Inside a ‘cybernetic newsroom’: Reuters’ robots are great at research but not at writing articles
Reuters, the historic news agency that once delivered its reports by pigeon courier, is building what it is calling the world’s first ‘cybernetic newsroom’.
The 166-year-old organisation is a global pioneer in using Artificial Intelligence (AI) in news gathering and has created two major programmes; one that enables it to laser in on breaking news from clusters of comments on social media, and another to unearth business stories by using algorithms to comb its vast banks of financial data.
Yet Reuters has concluded that early 21st century automatons still have significant journalistic deficiencies. “Machines write bad stories,” says Reginald Chua, Reuters executive editor for editorial operations (data and innovation). So the cybernetic newsroom is developing into a hybrid operation, in which the robots do vast and increasing amounts of laborious data-sifting, while real people remain in charge of composing articles.
“Where humans bring incredible value – and it is very hard to replace this in the short run – is in judgment about what’s newsworthy, and how audiences are changing in what they are interested in,” says Chua, speaking from the Reuters offices in New York. “Humans can talk to people and understand greater context and nuance and, of course, they are storytellers and to a large extent the job of journalism is to tell stories and engage people.”
This realisation, which must come as a relief to the agency’s army of 2,500 journalists, doesn’t diminish his excitement at the agency’s embrace of automated journalism. “Machines can dig through a lot of data, language generation is improving every day and stories can be more finely tuned towards audiences,” says Chua, a news veteran who spent 16 years with the Wall Street Journal and is a former editor-in-chief of the South China Morning Post.
In niche areas of Reuters financial journalism the robot is already writing the whole story. “We do full automations, where we automate headlines and we automate very short stories,” says Chua. “We do that mostly in the interests of speed where there are certain things that no human can do; where we need to get a headline out in a second.” These computer-generated snap articles are usually in relation to economic data that needs to be relayed immediately to the Reuters client base. “We serve a financial audience and they expect those numbers to come to them in frankly less than a second, we are working on millisecond ranges there.”
Somewhat worryingly for the future of journalism as a career, Chua says that “in some cases, when it’s a very straightforward story, readers have preferred the machine-written story to the human-written story”. But generally that’s not the case. “We are not looking to determine between the two (machine-written and human-written), for us the issue is simply ‘How can we make a better story?’ What readers really want is information, context and insight. And the best way to do that, I think, is to marry machines and humans.”
At Reuters the robot operates for the most part in the role of researcher, feeding material to real life correspondents. It is an editorial assistant with specialist knowledge, reliability and no inclination to complain at being tasked with drudge work.
Using AI to break news
At the hub of the cybernetic newsroom is the Reuters News Tracer programme, which has been designed to sift social media and flag up indicators of breaking news, from natural disasters to terror attacks. Chua and his team claim that this early warning system has given Reuters journalists between eight minutes and an hour “head start” on other global news outlets in identifying major stories, depending on the type of incident.
It was Tracer’s constant monitoring of social media keywords that instantly alerted Reuters journalists to the September murder of Belgian city mayor Alfred Gadenne, whose throat was slashed in a cemetery where he worked as a caretaker. When eyewitnesses in Moscow posted on social media that fire had broken out in a building used by Russia’s foreign spy service, Tracer picked out the story. It also scooped human journalists on the murder by gunmen of a Pakistani diplomat outside his residence in Jalalabad, Afghanistan in November.
In the Western world, Twitter is where Tracer focuses most, he says. “If you are standing on a street corner and a bomb goes off, heaven forbid, then your first instinct is not to go on Facebook, it is more likely be to take a photo and put it on Twitter. Twitter is the place where news breaks.”
Tracer scores over Twitter’s integrated notification service TweetDeck because it monitors accounts which could not be predicted as sources of news. “You don’t know what you don’t know,” says Chua. “But Tracer uses enough language and understanding that if a number of people start talking about the same subject in the same place that becomes a cluster which is then evaluated against newsworthiness.”
Tracer’s algorithms are designed to measure “newsworthiness” and “credibility” by matching geographical clusters of references to an incident with mentions by verified accounts. It’s a difficult science when fake news is rampant and the Reuters robot has to contend with other bots designed to corrupt social media with propaganda and fabrication.
But waiting too long for official confirmation from verified accounts can defeat the object. “If you tune [Tracer] only for extremely credible events the problem is that it will filter out lots of things you might want to know about,” says Chua, who seeks an “optimum signal to noise ratio” for the algorithm.
Ready-made sentences
It is Reuters’s new Lynx Insight programme that has the potential to be the real game changer for automated journalism.
Robots may still struggle with compiling elegant and evocative prose but Lynx Insight has language generation software capable of creating chunks of usable copy that can be used to build out a story or even provide a news angle. These robot-created sentences are structured in pre-prepared templates with relevant statistical information harvested and inserted by computer algorithm. The journalist simply pastes them into an article, a valuable tool in a job where speed is often at a premium.
Programmed with questions chosen by journalists to generate stories, Lynx Insight mines a vast trove of data to deliver up alerts that reveal “patterns, outliers, interesting nuggets of information” about any company that is being reported on, says Chua.
For example, in relation to Johnson & Johnson, Lynx Insight highlights that in the three months to 22 February, company insiders sold $20.17m worth of shares (excluding dispositions of indirectly held shares). It also records, in a ready-to-use sentence that “Among the 23 analysts that cover Johnson & Johnson, the breakdown of recommendations is 12 ‘strong buy’ or ‘buy’, eight ‘hold’ and three ‘sell’ or ‘strong sell’.”
The journalist can take these ready-made sentences and “pick out a nugget worth dropping right into a story or use it as a tip to go off and investigate something”.
Reuters is currently using these computerised tips to support its journalists who cover the markets but Chua says that Lynx Insight can be extended to other areas that Reuters covers, notably in sports and political coverage (Reuters has a long history in political polling). “We have a ton of polling data,” says Chua. “And down the road who knows what data is going to become available? As a species we are being inundated with data of all kinds.”
In an article last month in The New European, Geoff Sutton, a former journalist who became a Microsoft executive, recalled sardonically how “I had a brilliant genius of a boss at Microsoft who believed that we would be able to replace all journalists and editors with algorithms.” The process of automation, Sutton noted, “is continuing”.
But Reuters has decided this can only go so far. “We’re placing a bet that the future of automation in the newsroom is less around using machines to write stories than in using machines to mine data, find insights and present them to journalists,” Chua has decided.
The Reuters “idea”, he says, is “to marry the advantages we have as a newsroom and as a technology developer to create a system that will help our journalists get better”.