And when it comes to economic forecasts, the authors cite a study conducted by sociologist Chris Snijders, who “used 5,200 computer equipment purchases by Dutch companies to build a mathematical model predicting adherence to budget, timeliness of delivery, and buyer satisfaction with each transaction. He then used this model to predict these outcomes for a different set of transactions taking place across several different industries, and also asked a group of purchasing managers in these sectors to do the same. Snijders’s model beat the managers, even the above-average ones.”
Even politicians could be overtaken by algorithms. The online news site Politico.com recently published a story with the provocative title “Could a Robot Be President?” The article noted that, after the many blunders committed by President Trump, “a small group of scientists and thinkers believes there could be an alternative, a way to save the president—and the rest of us—from him or herself.” It also added that “unlike a human, a robot could take into account vast amounts of data about the possible outcomes of a particular policy. It could foresee pitfalls that would escape a human mind and weigh the options more reliably than any person could—without individual impulses or biases coming into play.” Furthermore, robots can’t be bribed, and they can’t be influenced by lobbyists. The article also suggested that, since robots are programmed by humans, elections wouldn’t be about deciding the presidency, but about deciding who would program the robot-president.
All this may seem like science fiction, and I myself am quite skeptical about the benefits of giving computers so much power. After all, machines are programmed by humans, and they can make very dangerous decisions if their programmers make a mistake. But it is important to note that growing numbers of scientists are saying that artificial intelligence–powered robots will be able to make much more judicious decisions than humans and that a growing number of those humans are beginning to believe them. If this quasi-religious faith in data continues gaining ground, how long will it be before algorithms break into the fields of investigative journalism, economic predictions, and political decision-making?
AT THE WASHINGTON POST, BOTS ARE ALREADY WRITING POLITICAL ARTICLES
Very few people realized it at the time, but The Washington Post reached a technological milestone in November 2016 when it reported that Republican congressman Darrell Issa had won the highly contested race in California’s 49th congressional district. At first glance, the story seemed much like any other that the paper’s reporters had written that day: it stated that “Republicans retained control of the House and lost only a handful of seats from their commanding majority” and put the facts in context by noting the results were “a stunning reversal of fortune after many GOP leaders feared double-digit losses.” The article went on to report that votes had been counted in 433 congressional districts and that Republicans had won 239 of them to the Democrats’ 194. Some time after the article appeared, the tech magazine Wired published an article noting that election dispatches such as this “came with the clarity and verve for which Post reporters are known, with one key difference: It was generated by Heliograf, a bot that made its debut on the Post’s website last year and marked the most sophisticated use of artificial intelligence in journalism to date.”
Indeed, the story didn’t carry any reporter’s byline. Instead, the byline read “From staff and wire reports, powered by Heliograf, the Post’s artificial intelligence system.” That line, in fine print at the end of the story, went all but unnoticed. Amid the aftershocks of the political earthquake of Trump’s unexpected election victory—against the predictions of nearly every poll—almost nobody noticed the fact that one of the world’s leading newspapers was now publishing political stories written by robots.
Up to that point, it was no secret that certain sports and financial stories that consisted largely of citing data—like football game results or companies’ quarterly earnings—were being generated by bots. The Washington Post itself had been experimenting with automated journalism during the Rio Olympics a few months earlier. And the Associated Press had been using the software known as Automated Insights for writing sports and financial articles for years. But when Amazon founder Jeff Bezos bought The Washington Post in 2013, the paper started experimenting with computer programs that could create more analytical articles. So during the 2016 elections, the Post began running—in plain sight, but without much fanfare—its first political pieces written by an intelligent machine.
Thanks to this new technology, The Washington Post was able to cover the results of roughly five hundred local elections that year, which otherwise would have required an army of journalists and cost a fortune in travel expenses. By publishing articles written by Heliograf, the paper was betting on increasing its readership by reaching new communities across the country. Instead of just trying to win over a large audience through a few high-impact national stories written by people who needed a lot of time to report them, Heliograf allowed the Post to target a number of geographically diverse audiences through a massive amount of automated stories on specific local races. Few New Yorkers would care about the results of an election in California’s 49th congressional district, but millions of readers around the country were interested in the outcomes of elections in their own districts. Heliograf could get that specialized information to all of them in a matter of seconds, and it could automatically update the data in real time.
THE NEWS THAT WRITES ITSELF
Jeremy Gilbert, the director of strategic initiatives at The Washington Post who oversees the paper’s usage of Heliograf, told me in an interview that the main goal of automating the news is to free up more reporters’ time for substantial stories and to increase circulation by creating local or superspecific stories that can’t be covered by humans, because it would be too expensive to do that. For example, it would be impossible for any newspaper to cover the roughly five hundred legislative races across the country in any sort of detail, he explained. Before Heliograf, the Post was only able to cover elections in the most important districts. But today, it can offer readers stories on every contest around the country, reaching new readers who previously didn’t have access to detailed reports about the races in their own districts. Plus, each article is constantly updated as the vote count trickles in. In other words, as Gilbert said, the news constantly rewrites itself.
