News media companies have been using Artificial Intelligence for years, particularly for things like personalising content and targeting advertising. But recent advances in computing power mean that it’s becoming more feasible for AI to actually produce content like images and articles, which is known as ‘generative AI’. A report by Elaine Sun and Sarah Schmidt
How can news organisations use generative AI responsibly? It is possible to take advantage of its time-saving capabilities while still maintaining journalistic standards for accuracy, transparency, and ownership, said Kimmy Bettinger, responsible data ecosystems lead for the World Economic Forum. She discussed the risks and benefits of this quickly emerging technology during INMA’s Product and Data for Media Summit.
The first of five modules also featured a discussion about top data trends by INMA Smart Data Initiative Lead Ariane Bernard; top product trends by INMA Product Initiative Lead Jodie Hopperton; and Niksa Gopcevic, digital strategist at Syria Media Group.
Artificial Intelligence in general can be defined as “a computer system that can learn on its own,” Bettinger said, and media publishers have mostly been using it for analysis and inference tasks like personalising ads and promoting editorial content. It is also commonly used to do back-end business intelligence work, like identifying target audiences. But as computers become more powerful, there has been a huge expansion in AI capabilities, allowing models to move from analytical to creative work.
There has been a lot of buzz around generative AI or ‘large language models’. Because they can look at an enormous amount of data, these models are able to learn context and meaning by tracking relationships in that data. AI can now generate text, convert text into images, and produce audio, sometimes with “superhuman” results, Bettinger said. Tons of funding is going into generative AI in Silicon Valley, and enthusiastic media coverage abounds.
Bettinger thinks it would be a mistake to dismiss generative AI as overhyped. It’s already being widely used in real-world settings. “We’re starting to see organic take-off among real users across industries,” including architecture, design, and gaming, Bettinger said. It’s even used by bloggers to produce content using a feature called “natural language text generation,” which Bettinger expects to play a larger and larger role in editorial content in the future.
Accuracy, transparency, and ownership
The possibilities for using generative AI in news media are far-reaching and publishers will need to put a lot of consideration and thought into using it responsibly. The complete picture is still emerging. Initially, Bettinger is starting to see publishers use it to create new visuals that might fill in for stock photos. In addition, content generation has the potential to do some of the journalistic tasks involved in producing summaries and breaking news. “But there are certainly risks involved,” she told summit attendees.
Accuracy will be the primary challenge, Bettinger said, particularly when trying to cover breaking news. To provide a simple example, she shared a task she tried using the OpenAI text generation platform, Playground. She entered the prompt ‘Summarise the outcomes of the World Economic Forum’s annual meeting’ twice and shared the two responses the tool generated. Both were inaccurate, and the resulting text showed different — and incorrect — dates, locations, and other details, likely because it relied on archived Web pages from past meetings, she said.
It’s also worth noting that generative AI is very cost-prohibitive to build, so most media companies that are considering it will likely be procuring it, not building it in-house. That means they will be responsible for looking deeply into a third-party’s product and considering things that won’t always be transparent: What data is the platform using? What’s it doing with the data? What’s the resulting output?
The Factual is a news-rating company and Web site that implements a scoring system measuring bias and credibility of articles. It addresses the issue of the readers’ trust – or rather the lack of it. Consumers discouraged by what they considered misinformation and slanted reports now have a guide by which to judge what to consume.
Arjun Moorthy and Ajoy Sojan, self-described news geeks who were among the frustrated audiences, started The Factual in 2016. It rates about 10000 articles a day from 1000 sites, with a score from 1 per cent to 100 per cent, with 75 per cent considered a high-quality story. “In the public’s eye, the No. 1 problem they see with the news is bias,” Moorthy told summit attendees. “It’s not that it’s outright false, it’s just slanted a certain way. You don’t feel like you’re getting the whole story.”
The Factual, which was bought by Yahoo in September, turned the readers’ wants (well-researched stories reported by writers who are knowledgeable on their beat, with little opinion, and from credible news media outlets) into algorithms that can measure each of the four categories. The Factual’s system will peel apart paragraphs, look at quotes, links and sources. Words that are ‘emotional’ figure into the algorithm that measures opinion. The Artificial Intelligence model builds a database of reporters (currently at 50000), what they cover, how much they write in their field, and the quality of their sourcing.
Moorthy also emphasised that The Factual’s system grades on expertise, not popularity, which is why not all of the stories that show up on The Factual Web site are from well-known national news outlets. A substantial number of stories with top scores are from smaller outlets that have reporters at the top of their field with superior quotes and sourcing. The Factual’s algorithm and AI do not favour one side of the political spectrum over the other, Moorthy said, because what is measured is the journalism and the writing. The stories with high scores come from all areas of politics.
The Factual has a number of products in addition to the Web site. It has the app, Chrome extension, and its newsletter. Moorthy said usage is high – and highly retained. The newsletter also is an example of how AI can serve as a guide to editors rather than replace them. Editors can override algorithms if they want to pick a different diagram or substitute elements and change what had been selected.
(Elaine Sun is a freelance writer/editor based in Memphis, Tennessee, US. Sarah Schmidt is an editor, writer, and content manager, based in Brooklyn, New York. This article was possible, courtesy INMA.)