about where this technology is going. Real‐time face recognition used in conjunction with body cameras is coming—it’s already in use in the UK. The company who deploys it in the UK has a contract here to use it with dash cams. Think about this—real-time face recognition where the final arbiter of the algorithmic match is not somebody sitting behind a desk who has the time and training to evaluate whether the algorithm is right or not. It’s an officer in the field with a weapon, who has a moment’s decision to make on whether he’s faced with a threat to public safety and to draw his gun. What if the algorithm is wrong?
China is very aggressively deploying face recognition technology on very minor crimes. They use it not only to shame jaywalkers, but also to report that crime to the police when it occurs. The Russian government is very actively using face recognition to crack down on anti‐corruption and anti‐government protesters. They publish protesters’ names online and subject them to harassment if not arrest and incarceration. And then a final note on realtime surveillance: China has 200 million cameras. They’re planning to implement 400 million more in the coming years. They have real‐time face recognition in a lot of these cameras. The BBC just did a report on this where they had the system enroll the face of one of their correspondents, and he was found by the face recognition system within seven minutes of walking out the door. These systems are far more advanced than what we’re seeing in the United States today. But without restrictions, without laws in place to limit these systems, without transparency, without public knowledge about this, these systems are being deployed, and there are very few practical limitations on a U.S. agency deciding to purchase them.