WASHINGTON, D.C.—It’s a brisk late November afternoon in an 8th-floor office overlooking downtown Washington’s Thomas Circle. The White House is an easy five block walk; the Hart Senate Office Building, a 15-minute cab ride. Outside, the streets are filled with people bustling about, protected against the chill in dark suits and authoritative shoes, moving between power centers with the confidence of essential players in the workings of the American government. Here, in his office, Tim Hwang is walking me through a piece of software that is already shaking the ground beneath their feet, even if they’ve yet to feel the rumbling.
Hwang is the CEO of a four-year-old firm called FiscalNote, which makes a kind of technology that is quickly raising questions about who—or what—is still an essential player in Washington. Hwang, in sharp-edged glasses and a blue blazer, taps on his MacBook Air, and what appears on the screen is a full assessment of the legislative record of Senator Orrin Hatch, the 83-year-old Utah Republican.
Hatch’s varied career is the longest ever for a Senate Republican; he’s been a video-game critic and an advocate for the “Ground Zero Mosque,” and in his four decades on Capitol Hill he has championed hundreds of bills and taken thousands of votes both obscure and important. Figuring out Orrin Hatch isn’t a trivial job, even for a seasoned D.C. hand. But FiscalNote has all that data distilled, analyzed and weaponized. The display tells us that Hatch is formidable not just for his seniority, but because he’s in the top 3 percent of all legislators when it comes to effectiveness—or at least he was, before he announced his impending retirement. When he throws his weight behind a bill, it’s likely to become law. What’s more, his effectiveness varies: It’s high when the topic is health, but drops some on tech issues.
The software drills deeper. One immediate surprise it delivers is that the lawmaker most similar to Hatch’s interests and patterns is Louisiana’s John Kennedy, a 66-year-old Republican who’s been on Capitol Hill all of 11 months. Then, with a few more clicks, it’s crunching the woeful record of a shall-remain-nameless member of Congress who occupies the bottom third of legislators in the house, and who, the software dryly notes, is “fairly ineffective as a primary co-sponsor.”
There’s more. Much more. Hwang’s system analyzes interests, not just people, and quickly summarizes everything knowable about who is trying to pass what kind of rules about the most obscure topic I can come up with on the spot: “dairy.” A couple more clicks after that, and we’re looking at a summarized version of a bill tackling cybersecurity that the software has considered and rendered a judgment on, when it comes to the probability that it will become law. We’re not talking a rough estimate. There’s a decimal: 78.1 percent.
This kind of data-crunching might sound hopelessly wonky, a kind of baseball-stats-geek approach to Washington. But if you’ve spent years attempting to make sense of the Washington information ecosystem—which can often feel like a swirling mass of partially baked ideas, misunderstandings and half-truths—the effect is mesmerizing. FiscalNote takes a morass of documents and history and conventional wisdom and distills it into a precise serving of understanding, the kind on which decisions are made. Here, the software is telling us that if we’re looking for an up-and-coming Republican to get on board a health bill Hatch is pushing, Kennedy’s a good bet. Want it bipartisan? The system will suggest likely Democratic backers, too.
If you’re an aide, one of the people walking on the street outside from a power breakfast to a meeting on the Hill, there’s another way to think about what FiscalNote is doing: It’s doing your job. Washington, D.C., is a notoriously imprecise place, trading on memory and relationships and gut. And a huge amount of what people do in the city, the way they make their living, is guiding others through the morass. The things FiscalNote is doing—sifting through murky bills and votes and patterns of behavior—is precisely why you hire an experienced staffer. Without much in the way of human involvement, says Hwang, the system can “enable the top attorney at McDonald’s to immediately understand every single law and regulation pertaining to their industry.”
That’s tremendous power, the kind that threatens to rattle the bedrock of the capital. If there’s one central cog in the modern city of Washington, with its bustle of influence and steakhouses and exorbitant home prices, it’s the in-the-know lobbyist or staffer or government-affairs liaison. There are thousands of them here, paid, often quite well, for that know-how. This machine handily replaces much of that, and without running up huge bills at Brasserie Beck.
