While there are a number of valid concerns about the impact of AI, its usefulness in the news industry cannot be ignored. Some use cases for the techology are more commonly known, but there are some yet to be discovered — and experiements to find new applications for AI are happening in newsrooms right now.

During the recent INMA Smart Data Master Class, media leaders from The Guardian, Ekstra Bladet, Schibsted, and Bonnier shared insights into how their companies are using AI to improve coverage, products, and workflows.

The Guardian uses machine learning to power annotations

At The Guardian, lead data scientist Anna Vissens and her team are using machine learning to speed up the process of annotation, making the company’s journalists and editors more efficient.

For the project, the team decided to focus on quotes in the outlet’s articles. Vissens stressed the importance of what might seem like the tedious task of answering basic questions like, “What is a quote?” Even when the team felt that their definition of a quote and their list of exclusions was good, she said they realised that the way editors use quotes in their articles might be very different.

“We counted at least 15 different styles of quotes in The Guardian’s content,” she said, adding that her team knew they “needed to show the model enough examples of each style to teach it to properly identify quotes in our content.”

The next step was to develop the annotation workflow. They worked with three entities — content, source, and cue. The annotation tool they chose is called Prodigy (built by Explosion AI), and Vissens said the “interface is super flexible and easily customisable.”

Her team’s customisations included a flag in the upper righthand corner to indicate a difficult example that should be revisited later, as well as a free field at the bottom of the screen for annotators to leave feedback or notes.

It was interesting, Vissens said, to see “how the model was improving over time,” but that it was perhaps even more fascinating to see that “we were experiencing the same learning process ourselves, as humans.” She said that they found that “the more examples we looked at, the better we became at annotating.”

Ekstra Bladet Maximises the live chat experience

Tore G.C. Rich, content automation manager for Ekstra Bladet in Denmark, shared how the company is building customised tools for its journalists to make better use of generative AI and create new possibilities. One of the most popular tools is its live chat feature.

“We have journalists and experts that provide real-time updates on unfolding events and at the same time the readers can submit questions for the journalists to answer,” Rich said, adding that it is particularly useful for live sports, courthouse coverage, and updates from events like concerts.

Some of the biggest challenges of live chat come from its popularity.

“At the beginning of the live chat, we provide a summary for new readers to read so they can get up to speed with what has happened so far,” Rich said. “So the journalists have to update this summary continuously throughout the live chat process.”

After the event has finished, the journalist writes an article to summarise the key points, but Rich said there are often several hundred questions asked during a chat and it can be difficult to decide which ones to focus on for the summary.

Once a summary is written for the live chat, journalists use AI to generate suggested headings that it groups by themes. Journalists are asked to double-check the AI’s work and note whether it is accurate or not. They also give feedback on how good the generated text is, Rich said: “We want to have good or negative feedback to train the prompts and make the tool better.”

Bonnier debuts aggregated news initiative

With the creation of the news product named +Allt (which translates to “+Everything”), Bonnier News in Sweden found a way to offer subscribers access to its more than 50 news brands using a single subscription. Although +Allt was just launched in February, it boasts more than 100,000 paying customers.

But +Allt wasn’t on the market long before the team “started to see some problems” with it, Hans Hjelm, product owner for data science at Bonnier, said. Two of the common reasons for churn were ones the team knew it could resolve: People said they were not satisfied with the content they were getting, and people didn’t know where to find the content locked behind the paywall.

“This led us to create what we call the aggregated news initiative,” Hjelm said, adding it was launched about three months after the launch of +Allt. The first thing the team did was produce a newsletter.

“We have a bunch of articles and for each article in the newsletter, we produce an AI-generated summary using ChatGPT-4,” Hjelm said. “At the beginning of each newsletter, we create a summary of all the articles that are part of the newsletter, using an editor.”

In this case, the editor’s role is to fact-check stories and make sure the language flows properly. An explainer box tells readers they’re getting the newsletter because they’re a +Allt customer, and it also tells subscribers that the content is partly generated by generative AI with human fact-checking.

One important part of this process was to involve editors and ask for feedback about the AI-generated summaries they were receiving. That feedback was used to update the prompts being used to generate the summaries and has improved what the AI provides.

Schibsted experiments in connection

At Schibsted in Norway, the Futures Lab is filled with physical and digital experiments aimed at leading the news organisation to discover new and useful AI use cases to bring their business into the future.

Yifan Hu, Futures Lab’s UX designer, and a team of seven were inspired by ChatGPT plugins that allow users to do things like ask for parenting advice. They also looked for inspiration from other apps, like a video interface that helps seniors who live alone stay connected with friends and family.

“While all these new technologies are racing ahead, we also keep an eye on our society,” Hu said. “Aging population, loneliness, and education gaps are also expanding. So taking all these signals and trends together, we are seeing strong signals of futures where we believe conversational user interfaces or CUIs will play a very important role in our daily lives and we want to learn more about it.”

Taking inspiration from Siri, Alexa, and even science fiction, they started experimenting with a virtual personal assistant. Hu drew up a storyboard to envision the life of an older woman maybe 10 years from now, and the team built two prototypes.

These types of experiments are important to bridging technology and people’s needs, now and in the future. When exploring a new idea, Schibsted has three criteria it uses when drafting research questions: feasibility, desirability and viability.

“Feasibility is very simple,” Hu said. “How much are the technologies? Can we build a prototype with them? What are some of the technical limitations in today’s context? When it comes to desirability, in terms of use interactions and experiences, what are the key principles and design patterns for creating a good CUI experience? And the last thing, viability. Basically, will the service of products be viable? Can we find good use cases for it within our organisation?”

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .