Burn-In

by P. W. Singer

  Strange. It hadn’t projected onto the SUV’s screen. Keegan brought up the local police net on her vizglasses, but the checkpoint wasn’t there either; maybe it was a glitch or just not logged yet . . . or given how the cops were wearing masks to cover their faces, maybe this one wasn’t official.

  “We’re not in Kansas anymore,” she said, talking to herself as much as to the robot.

  The cops recognized the government-issue SUV and waved them up to the front of the line of vehicles. At its head, a policeman was standing beside a purple two-door electric car, holding the license of its young driver, his other hand resting on a pistol. This was the sort of unplanned opportunity she had hoped for to test TAMS’s ability to read the environment.

  “Tell me what you notice about both the officer and the person who they’re questioning.”

  “Both are exhibiting physiological indicators consistent with high levels of nervousness,” TAMS said.

  “That’s fear,” said Keegan. “Both have a valid reason for it, which is what can lead to a very bad outcome.”

  As the SUV began moving again, Keegan switched on the onboard ThreatView. Modeled after the heads-up display used by Air Force fighter pilots to track, it projected various security information collected in the area onto the screen.28 She tabbed it to track any aerial criminal activity. ThreatView showed a writhing column of climbing and descending drones working the area. Most likely a drug dealer sitting on a rooftop sofa in the shade, operating a distribution network.

  Keegan gamed out what she would do if she were dealing. She’d zig when everyone else was zagging, go underground; use squid crawlers to navigate the sewer systems so you were going straight to the customer. Every building, be it a luxury penthouse overlooking the National Mall or a boarded-up shooting gallery, already had a perfect clandestine delivery system, standardized and usually working: a toilet connected to a sewer main. You could have squids learning on every run, teaching one another as they worked their routes. As she worked through the crime in her head, she thought that this was one of those moments when it was good bots couldn’t yet read minds.

  They drove on another few blocks into Southeast DC. Stopping at a red light, Keegan scanned the corner. A dozen people were boarding an automated trolley bus, the spiderwebs of its cracked windows glowing with refracted light.29 Maybe someone had thrown rocks at it for fun. Or maybe it had been an act of defiance, lashing out at what the driverless vehicle represented. It didn’t matter, though. The buses were essentially disposable, their likely loss baked into the plan. There was a term for that, Keegan knew, because they had also used it to describe the bots she’d used on deployment and even sometimes the human troops. “Attritable.”30

  “Check out the bus,” she instructed TAMS. “Are all its systems running normal?”

  A few months back, some hackers had run a ransomware hit on the bus system, locking the passengers inside as the automated vehicles drove across the city. As far as attacks went, it was a pretty dumb one; people who ride public transit aren’t flush with extra cash. The bus manufacturer had eventually paid up for them, mostly to get the story of people being taken hostage on a robot bus to stop trending.31 But the fact that the hack had been traced back to Romania had made it international, and thus put it on the FBI’s radar screen.

  “All systems running within normal parameters,” TAMS replied.

  In the end, the lawsuit meant that the passengers would likely make out better than the hackers had.

  The SUV automatically slammed on its brakes, letting a red Volks­wagen hatchback driving without its lights on cut across the lane for driverless cars. The driver slipped a hand out the side window with a quick wave of thanks. Keegan responded with her middle finger.

  Fully aware of her own hypocrisy, she looked over to gauge TAMS’s response. Nothing. It’d be interesting to know how it factored that. Was it going to use it as data for understanding Keegan’s psychology? Or had she just created the first kernel of some future moment of robot road rage?

  Keegan tapped in a new destination. As the SUV turned off the arterial, its dash screen shifted—red and green dots popped up in clouds, showing geographic concentrations of past crimes and predicted crimes based upon law enforcement data, as well as private-sector information that fed into the same algorithms.32

  They drove up a street of low buildings that were a cluster of both colors. The shop fronts were covered in plywood, topped by makeshift apartments hidden behind thick curtains and cardboard. A light wind lifted up translucent food wrappers that drifted in the breeze.