Could the swamp really be automated? The question feels almost alien. At the moment, if “automation” and “Washington” are used in the same sentence, it’s usually to decry how behind the curve policymakers are on a transformative economic issue like industrial robots or self-driving cars. In its own workings, Washington seems almost a uniquely un-automatable place, a constitutionally erected edifice of institutions and people driven by irreplaceable experience and relationships.
Hwang is demonstrating that’s not true. FiscalNote isn’t some pie-in-the-sky, grad-school project. The firm employs 160 people today, with 1,300 clients and upward of $28 million in backing from hugely prominent tech industry investors (Mark Cuban, Jerry Yang, Steve Case). Toyota and the National Institutes of Health use FiscalNote to keep tabs on the political realm. Hwang’s also got a healthy client roster among world governments, which need to understand D.C. for their own reasons: FiscalNote is used by the foreign ministries of Canada, Mexico and South Korea. He has a competitor, a bootstrapped firm called Quorum in nearby Dupont Circle, that specializes in giving clients the ability to respond instantaneously to what the political world’s talking about right now.
FiscalNote sits at the front edge of a change that goes far beyond the lobbying world. “Washington” writ large is a dense entanglement of politics, rulemaking, legal work, journalism and jurisprudence—all fields that have seen significant, if often quiet, incursions from machines. Washington’s law firms, a linchpin of the local economy, have already automated much of their paralegal work. Journalism, another mainstay here, is more of a challenge to automate, but that’s happening too: The Washington Post experimented with machine-written coverage during the 2016 Rio Olympics, and is now trying to do the same thing with House, Senate and gubernatorial races in every state in the Union. Stranger still are attempts to inject automation into the judicial branch, inspired by those who argue that computers are better and fairer at some kinds of decision-making jobs than human beings in black robes.
As quickly as technological change is coming to Washington, the profound questions it raises about both ethics and economics—what is “democracy” if it has machines at its core? whither the United States’ capital city if there are far fewer people left?—are lagging behind. It might be time for us to take them seriously. “We’re still going to need a lot of them,” Hwang says of those professionals hustling down the streets outside, “but I don’t think we’re going to need them at the scale at which Washington operates today.” When it comes to the nation’s capital, he says, “People vastly, vastly underestimate what automation is going to do.”
Washington, D.C., to be fair, was never all that congenial for humans. It started as nothing—a swampy spot on the Potomac River conveniently located near Mount Vernon—and grew in fits as the nation did, and as the federal government become more ambitious. When Civil War General George Henry Thomas, whose statue stands outside Hwang’s window on Thomas Circle, was at the height of his career, the city’s population hovered at around 130,000. The population is now almost quintuple that—if the District were a state, it would be larger than both Wyoming and Vermont—and the booming metro area is one of the nation’s largest and wealthiest.
For all the changes, many things about Washington would still be recognizable to those who lived here in Thomas’ time. The name of the game is still trying to influence how the national government makes and carries out its laws. Nearly every job tied to official Washington falls under that banner; a 2013 study by the economic modeling firm Emsi found that half the jobs held by people in Washington can be “explained by the federal government.” When it comes to the lobbyists threatened by Hwang’s software, their numbers are, on paper, small: Only 10,960 people are at the moment registered as lobbyists in the city. But that figure is supremely deceiving. Influence is a far larger ecosystem of “consultants” and law partners and their many junior staffers; the Center for Responsive Politics found that in the first quarter of 2017 alone, companies spent some $838 million lobbying in Washington.
The people doing those jobs likely feel irreplaceable, but the pace of technology suggests otherwise. As artificial intelligence grows more sophisticated, it gets better-positioned to take on the sort of professional jobs that we’ve long thought were beyond automation’s reach. In his 2015 book, Humans Need Not Apply, Jerry Kaplan of Stanford University warned, “Automation is blind to the color of your collar.” Bill Welser, director of the engineering and applied sciences department at the RAND Corporation, co-authored a study that included a matrix for determining which jobs are more likely to be done by machines in the future. Its conclusion: The jobs at lowest risk of automation are ones where there is a high degree of “chaos” and little time for the humans who fill them to react. Kindergarten teachers, says Wesler, seem around for the long haul.
