The Washington Post has a big team of journalists covering the Rio Olympics.
Also covering the games for the paper: Robots.
The Post is using homegrown software to automatically produce hundreds of real-time news reports about the Olympics. Starting tomorrow morning, those items will appear, without human intervention, on the Post’s website, as well as in outside channels like its Twitter account.
The idea is to use artificial intelligence to quickly create simple but useful reports on scores, medal counts and other data-centric news bits — so that the Post’s human journalists can work on more interesting and complex work, says Jeremy Gilbert, who heads up new digital projects for the paper.
"We’re not trying to replace reporters," he said. "We’re trying to free them up."
Gilbert and Sam Han, the paper’s head of data science, have a team of three engineers working full-time on Heliograf, the Post’s AI software. A few more product analysts are spending about half of their time on the project, and four or five newsroom staffers are also spending time shaping the software.
Here’s a mockup of the kind of the thing Heliograf is going to produce. This example uses data from the 2012 Olympics, but we will replace it with current stuff once the games get going on Saturday.
This content isn’t going to blow your mind. But it gets the job done so someone else can do a different job. Which is the point.
The Post isn’t the first group to create machine-generated journalism. Narrative Science, which started out as a Northwestern University experiment that turned baseball box scores into stories, now does earnings reports previews and real estate listings; Automated Insights does similar work for clients including the Associated Press.
But Gilbert, a former Northwestern journalism professor who taught a class that helped create Narrative Science, says the Post has bigger ambitions than short, robot-written stories.
For starters, it will use Heliograf to augment its coverage of the November elections, where it will generate a series of stories for some 500 races, to be updated in real time as results come in.
In 2012, Gilbert said, some of the Post’s election coverage was up to 16 hours behind, as a handful of humans waded through returns.
The plan is to also use the same software to look for interesting data points, like trends in voting patterns across the country, that it can flag for human reporters to build upon.
And at some point, Gilbert says, the Post wants to be able to "inject" contributions from its AI into stories its flesh-and-blood journalists are creating. Ideally, readers won’t be able to tell who made what.