According to Gilbert, who was hired by the Post shortly after Bezos had purchased the paper, “the purpose of Heliograf is pretty straightforward: we want to take rogue and mundane tasks from our reporters and allow them to focus on much more interesting and sophisticated stories. For example, for the 2012 elections, we had human reporters and editors who wrote about 15 percent of all the stories related to congressional races, most of which we had no more information than was available through the Associated Press’s vote totals. So in that case, that’s not a good use of human time to read through a data stream and write fairly similar stories every time. So instead, what we did in 2016 was to build a system that could write those fairly template-driven stories and instead allow our humans to work on much more interesting stories for our audiences that still give us scale and reach.”
How did they do this? It was really very simple. Before the 2016 elections, journalists wrote a number of templates, which are basically prototypes of articles that describe various possible outcomes, along with several analytical paragraphs giving the voting history of each specific district and detailing how each particular outcome would influence the general election. On the night of the election, an editor fed Heliograf with the results of each contest as they came in via the Associated Press, inserted them into the templates, selecting the analytical and historical context based on who was ahead, and updating—or, in some cases, changing—the story as new information flowed in. If the Democratic candidate won, the computer automatically added his biography, his stance on various political topics, and the impact that election would have on the
balance of power in Congress. If the Republican candidate won, the computer would do the same thing with the corresponding information.
But were these stories written by the editor or by the computer? I asked. “By the computer,” Gilbert replied. “But at the end of the day, it was Washington Post editors who were creating the different types of narratives and structures that the computer was combining to make up the story. So ultimately it’s all human powered, but the computer is determining how to assemble the story based on the different options that are available to it and as it reads the data coming in. So on election night, when we say what happened in California’s 49th congressional district, that was all determined by the computer, but the options the computer had were originally coming from journalists.”
“FOR 2020, WE’RE GOING TO BE TELLING MUCH MORE SOPHISTICATED AUTOMATED STORIES”
For the 2018 midterm elections, The Washington Post was planning on using Heliograf to not only to write every article about the outcome of the roughly 500 local contests, as it had done in 2016, but also to write stories about the campaign contributions to each congressional candidate. For instance, Heliograf was scheduled to write stories on how much candidates and incumbents had received in donations from powerful lobbying groups such as the National Rifle Association. And Heliograf was programmed to turn out more articles based on opinion polls focusing on specific issues that could affect the elections or the fate of bills in Congress.
“Just to give you a sense, we have about sixty political writers, and on [the 2016] election night we probably had eighty or ninety people, including video coverage and the live blog,” Gilbert explained. “For the most part, what we try to do with our human reporters was not have them do simple, tallying-type stories. They were free to look at broader trends, to focus on individuals, to do deeper reporting. We just didn’t have them write stories that said, ‘Here’s who won in which district.’ For 2020, we’re going to be telling much more sophisticated automated stories. I absolutely believe that. I absolutely believe that it will allow us to focus even more of our human journalists on telling really unique, really human types of stories.”
Gilbert isn’t a technocrat spitting out the supposed benefits of Heliograf from the company’s public relations manual. He seemed to be sincere. His background is in journalism, not technology. He had started his career as a reporter and art director at The News-Press in Fort Myers, Florida, before moving on to become a professor of journalism at Northwestern University, teaching artificial intelligence and machine-generated content, and to serve as a consultant for newspapers that were trying to adapt to digital journalism. After a few years of teaching and consulting, he was hired to be the National Geographic’s web page deputy editor, before settling down at The Washington Post in 2014. He considered himself primarily a journalist, or at least an advocate of good journalism. He told me, “My goal is to put as many reporters out in the field as possible. People interviewing people, writing analysis, helping the reader understand not just what happened but why it happened. That’s what we should do with our humans. The machine is very good at telling you what happened, but it’s not as great at understanding why that thing happened.”
COMPUTERS WILL COVER FOOTBALL GAMES
When I interviewed him in 2017, Gilbert was about to launch his latest automated journalism project: articles generated by Heliograf about football and baseball games from hundreds of high schools in and around the Washington, D.C., area. “Right now, our reporters cover probably five or six high school games every Friday night,” he explained. “Those are the most interesting games, which is great, but it doesn’t allow us to cover all the other teams in schools in our area that might want coverage. Now, we could hire or relocate a bunch of journalists to cover those games, but instead, we can create a template-driven and algorithm-data-feed-driven approach with Heliograf that allows us to cover hundreds of high school games in a single night fairly instantaneously. That way, if your child goes to that school, and we don’t happen to be covering that game with a human reporter, we can give them some type of story. So in that case, it’s very much about scale.”