But, Welser says, where there are “tremendous numbers of white-collar, college-education positions that likely to date have felt very insulated from automation,” RAND’s matrix suggests that plenty of them should feel exposed. CEOs, for example, says Welser, often operate at low levels of chaos, and by virtue of their positions, are granted the freedom to make decisions on a longer time scale. (Given what CEOs are paid, perhaps it’s surprising that no board of directors has tried to replace one with an algorithm.) Drilling down in Washington, Welser says, if you look at that low-chaos, low time-pressured corner of the matrix, you find the policy analysts—those professionals who fill the ranks of institutions across Washington, from, say, the Department of Education to the Council on Foreign Relations.
John Fernandez is an Obama-era Commerce Department official who is now the global chair of a venture capital firm called Nextlaw Labs, which focuses on applying cutting-edge technologies to the legal world. When it comes to automation, he says, government is “a target-rich environment, there’s no question about that.”
Some of Washington’s most iconic jobs are especially low-hanging fruit. For one, there are the entry-level workers on the front line of Capitol Hill, the ones dealing with the influx of constituents who call or email or write letters. Philip Resnik is director of the computational linguistics and information processing lab at the University of Maryland. “My son Ben had a friend who interned in one of those offices and was dealing with that,” says Resnik, who also advises FiscalNote, “and I was like, ‘Please, let me get my hands on the data,’” envisioning how easily machines could process it for meaning. If a member of Congress wants to take the temperature of his or her district, an algorithm might be of more use than the legislative correspondents who dot Capitol Hill.
On the other end of the spectrum is the White House, says Christopher Lu, who served as President Barack Obama’s point man with his Cabinet from 2009 to 2013. At 1600 Pennsylvania Ave., Lu says, “You’re really doing the higher-level decision-making. I’m not saying everyone’s indispensable there, but it’s much harder to get rid of those people.”
Lu is a classic D.C. figure: a high-level liaison with an intimate view of how power works. He’s not, though, ensconced in a lobbying firm: He now works at FiscalNote as its senior strategy adviser. Lu and Hwang both attended Wootton High School in Montgomery County, Maryland. Both attended Princeton too, and Lu, Class of 1988, spoke at Hwang’s 2014 college graduation there. Lu says that when he’s out trying to woo would-be clients for FiscalNote, he tries to deliver a realistic, and perhaps comforting, pitch: “I can’t replace the conversation you had with a [congressional] staffer, or if you’re at a cocktail party and a member of Congress says, ‘Hey, this is going to happen, and this is not going to happen.’ But we can give you a more comprehensive landscape” of how things are likely to transpire.
The company’s chief competitor is a 10-minute walk from FiscalNote’s office, just south of Dupont Circle. Quorum was founded in 2014 by Alex Wirth and Jonathan Marks while they were still Harvard undergrads. Now employing up to 46 people, stuffed into a buzzing glass-partitioned office space marked by caramel-colored wide-board wood floors, the office has a map of the D.C. Metro system embedded in tile on the wall of the open kitchen space. Next to the laptop of one programmer feeding data into the system is a stress toy in the shape of the Capitol.
The Quorum platform—which looks a bit like Facebook in a suit, its crisp aesthetic characterized by a sober army blue—pitches itself as operating on the forefront of “data-driven politics.” Its specialty is another mythical D.C. work product: buzz. Quorum tries to track, in real time, the topics people in and around Washington are talking about, whether that’s a tweet or an email to constituents or a Congressional Research Service report. What’s more, it gives users the tools to respond, a sort of actionable intelligence that has attracted a broad client roster. Quorum’s integrated monitoring and outreach system is how both Walmart and the Sierra Club engage with Capitol Hill. Same goes for Etsy, Apple and the National Restaurant Association.
The human version is staffers scouring Hill websites, surveying Twitter, tracking co-sponsorship lists, trying to get a sense of where, exactly, Congress is settling on an issue. Says Wirth, “We can automate all of that, and create the same spreadsheets in 10 minutes that used to take an intern a week to do.” Once an organization gets an alert that its issue is being talked about, it can respond en masse with a few clicks via email or tweet to any slice of the 31,000 federal employees in the Quorum system, with messages automatically customized to pull in what the system knows about the target’s past activities on an issue. “It’s just like Amazon,” says Wirth, pointing to checkboxes that allow a Quorum user to dictate to the machine the contours of its desires. “I want to filter it to be just the chiefs of staff,” for example, “and there we go. I just hit submit.” Boom, done.
The system seems shockingly knowing. How many staffers sitting in their offices on Capitol Hill, I wonder aloud, get one of these machine-customized emails and have no idea that the only human involved was the one who set the filter? Wirth and Marks laugh, and the former says, “That’s the goal.”
Take a step back, and it seems exceedingly likely that there’s almost no slice of the Washington economy that won’t be reshaped by automation. The capital is one of the biggest job markets for journalists in the nation, and much of that work isn’t much more than quickly getting information to the right people. Do reporters really need to stay on Capitol Hill past midnight just to note that, say, H.R. 1789 has passed on a party-line vote right before Congress left town for recess? How about election results? The Washington Post’s in-house automation tool, called Heliograf, currently cranks out simple, formulaic sports stories built on box scores, and the publication is experimenting with using it for elections. (Human editors have the option of adding to or overwriting the bot’s prose.) The publication is experimenting with robots as information scouts—for instance, alerting reporters via chat messages when a campaign is playing out in unexpected, coverage-worthy ways, like a “safe” incumbent suddenly struggling. The Associated Press, meanwhile, produces automated reporting on corporate earnings calls, and is looking to get more ambitious. One project floated by AP researchers would use artificial intelligence to constantly read satellite images, letting environmental journalists know when there are signs of deforestation to follow up on.
Quorum says its system also powers one well-known late-night comedy show—it won’t reveal which one—helping staffers scan YouTube to pull together almost instantly a selection of footage about topics they might use on that night’s show. Says Wirth, “They’re like, ‘What kinds of Roy Moore clips do we have?’” The software has already machine-read heaps of video footage, turning them into tagged transcripts, and sends off exactly the correct clip to that evening’s show. Millions of Americans, without knowing it, are getting their political news curated for them in part by a Quorum bot.
Law firms across the country have begun depending on automation to churn through the sort of “discovery” work that used to be the job of new lawyers, gathering and making sense of evidence from the opposing party. But that’s just an early foray. The law firm BakerHostetler—the first firm of record in House Republicans’ 2014 Obamacare lawsuit—has adopted a system called ROSS that, using IBM’s proprietary AI platform Watson, answers queries about complex legal issues much as an Amazon Echo answers our dumb questions about movies and the weather.
As these examples suggest, a key function in D.C.—one that stretches from law firms to federal agencies—is the routine processing of huge amounts of data, and automation is already at work there. Back during the implementation of the 2009 stimulus package, the oversight board that was created to keep tabs on the bill’s roughly $800 billion expenditure contracted with the Palo Alto-based company Palantir to use automated data mining, looking for relationships or patterns of behavior that might help investigators discover fraud—what the company calls “human-driven, machine-assisted analysis.”
More than that, though, Washington runs on the sort of judgments made by agency staffers day in and day out, such as whether an American qualifies under the rules of one of the scores of government programs. That’s a step up the ladder from just crunching information: It’s a judgment call, and it could very well be done by a machine. Jennifer Doleac, an assistant professor of public policy and economics at the University of Virginia, is one of the handful of social scientists in the United States closely studying the practical and ethical questions of extending automation into public decision-making. “You could imagine feeding information into a computer that says, ‘Yes, this person’s eligible for benefits or not,’ instead of just looking at a file and say ‘yes’ or ‘no’ based on their hunch about whether the person needs the money,” says Doleac.
One of the biggest information blizzards in D.C. comes down when agencies open their public comment period on a new rule or regulation. The system is supposed to give the public a say in how the federal government interprets the laws Congress passes, but the ease of Internet commenting has opened the floodgates, making it exceedingly difficult in some cases for a federal agency to make sense of them. The Federal Communication Commission’s push, for example, to write new rules on so-called net neutrality brought some 22 million comments, or 13,000 for every one of the FCC’s employees. It’s a broken system, one that perhaps only automation can help fix.
