How computers are now writing business news stories
The next business news story you read may have been composed by a computer software program.
Major business news organizations such as the Associated Press and Bloomberg have either already begun using computers to write basic business news stories or are exploring ways to use computers to help them write stories.
Some of them are using Automated Insights, headquartered in Durham, North Carolina. The company has created a product called Wordsmith that allows news organizations to generate unlimited pieces of content from a single story structure and data set. Since its start in 2007, it has worked with many big-name clients including the AP, Yahoo and Samsung.
The stories that Wordsmith produces sound like a human wrote them, and depending on the tone of the story, can contain sarcasm, jokes, or even sympathy. Last year, Automated Insights produced over 1.5 billion pieces of content for its customers.
“I think that smart news organizations should find ways to use computers to increase their reach and lower their costs,” said Ryan Thornburg, an associate professor of journalism at UNC-Chapel Hill and data journalism expert. “That is going to come with some human costs just like it has come with the automotive industry and grocery store clerks where jobs are replaced.”
Automated Insights is primarily working in sports news and finance news. Despite the rapid success that the company has experienced, it does face several challenges, including journalists’ fears of losing jobs, continued imperfection due to human error, and increased competition.
Although Automated Insights is a key player in automated journalism, it isn’t the only one doing it. Narrative Science, headquartered in Chicago, started producing automated stories in 2010. In contrast to AI, Narrative Science is product heavy rather than service heavy and focuses more on helping smaller markets.
In addition to Narrative Science, news organizations across the country are incorporating automation into their structures.
For example, the Los Angeles Times has created Quakebot, a robot that takes earthquake data and produces hundreds of stories about minor earthquakes in the California area. This keeps the average citizen in tune with what is happening on the tectonic plates, but frees up journalists’ time to cover more important stories.
Bloomberg is also jumping on the automation train. On April 26, Bloomberg editor in chief John Micklethwait sent out an announcement stating that Bloomberg will be expanding its automation efforts. The company is forming a 10-strong team to lead the initiative in-house, and Bloomberg will focus on using automation to unveil interesting stories. The company will do so by exploring extensive data sets in various markets.
Micklethwait thinks that this will help Bloomberg reporters use their time in a more beneficial way. “The time spent laboriously trying to chase down facts can be spent trying to explain them,” said Micklethwait.
AI and the AP
Automated Insights’ biggest success story has been with the Associated Press. In 2014, the AP hired AI to produce automated earnings stories.
Before the partnership with AI, AP financial reporters were covering only 300 stories per quarter. Now, with the help of Automated Insights, AP publishes roughly 3,700 automated earnings stories per quarter.
Lisa Gibbs, AP business editor, wants her reporters to work on the highest value journalism that uses their creative sources and investigative abilities. Gibbs, a big supporter of Automated Insights, said, “By setting up a different kind of structure, we are able to free up time.”
Still, there is concern that computerized business news will diminish the journalists’ voice and result in lost jobs.
Robbie Allen, Automated Insights’ CEO, spoke at FutureWork in Raleigh on Feb. 8, and in response to this concern, he said, “I’m happy to report that the number of jobs that to our knowledge have been lost as result of an implantation of Wordsmith is still zero today.”
The CEO explained that Automated Insights is actually a net creator of jobs, and he claimed that AI knows several people who have been promoted after a successful implantation of Wordsmith.
Gibbs supported Allen’s claim and explained that automation certainly doesn’t have to eliminate jobs, it’s about how the organization chooses to use it.
“We sat down after fourth quarter earnings, and we did a bit of an audit,” said Gibbs. “We asked: ‘Were there little errors? What were they?’ ‘Are we noticing certain issues?’ You have to pay attention to it.”
Allen believes that although it is impossible to say no one will ever lose their job because of automated journalism, we should take a step back.
“The 20th century was all about automating repetitive physical tasks,” said Allen. “So if your job prior to that was doing something manually repetitive, that was prime for being automated…The 21st century is going to be all about automating repetitive intellectual tasks
There are still errors in the stories that Wordsmith produces. However, misspellings and grammatical errors are no longer an issue.
“We promised people that we would automate earnings reports,” said Lou Ferrera, former vice president of content verticals at the AP and now chief content officer at Bankrate.com. “I didn’t promise them it would be 100 percent perfect, but the difference is now I have a much lower percent of error.”
While the software follows its grammatical instructions 100 percent of the time, it is still ultimately under the control of the human, who is the one writing the algorithms that produce the stories. Therefore, errors involving the data can occur. If the human enters in the code wrong, then the robot does not have the ability to correct the mistake.
