Sign up for The Media Today, CJR’s daily newsletter.
Mentions of automated news in the current climate likely conjure images of tech firms racing to develop AI tools capable of producing news stories comparable to the work of a journalist.
Google began pitching one such tool—dubbed “Genesis”— to media organizations last summer, a presentation that some of the executives present described as “unsettling”, according to the New York Times.
According to Google, Genesis is able to overcome the pitfalls of generative AI. It is reportedly being tested by small publishers, but its capacity to handle the nuance and unpredictability of journalistic practice is unknown and will ultimately be judged when – or if – it is released.
In the meantime, newsrooms that engage with automated news—that is, computer-generated text for journalistic purposes—do so relying on a more basic utilization of the technology.
Filling the blanks in templates
The Natural Language Generation (NLG) used in much automated journalism has been around for decades. It involves algorithms fetching data online or in internal databases, then pasting it in blanks left on pre-written templates (some say it resembles the game Mad Libs). As such, it has proven to be particularly well suited to domains such as sport, finance and weather.
Automated news has progressively made its way into newsrooms, probably due to greater accessibility in coding languages and open datasets.
It is typically implemented in one of three ways:
- Internally via in-house development
- Externally via a third-party NLG provider, or
- Semi-externally by journalists using low-code or no-code third-party tools to draft their own automated news.
In August, Alessio Cornia and I published an article in Journalism – ‘The cultural capital you need to work with automated news: Not only “your beautiful piece of work” but also “patterns that emerge”’ – that drew upon 30 semi-structured interviews with media practitioners, technologists and executives, conducted between 2020 and 2021, to explore the skills and mindset needed to produce this kind of automated news.
“Structured journalism”
We found that newsworkers need to become acquainted with a thought process reminiscent of ‘structure journalism’, or the process of “atomizing the news”. This process requires them to envision an ideal story, then break it down into smaller components fit for automation.
“If you didn’t have any automation involved, what would be the story you, as a human being, would want to write, or what would be the elements of the story that you would want to write?” asked a BBC editor.
“Having established that, we then looked at what data we could get to fill that.”
In the runup to the UK’s 2019 general election, the BBC had to pre-identify as many variables, or scenarios, as possible: candidates who could win or lose a constituency, comparisons of victory margins against the previous ballot, identifying candidates that could qualify to have their deposits reimbursed, etc.
An executive at the Associated Press outlined a similar process of “working backwards” when engaging with automated news.
Describing their approach to reverse-engineering automated stories, they said, “What does the story need to look like? And what are all of the possibilities? You know, earnings go up, earnings go down, earnings stay flat. There’s all of the branches that follow depending on the data that you’re using.”
Templatizing
This mindset was also exhibited at the Bavarian public service broadcaster, Bayerischer Rundfunk.
“We usually start with an ideal article, […] for example a perfect basketball match report, and then we kind of try to templatize it—make it into a template—and find out what’s possible and what’s not possible,” said a senior technologist.
However, not everyone was able to comprehend the level of abstraction necessary to hone this skill.
A senior technologist from the BBC noted that some in NLG familiarization training were unable to seize all the possible permutations to a story.
“That seemed to be the distinguishing thing: some people were kind of in tune with that complex tree of a story and other people were much more reluctant to do that,”
they said.
This was not exclusive to the BBC. Describing training held at the German newspaper Stuttgarter Zeitung, one editor said some participants struggle to “think like a computer would do, like a program would do”.
This has implications for hiring practices, we found. Some interviewees said that when recruiting journalists to work with automated news—or computational journalism projects more broadly—their organizations seek out talent proficient in this type of abstraction.
Aside from having to adopt a structured journalism mindset, media practitioners were presented with another challenge: that of translating journalistic norms and rules into computer code for automated news.
Norms and rules
According to a BBC manager, this required these principles to be articulated even more precisely.
Programming the algorithm for the corporation’s automated news coverage of the 2019 general election required determinations about the magnitude of an electoral victory, for example.
“If it’s by 50 votes, then you might call that [a] very narrow win or whatever; if it’s by 50% of the votes, you might call that an enormous win,” said the manager. “You’ve got to figure out where the boundaries are for the words that you use.”
Elsewhere, journalists’ expertise was also mobilized to anticipate editorial issues that could come up during an election.
An executive at the Washington Post mentioned the possibility of “really odd outcomes”, such as if an election goes to a runoff or if no winner is declared on election night. “We needed a ton of extra help from reporters and editors to essentially figure out what those edge cases were and then how we would like to handle them using Post’s style.”
Editorial questioning was sometimes brought by the utilization of automated news.
Two BBC staffers held somewhat different views on the handling of pre-gathered quotes in automated news.
The senior technologist said that interviewees could only ever have “hypothetically responded” to a situation that has yet to occur.
By contrast, a computational journalist argued that such quotes were editorially valid as long as the logic is explained to participants.
“It’s just making sure that you’re not going to misquote the person by putting it in the wrong scenario or context. So basically that you are going to do what you’ve told them you’re going to do with that quote, or why that quote is relevant.”
Unique disposition
In essence, our study argues that a unique disposition characterizes work with automated news.
It involves seeing journalism as both a unique endeavor—a “beautiful piece of work that’s completely owned by you and [has] nothing to do with anybody else,” as the BBC editor put it—and a procedure that can be deconstructed in an abstract way, just like computer programming.
Drawing on Bourdieu’s notion of “cultural capital”, which addresses unique abilities proper to a socio-professional sphere, we qualify this disposition as distinct-abstract capital.
Given the more prominent role taken in newsrooms by computational journalism (automated news, data journalism, algorithmic accountability reporting, etc.), this capital may give relevantly skilled practitioners a competitive edge, we argue.
But for those who don’t, there remains a risk of being sidelined.
With generative AI’s rise to prominence, new, but related, questions emerge about whether this type of capital can be found in other areas barely even considered a few short years ago.
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 765140.