Lu, the FiscalNote adviser, thinks back to the Department of Labor’s push to write new rules for a controversial expansion of overtime benefits, back in 2016. “You get thousands of comments, and we had no ability to process them,” says Lu. “We had outside contractors, and I don’t even know how they did it, but I strongly suspect some poor person printed them out and then sat in a conference room. We’d periodically call and say, ‘Hey, which way are the comments going?’ And I guarantee some poor person had to look at the stacks of paper.”
FiscalNote has pulled in every one of those 22 million FCC comments and organized them for clients on a scatterplot showing what the company calls stance detection—not just whether the writer is for or against the new rules, but, if they don’t like them, why not. It piles them into buckets—outside the scope of authority, too broad, too narrow. Says Lu, “You hit the thing and just instantly scan the comments. Hundreds of man-hours were saved overnight.”
Although automation is sometimes seen as a threat, the existence of 22 million comments on anything suggests that it’s also a way to level a playing field that automation has already tilted out of human control. Reports have found that millions of those comments appear to be generated by bots, some based on fraudulent identities. Even that question is complicated: Is a minimally customizable form letter a “bot”? FiscalNote, for its part, tried to sift through this with an AI hunting for telltale patterns of “natural language generation” in the comments, or wording likely produced by machines trying to look human. We aren’t far from a future where public commenting on regulations—the process for individual American citizens to offer feedback to their elected government—comes down to a bot vs. bot fight.
The economic argument over automation hinges on one key question: Overall, do new machines replace humans, or just change human roles to make us more productive? Hwang, the FiscalNote CEO, takes the latter view: “You can imagine it like having a bionic arm,” he says, pitching his software as a way to give savvy operators an advantage over their competitors. But if you’re an employer and soon you have enough employees working with bionic arms, the day comes when to do the same amount of work as in the past, you simply need fewer workers.
When it comes to governance, the questions are ethical as well as economic. We elect lawmakers, and appoint judges, to make human decisions about the biggest questions that affect our lives. Robots are making their way into those fields, too. Across the country, criminal courts have started to rely on algorithms to help judges decide things like whether a defendant should stay in jail while awaiting trial. Doleac, the UVA professor looking at the future of algorithmic decision-making, founded and runs a center at the school called the Justice Tech Lab in response to requests from law enforcement and others for help figuring out how to best apply new technologies to criminal justice. “Humans are easily distracted by irrelevant information,” says Doleac, “and computers are much less easily distracted by that stuff.”
The factor that limits whether robots start rendering more decisions in the criminal justice system is less technological than that “people aren’t totally confident in computers making decisions yet,” says Doleac. But that doesn’t mean robots can’t augment those judgment-renderers’ brains. While the Supreme Court, Doleac says, as a rule hears cases too complex for algorithms to help much, one place they could is on deciding stays in death-penalty cases. A computer could tell the justices how likely the case before them is to eventually be overturned, giving them insight into the wisdom of stopping that particular execution. (Robots are already watching the Supreme Court with increasing precision: AI forecasters, digesting details of the cases, how judges interact with the lawsuit participants and other granular data, say they are now able to predict how the court will go 75 percent of the time.)
Could there be robots usefully serving in Congress? Writing laws? Making, with no human oversight, the day-to-day decisions of governing? Technologists are quick to say that the technology isn’t there yet, but academics aren’t waiting around to consider the complex implications of the likely day, not too far off, when citizens will be forced to wrestle with the question. A February 2017 workshop at the University of Pennsylvania looked at the question of government “for the people, by the robots.”
Transparency is already emerging as a concern: At the workshop, experts discussed how digital systems can be easily designed in ways for which no human, in the end, can objectively be held immediately responsible—the digital equivalent of an unaccountable firing squad. What’s more, they said, there are cases in which algorithms simply can’t be made fully transparent to the humans they affect, either for reasons of safety, such as the formulas that pick people out of the line at airport security, or simply because they’re too complex and by the nature of artificial intelligence always changing.
And the human animal might not be all that well-adapted to trying to demand even a minimal level of accountability. We as people, RAND’s Welser says, have proved ourselves too eager to cede authority to these machines, not paying much attention to not being able to see inside what amounts to black boxes. Automation bias, he calls it: the idea that people assume that what an algorithm spits out must be logical, right and good.