  She slipped her vizglasses on and tagged one of the buildings, a closed-down butcher’s shop, whose windows had been replaced with what looked like garage doors. Small black crescents covered the front. The bottoms of the protective shielding looked burned, maybe melted. Next door was a burned-out shell of a small industrial space, fingers of charred wood and steel stabbing skyward like a rib cage. A calico cat lounged in the cool shadow of what had been the entrance, lying contentedly in the way that only a cat could, as if the destruction around it had been done by the humans just for its benefit.

  “Describe what you see over there,” Keegan said, tagging the scene with her glasses. “Tell me the highlights.”

  “In the foreground is a concrete sidewalk. On top of it is a male cat with gray, black, white, and orange coloring. Behind it are two buildings, both 7 meters in height, 12 meters in width. They were damaged by fire. Signs identify the location as the former business of A-1 Butchers and the Capitol Technology Corridor Site 32.”

  TAMS had not actually “seen” that, but rather used a neural network. Modeled after the human brain, it corresponded to the information that it actually detected, the digital pixels in the image, to extract information on what those pixels might represent, like the color of the fur of the lazing cat. Indeed, it didn’t even detect the color originally, but rather turned the whole image into grayscale, which represented the intensity of light matched to the standard color spectrum a human could see. Just as the neurons in a human’s brain are dedicated to detecting specific objects, that grayscale had then been turned into a matrix of numeric values in TAMS’s calculations, each and every pixel in it given a number between zero and 255. Then, in a little over a picosecond, the neural network had gone to work, each pixel filtered again and again by a set of votes on what was in it, each layer of winning votes then pulling out information on the tiny details of each part of the image. In this way, the network determined everything from whether the pixel depicted an edge or a curve or a color, which was then fed to the next layer. Ultimately what was known as a Densely Connected Convolutional Network pooled all those pixels, votes, and layers together to recognize what the data matched of most likely objects.33 All just to “see” a “cat.”

  “Tell me more,” Keegan demanded, nonplussed by the recognition software. “What additional information can you determine about what you are seeing?”

  “The cat is of mixed breed, popularly known as a calico,” TAMS replied. “Its fur pattern and size indicate a mix of Maine Coon and domestic shorthair. It lacks a collar and is unchipped, classifying it as feral. Would you like me to notify Washington, DC, Animal Care and Control Field Services?”

  “No, forget the cat,” Keegan said. “What can you conclude from the writing on the wall?”

  The bot emitted a faint hum, as if considering the question’s validity and the energy required to respond. “The unoccupied retail space was formally leased by A-1 Butchers. It is currently unoccupied and has been renamed Progress Equals Pain, Fuck Your Future.”

  Keegan snorted. “That’s graffiti. Somebody painted that with a spray can, to deface it, not to rename it.”

  Silence.

  She tabbed the image she captured and blew it up larger on the SUV’s screen. “Here. This is a sign, just not the kind that you think it is.” She drew a circle around “Progress = Pain” and another with her finger around “FUCK YOUR FUTURE.” “What d
id the creators of this imagery mean by that?”

  “Graffiti writing in paint is an expression of anger, of individual feeling,” said the bot.

  “Close. But this is collective, not individual,” Keegan responded. “They are expressing what millions of people feel. Anger at the local scene, but also the larger environment they feel caught in.”

  The machine looked back, calculating what she meant. Or just gathering more details on the cat.

  “And how about those black marks on the door, the round ones?” Keegan asked.

  That she didn’t know, and she was curious to see what TAMS would guess.

  The response was immediate, a faint vibration, almost purring, as if this was the kind of problem the bot knew it could solve. It collected billions of pixels and then evaluated them across not just the visual spectrum, but with comparisons, using all the tools that a single human brain lacks.

  “Those are impact marks consistent with a hockey puck, most likely the activities of a child athlete at practice.”