The mechanics of automating high school football results would be similar to those used for local elections, Gilbert said. The scores and stats will be sent in to Heliograf by coaches from each team, and a newspaper editor will verify the data before it gets published to prevent anyone from cheating. Then in a matter of seconds, Heliograf will churn out a story on each game.
Gilbert is also preparing to publish automated Heliograf articles for the Post’s book section. “We are going to create short little narratives to accompany the titles on our bestseller list,” he told me. “They will tell you what books are new to the list, who the new authors are, and we will be able to tell you how these books fare against historical data of other books we have in our database.”
ARTICLES WILL BE TAILORED TO INDIVIDUAL READERS
One of the most interesting—or maybe disturbing, depending on how you look at it—things that Gilbert told me was that the automated articles produced by The Washington Post, and eventually just about every other paper, will be tailored to each individual reader’s needs and interests. Using the data and preferences of everyone who subscribes to the paper’s digital edition, Heliograf will create articles taking into account not only the city and neighborhood in which each subscriber lives but also his or her knowledge of the subject. If someone has already read hundreds of articles about Russian president Vladimir Putin over the past year, Heliograf will assume that this person is already an expert on Russia and won’t bore him with information on the Russian leader’s background. But if, on the other hand, someone is reading an online story about Putin for the first time, Heliograf will automatically tell the reader that the Russian president is a former KGB officer and give him several other details about his life and work.
“I think Jeff Bezos has pushed us very hard to think about the consumer first,” Gilbert said. “In that light, there’s a lot we could do with personalization around locations or reading habits. Meaning, if you’ve read everything there is to know, or everything at least that the Post has published, about the situation in Syria and there’s a development today, a potential cease-fire, then we should tell you the thing that’s new as opposed to telling you who the main characters are, what nations are involved, because you already know all of that,” he explained. But if the reader hasn’t read anything about Syria other than that there is a potential cease-fire, Gilbert explained, then the paper will give the whole backstory leading to this latest development.
Does that mean that The Washington Post’s intelligent machines will track people’s reading habits and know exactly what each reader wants or needs to know? I asked him. “We’ll have some very educated guesses,” Gilbert said. “We’ll certainly know how much they used our coverage. We’re not tracking everything you read online, but if you’re a subscriber to the Post—if you read a lot of the online content—then we’ll know if you read any articles on Syria in the past week, or if you’ve read any of them at all.”
So how far are we from getting personalized news? I asked. “I think we’re talking a year or two,” Gilbert replied. “For example, we’re already starting to look at what kind of media each subscriber is consuming, whether you watch videos or not, so we can make videos a bigger or smaller part of the news you’re getting. It’s all about saving the reader’s time. If you’re not into watching videos, we’d much rather jump you into the story than force you to watch a video. If you’re never going to interact with our informational graphics, then we shouldn’t slow you down by putting that at the top of your story.”
I was tempted to ask Gilbert about the political risks of this new technology. If the technology for tailor-made news expands, as it certainly will, won’t criminal organizations, unscrupulous politicians, or governments use it to manipulate the news we get? Wasn’t the 2018 scandal o
ver the Cambridge Analytica technology firm’s mining of Facebook users’ data to compile psychological profiles only a prelude of what is coming? How long before news organizations and governments scour the entire Internet even more so than now, scooping up as much data as possible on what we’re reading and watching so they can pump us full of whatever information reinforces or disputes our beliefs? I didn’t ask because the answer was all too obvious: every new technological advance comes with potential dangers. Gilbert would surely tell me that if mankind had stopped developing nuclear energy for fear of atomic weapons, some of us would literally be in the dark. However, tailor-made news will have a huge political impact, one whose consequences still aren’t clear to anybody. It’s something that’s already happening—as we’ve seen with the Russian disinformation campaign that helped Trump win the 2016 elections—and it won’t be slowing down anytime soon.
WHAT THE NEWSROOMS OF THE FUTURE WILL LOOK LIKE
According to Gilbert, newsrooms will be much more automated five or ten years down the road, but it will be a gradual process. The most surprising thing for many veteran reporters will be the growing use of virtual reality headsets, the proliferation of augmented reality apps on smartphones, and the phasing out of computer keyboards. Much of our work will be done orally, thanks to advances in speech recognition software. According to Gilbert, instead of having keyboards, computers may simply have built-in speech recognition devices, and we’ll edit our articles by speaking to them. “Judging by the number of people I see just talking to their cell phones to dictate text messages and emails, and the number of people who feel quite comfortable using virtual assistants like Echo and Alexa, I think this technology is moving much faster than, for example, virtual reality,” he said. “I’d be surprised if most cars don’t allow you to talk with them, or if most homes don’t have some sort of voice-activated device, in the next two or three years.” And the very same thing will happen with the machines used by many journalists, he added.
The Robots Are Coming! Page 9