On the flip side, though, is an argument UVA’s Doleac makes strongly, one you hear a lot in the debate over self-driving cars: The machines we’re coming to trust with our lives don’t have to be perfect. They just have to be better than humans. And when it comes to the courts, says Doleac, judges make mistakes. They can, like other humans, be ruled by hunger and anger and misunderstandings. “It’s important to remember that the counterfactual”—that is, human-driven decision-making—“is not some sort of perfect truth,” she says. And even if robots are only just as good as their human counterparts, she points out, they’re cheaper and easier to replicate.
After automation comes, what’s left for humans to do? Vlad Eidelman leads the data science team at FiscalNote. He argues that AI is no different from other major technological advances. “Each time, the types of jobs that exist grow closer to the more philosophical and cognizant,” he says, saying that’s what he predicts will happen in Washington.
But his boss, Hwang, isn’t so sure. What makes us think, he asks, that members of Congress will use all their new free time for nobler purposes? They may well spend it engaging in even more intense scrambling to become, say, the chair of a congressional committee.
Hwang says he also gets to pondering whether the judgments that FiscalNote renders on just how good—or “effective”—a lawmaker is robs the American voter of his or her right to decide how well an elected representative is behaving. And what if, he asks, FiscalNote gets into the hands of even more clients: Is a legislator doomed to a life of ineffectuality just because the company’s machines have told the world that he or she is a lost cause?
“And we can be discussing that,” Eidelman jokes, “because now we have time to discuss metaphysical questions.”
Talking to Eidelman, I realize that I’ve been gripped by automation bias, so caught up in effectiveness that it was easy to forget that it’s just one algorithm’s understanding of the point and purpose of the legislative branch. Eidelman built it, and though he says with a laugh that it’d be “very egotistical” to say that it’s Vlad Eidelman’s version of what constitutes effectiveness, he acknowledges that, yes, at the end of the day, it’s largely his interpretation of how Congress does and should work.
Voters, on the other hand, might have very different definitions of effectiveness. “Someone can come to me and say, ‘Well, I got elected because I want to stop everything from passing, and I’m effective because literally nothing has passed on my watch,’” says Eidelman. Indeed, out in America, voters have fairly often proved themselves perfectly happy to send and re-send to Washington lawmakers paid close to $175,000 a year who refuse to actually make laws. Both Senator Tom Coburn and Representative Ron Paul were often nicknamed “Dr. No” by detractors and fans alike. Coburn was elected twice by Oklahomans before stepping down in 2015, and Paul represented a slice of East Texas for 37 years, a remarkable run. In those cases, FiscalNote’s robots would have deemed them remarkably ineffective. But their constituents back home might have thought them heroes, for precisely that reason.
In D.C. itself, with a workforce so heavily dependent on the federal experiment, local leaders have begun thinking through how to diversify the city so that its future vitality might be something more than just a gamble on government. One place some thought leaders are looking is to the north, New York City, where both lingering economic uncertainty and the quick entrenchment of automation are leading to fewer jobs in the financial sector that serves as that city’s foundation.
How Wall Street comes to terms with the impact of AI, Hwang says, should provide some lessons for Washington—or at least some motivation to start coming to terms with the idea that change is imminent. “What does Washington look like,” Hwang ponders, “when you hollow out large portions of trade associations?”
D.C. Mayor Muriel Bowser has put a focus on growing the city’s burgeoning tech sector. In 2016, she set out a plan to create 5,000 jobs in the local innovation economy that can be filled by “underrepresented workers.” A spokesperson for the mayor said, “We want people see the District as more than a government town.”
As part of that push, the city has lured Apple into opening a flagship store in the old Carnegie Library downtown, which the company will help renovate. It has also worked with San Francisco-based Yelp to open an office here, and given a whopping $750,000 grant to help one of the city’s most promising tech firms keep its headquarters in the District.
That company? FiscalNote, and the irony is that should Hwang’s company be wildly successful, it could, in the end, contribute to some of the very hollowing out of Washington that the mayor and others are worrying about.
At least Bowser is worried, Hwang says. It’s exactly the kind of concerned leadership, he says, that we need from the humans we’ve elected to see around the corner and prepare for the future. “Otherwise,” Hwang says, “we end up like any other city, in decline from an automated industry.”