For example, in July 2015, Automated Insights produced an earnings story for the AP on Netflix. However, when Netflix released its second quarter earnings, the stock underwent a 7-1 split, and the data that Wordsmith collected did not reflect the switch. This caused Wordsmith to report that the individual share fell 71 percent and noted that the company had missed analyst expectations for per-share earnings.
Celeste Lecompte, a journalist for Nieman Reports, wrote a report called “Automation in the Newsroom” and she reflected on Wordsmith’s Netflix error. “Your data have to be bulletproof,” she said. “You need some form of editorial monitoring to catch outliers.”
Because of these possible mistakes, it is essential that humans have a strong role in automated journalism.
The human touch
Andy Bechtel, an associate professor at UNC-CH and expert in digital journalism and editing, is a critic of automated journalism. He understands the efficiency of it but believes the context is often missing, and that is a detrimental fault to the value of the automated stories.
“When you get to more complex stories, I still think a human touch is necessary,” said Bechtel. “The robot got the basics down OK, but it falls short with quotes.”
Although earnings stories are formulaic, the AP does adding context. At the FutureWork forum, Allen explained the way the AP inserts context into its earnings stories.
“In some cases we automated 80 percent of the story and enabled the journalists to focus on the top 20 percent, on the company analysis, the executive team, etc.,” said Allen.
Gibbs confirmed this editing structure. She said that for the top-tier companies, “Your Googles, Apples, and Walmarts,” her reporters still basically write their own story. Maybe they incorporate the earnings into their larger piece, or maybe they “write through” the automated story.
“Essentially with those companies we’re saying these are important enough and consumers care enough about them for us to provide a meaty look at their strategies, or products, or what have you,” said Gibbs.
The AP has about 100 of those top-tier companies.
In addition, it has a second tier of companies where it checks facts and adds context to earnings stories. There are 250-300 second tier companies.
In addition to adding context, Gibbs believes it is also key to advise the reader that the automated story was computer generated. The AP does this by adding a disclosure message to the end of every story.
Thornburg thinks that automation is a smart business strategy for companies to use. “If it’s replacing people with machines and the quality and scope of the stories are the same,” said Thornburg. “Then that’s also useful, because as advertising revenue goes down, news organizations have to find a way to do the same quality work they did by lowering the costs.”
While other news organizations are trying to incorporate automated journalism into their businesses, Automated Insights is actively seeking other markets to break into.
Allen explains that computers are well suited for quantitative analysis, so when looking for potential markets the best place to look is areas where there is clean and consistent data.
Recently, Automated Insights has announced that it is extending its partnership with the Associated Press and are going to be expanding into covering sports and other areas.
Gibbs said that the AP is looking at automating stories about state unemployment rates, and it could consider automating stories about local gas prices and real estate sales.
Thornburg sees potential for Automated Insights in politics. “I think politics is a good place to get involved,” he said. “Night and day messages fly under the radar, so I can see political campaigns wanting to send emails to personalize messages.”
Thornburg believes politics is a time-sensitive topic and that automation could be a key resource in getting stories that include data about polls, elections, etc., out fast.
On the contrary, Bechtel thinks that politics may be too tricky.
“For basic things like who won the primary and earnings reports of politics, I could see a role for that,” said Bechtel. “But when you get into the nuances of campaigns and politics, it may be difficult for automated journalism to see that.”
While these basic stories may be helpful for simple knowledge, Bechtel believes it is not helpful in the decision making process. “The removal of that context makes it harder for people to make decisions on voting, what to buy, how to run their lives,” said Bechtel. “You need contextual information not just broad numbers.”
When asked about expanding into politics, Gibbs said, “Definitely not because there are too many nuances.”
Allen believes that in five years automation will be a critical aspect to every news organization in the country. Allen thinks that in the next 50 years the future of automated journalism will consist of a much higher presence of humans plus software.
“Not the notion that there will be a black and white reality we are going to face where software comes in and takes jobs,” Allen said. “It is much more likely that humans plus software are going to make a better world than just having software or just having humans.”
Thornburg believes that smart news organizations should find ways to use computers to increase their reach and lower their costs.
Bechtel says he’s not afraid of automated journalism. However, he adds, “If it gets to be a point where somehow they can make it so a robot could write a 2000-word New Yorker article, then I’d be scared.”
Armstrong is a senior business journalism major at UNC-Chapel Hill