  “Makes sense,” said Keegan. “Good observation.”

  “Thank you for the compliment,” TAMS replied.

  “Setting change,” Keegan ordered. “When on mission, no acknowledgment of praise or other social convention protocols.” She saw no need for the faux conversation tools that they often packaged the machines’ user interfaces with. It was an implement for work, utilitarian; why make it a Chatty Cathy as if the machine were your friend?34 TAMS didn’t actually feel gratitude, so there was no reason to fake it.

  Keegan slowed the SUV to park in front of the burned-out building. “So what happened here?” she asked the machine. “Give me the recent history of this site.” She wasn’t sure, but she could guess and wondered how her instinct would align with what TAMS pulled.

  “Sixteen months ago, the Capitol Technology Corridor opened a micro-factory at this location,” TAMS began. “Nine hours after the opening of the micro-factory, a fire broke out that destroyed the facility. The cause was arson. Five people were charged.”

  “Timing is everything,” said Keegan. “Anything further of note?”

  “The building was going to print actuated pincer-arms for Task-Shop bots,” said the bot.

  And that explained it all to Keegan. Task-Shop bots had been developed to deliver items from tractor trailers directly to the shelves of stores.35 Combined with already automated warehouses and inventory check drones, the entire supply chain that the old grocery and convenience stores ran on had been disrupted.36 The savings had been immense; what took human grocery store workers hours, a machine could do in minutes, and it allowed the big tech firms to do to supermarkets what they’d already done to bookstores.37 It had also knocked out of the market some eight million jobs that people in neighborhoods like this used to depend on.38

  “However . . .” The bot paused, processing the answer, but also adding to the perceived importance of its finding. “There are indicators of a larger criminal conspiracy. The participants in the arson were subsequently paid three Bitcoin each by a member of Local 400 of the United Food and Commercial Workers International Union.”

  “Now that is useful,” Keegan said. “How do you know this?”

  “An all-source synthesis of the case’s court documents, cloud-harvested personal biographical data sets, and financial transaction information,” said the bot.

  “Good,” said Keegan. “Why didn’t they rebuild the site then?”

  “Capitol Technology Corridor deemed the location to be no longer viable as an automated activity site, due to security reasons,” said the bot.

  Meaning, Keegan thought, the industry of the future had run into the politics of the future.

  “If you gather enough information, everything can be linked to together. That’s not the hard part. The hard part is figuring out what links actually matter and what else is just random noise,” she said.

  “Instances of coincidence are commonly cited in popular media, which indicates—” said TAMS.

  “Stop. There’s no such thing as a coincidence when you’re on a case,” said Keegan.

  As they drove past a low block of cork-colored foam-fab apartments, TAMS pushed a notification flagging a man sitting in an old office chair pushed out onto the sidewalk: Person of Interest: William Leonard of Washington, DC. 71 years old.

  “Why did you send me that?” Keegan asked.

  The man’s face popped up on ThreatView. Leonard had an arrest record that showed charges for armed robbery and trespassing.

  “What is he doing now?” she asked TAMS.

  “He is drinking a 32-ounce can of Ox Blood Slambucha, purchased at the Crazy Eight store.”

  “Where did you get the source?”

  “It is contained on the bottle’s product label.”

  “You read the barcode?”

  “Yes.”

  Keegan thought about how somewhere in the police bot there was the same software used by stores for automated checkout. One scanned the items as you placed them in your cart, rather than going through some line with an underpaid human, while the other allowed the police to backtrack the purchases and past locations of somebody on the street. The whole surveillance cycle very likely funded by a VC firm.

  “So, you’ve got an old guy with a record drinking in public. What is the reason to make him a person of interest now?”

  “Leonard has a 66 percent chance of being arrested during the next thirty days,” TAMS said.

  “Arrested for what crime?”

  “That is not identified in the model,” said TAMS.

  “So what are the key causal factors then?”

  “Mr. Leonard’s past criminal record and his current public consumption of alcohol create a risk matrix of significant accuracy.”

