Rise of robot reporters: when software writes the news
Just three minutes after an earthquake hit California on Monday, the Los Angeles Times broke the story on its website.
The short article seemed fairly ordinary. It covered all the major details – when the quake hit, its magnitude and how far it spread. The only sign of anything unusual was the final sentence: “This post was created by an algorithm written by the author.”
In other words, the article was put together by a robot.
Once readers realised that the story was computer-generated, it attracted a bit of attention. But quite a few machines already write the news. Forbes magazine uses a company named Narrative Science, based in Chicago, to report on corporate earnings. The same service partnered with ProPublica in January to create a news application concerned with equal access to education.
The earthquake algorithm, nicknamed Quakebot, is not even the only robot reporter used by the Los Angeles Times. The newspaper’s Mapping LA project uses bots to compare neighborhoods, and its website, The Homicide Report, automates posts about murders in the city.
While robot reporters are not yet able to file fascinating 2000-word features for New Scientist, early research suggests that they are not bad at all. A study published last month in Journalism Practice (doi.org/r2g) found that a small group of readers could not reliably discern whether a sports article was written by a human or a bot. Those assigned the automated article found it trustworthy and informative, albeit a bit boring.
“Would this replace real reporters? I would say no,” says Christer Clerwall of Karlstad University in Sweden, who led the study. He believes that robots will generally be stuck with the mundane types of stories that they already do, leaving more complex tasks like narrative and eyewitness reporting to the human journalists.
Robot reporters have plenty of benefits, says Quakebot creator Ken Schwenke.
“We had it up and we had it first and we had the information out for people,” he says. “If we can automate it, why not?”
Just like famous statistician Nate Silver, who correctly predicted the last US election, robot journalists are all about data. They write stories by crunching spreadsheets full of sports scores, sales numbers, or stock market fluctuations. When it comes to pure computing power, bots will beat humans every time.
In Quakebot’s case, the process is a bit like the parlour game Mad Libs, in which players fill in gaps in ready-made sentences. When the US Geological Survey sends an email alert about an earthquake of significant magnitude, the information zips over to Schwenke’s web server. His bot parses through the email for the data, slots it into a prewritten template, and uploads the article to the newspaper’s content management system. It even sends an email reminder for the editors to look it over.
Other approaches are a little more complex. Automated Insights, based in North Carolina, builds robot reporters that scour data for interesting trends. They focus on personalised stories that are only of interest to a small audience – like recaps of fantasy football games on Yahoo, or summaries of recent web statistics.
In 2013, the company churned out 300 million pieces of content. Most journalists, says CEO Robbie Allen, want to write one article that will be read by lots of people. Automated Insight’s goal is to do the opposite.
“We’ll create a million pieces of content that we hope a million people read just one of,” he says.
And as the number of sensors in the world grows – from lifelogging to environmental trackers – so too will the number of niche stories that robot reporters could potentially write about.
Now that robots can handle the basics of reporting, could editing be the next frontier? At a panel on automated storytelling at Columbia Journalism School’s Tow Center for Digital Journalism last month, Narrative Science co-founder Larry Birnbaum speculated on a system that could exercise editorial judgment. The bot would decide which stories were worth writing, how the stories should be written, and which readers to show them to.
“I would love to give readers and editors and algorithmic story generators an interface like this that says, you know, ‘I understand there’s a trade-off between brevity and content, or between timeliness and analysis,’” says Birnbaum. “This is just a gleaming in our eye, really, but this is something I’d love to build.”