  “That’s a risk, nothing more,” said Keegan. “You don’t know what goes on in a person’s head. We change day to day, minute to minute, second to second. Maybe that next sip . . .” Keegan pulled up a close-in view of the man drinking on the bench. “Gives him that clarity, some insight. Maybe he finds God and, fortified with liquid courage, hangs out with the church choir from here on out.”

  TAMS’s head spun to keep tracking Leonard as they pulled away. “He does not show any association with a church or religious institution.”

  The robot was probably right, but she didn’t want to cede the argument it was having without intending to. “What makes you think you can forecast a person’s actions by their connections?” she asked.

  “An individual’s connections impact their life choices.”

  That sounded like a canned line, like some wisdom that TAMS had stripped from an online self-help article. The machine’s intellectual tendrils could reach far into the cloud, but that also made it unclear when it was an original thought.

  “No, it is not that simple,” said Keegan. “Take me. I’m connected to plenty of people, but which ones were my friends? Which ones did I actually listen to? And which ones did I just tolerate when I ran into them in college or at a party?”

  The screen displayed an image with Keegan in the center of it. It was a photo from an Instagram account she had closed down years ago. Then tens . . . then hundreds of profile pictures popped up around her photo. School class photos, selfies from parties, headshots from online resumes, each popping up and then morphing into colored dots. As a link analysis began to build, the dots connected into lines so dense they appeared as clouds.

  Shit, thought Keegan, seeing faces with hazy familiarity and one in particular she wished that she could forget. “Take it down. I didn’t mean that question as an actual command,” she said, and the image disappeared from the screen.

  Frustration was not an emotion that TAMS would ever feel, but it was welling up in Keegan.

  “Let’s walk,” she said, pulling the SUV into the parking lot of a boarded-up McDonald’s on the corner. As they exited the vehicle, Keegan could swear she still smelled fries in the air. A siren blared in the distance and then
faded.

  TAMS flashed to Keegan’s vizglasses that it had lost connection with the cloud, that access to Wi-Fi or 5G was not available.

  She scanned the area and noticed an unusually large bird’s nest in the crook of the golden arch.

  “There’s a jammer, up there,” said Keegan, tagging it with her vizglasses for DC police to take down later. They probably already knew and just avoided this block for what it signified. Someone wanted this to be a dead zone for a reason.

  She led TAMS two blocks to avoid whatever was going on.

  A pair of elementary-age boys riding a BMX bike, one astride the rear pegs, chased a middle-school-aged girl on a unicycle. She bunny-hopped it off curbs by using her body weight to bounce the underinflated tire, laughing each time she lifted off. Seeing Keegan and TAMS, the kids skidded their bikes to a stop. Keegan thought they would come over for a look, naturally curious about the bot. Instead, they shook their heads as if in disapproval, and then rolled off up the street and turned at the intersection. She couldn’t blame them. So much bad data had been fed into the policing platforms that the machines usually turned out to be even more biased against the people in this neighborhood than their human creators.39

  From a nearby townhouse, Keegan heard someone sitting on the stoop cough up phlegm and spit onto the sidewalk. She slowly pivoted to see a young man in his twenties wiping his mouth with the back of his hand. She quickly looked past him, off into the distance, avoiding eye contact to defuse the situation. It just wasn’t worth it—no different than if she were back patrolling with a squad of Marines.

  It didn’t work. The man stood up slowly, as if measuring the moment. Arching his back to stretch, he drew up to his full height, at least six and a half feet tall. He wore black sweatpants, a faded gray and blue camo Crye combat shirt with the sleeves cut off, and a pair of neon blue indoor soccer shoes. TAMS flagged Keegan on their closed-network connection that its gun-detection algorithm had identified the bulge in the man’s waistband as matching the outline of a Glock 43.40 The subcompact pistol was marketed by the maker as “ultra-concealable,” but object recognition software used changes in light patterns and angles to work around that.41