“The article will die, but storytelling will not”: Notes from the Nordic AI in Media Summit

COPENHAGEN, Denmark — Last week, reporters, newsroom leaders, product managers, and AI strategists gathered in the former printing press of Politiken, Denmark’s largest newspaper. The setting was apt for the third annual Nordic AI in Media Summit (NAMS), which brought together journalists from Norway, Sweden, Finland, and across Europe, to Copenhagen for a two-day conference.

Far from mourning the days of print newspapers, the room was filled with technologists and futurists embracing AI technologies and calling for major changes to the way digital news publishers do business. Asked if generative search and AI transformation would kill the article page, Gard Steiro, the editor-in-chief and CEO of VG, the most-read online newspaper in Norway, said, “I think we should’ve killed it several years ago.”

As Steiro framed it, along with several other speakers across the summit talks, incremental changes to the article format and news distribution will not be enough for news media to survive the rise of AI technologies. Steiro encouraged others in the room to start implementing radical changes now, for a new generation of readers — “The article will die, should die, but storytelling will not.”

Provocations like this were par for the course at NAMS, with speakers like Aimee Rinehart, the Associated Press’ senior product manager for AI strategy, saying that Nordic newsrooms are “two to three years ahead” of U.S. publishers in terms of their adoption and adaptation to AI.

Across both days, representatives from some of the largest publishers in the region presented AI use cases from their respective newsrooms, and keynote speakers and AI strategists parsed buzzy topics like AI agents, “liquid content,” and European tech sovereignty. Here are a few of the highlights.

How to build your own model

TV news broadcasters in Sweden are training their own speech recognition models to rival OpenAI. The project was laid out by Mikaela Åstrand, a machine learning engineer at SVT, Sweden’s national public television broadcaster, and Leonora Vesterbacka, senior data scientist at KBLab, a digital humanities arm of the National Library of Sweden.

SVT has been experimenting with fully automated live captioning for its broadcasts for some time, but the system was plagued by chronic errors and viewer complaints. Originally, SVT explored using OpenAI’s Whisper, a multilingual speech-to-text model released in 2022, for its AI captioning. But less than 0.5% of Whisper’s training data — just 2,119 hours — is in Swedish.

In 2023, SVT and KBLab teamed up to develop an alternative speech recognition model primed for the Swedish language. SVT contributed its archives of millions of hours of audiovisual broadcasts, dating back decades, and KBLab successfully petitioned for access to 23,000 hours of recordings from the Swedish Parliament’s proceedings. They combined these datasets, among others, to fine tune a new model called KB-Whisper.

In certain benchmark tests, KB-Whisper is as much as 47% more accurate in its outputs than the original American-built Whisper. The highest rate of improvement was for the “tiny” version of KB-Whisper, which requires far less compute and is easy to run on even a basic smartphone.

KB-Whisper is now being integrated into SVT’s broadcasts for automated captions, and lives on Hugging Face for any other developer or newsroom in Sweden to use.

Giving advice to others in the room who might be daunted by model training, Åstrand said to look outside your own newsroom for collaborators — “You maybe don’t have the computational resources you need, you maybe don’t have the expertise or experience, but there might be others that do and you have the data.”

Can AI develop news sense?

Knowledge sharing and collaboration were themes throughout the summit, including in presentations about AI tools for story discovery. One spotlight was on the tool DJINN, developed by the AI lab at iTromsø, a newspaper serving the city in northern Norway.

The team behind DJINN built scrapers and APIs that pull data in from Tromsø’s municipal archives. DJINN then uses natural language processing to evaluate thousands of PDFs each month, rank each document’s relevance, and generate short summaries for reporters. One key to the tool’s effectiveness has been the use of supervised machine learning, in which journalists are asked to give a thumbs-up or thumbs-down on any filing flagged by the tool. The prompt is simply: “Is this a news story or not?” These rankings continually refine the tool and train it to spot newsworthy filings more effectively.

Rune Ytreberg, head of the AI lab at iTromsø, says the training is an attempt to mimic the “gut feeling,” or news sense, of investigative reporters.

The tool has been the source of several major leads, but not only for stories published by iTromsø. Stavanger Aftenblad, another daily newspaper in Norway, licensed DJINN to use in its own newsroom, where it uncovered the story of an old hospital building that was illegally converted into an apartment complex.

DJINN has since been key to reporting a series by Stavanger Aftenblad on the Norwegian floating sauna industry, which has left a paper trail of permit violations and police complaints by neighbors as these saunas began popping up in protected fjords and bays across the country.

Elin Stueland, head of AI at Stavanger Aftenblad, leads the partnership with iTromsø. She said one of the fastest ways to get even the most AI-critical journalists to consider adoption is through discovery tools. “You can make journalists happy by helping them win the news race,” she said.

Europe’s big platform question

From comments made on-stage and off, it’s clear the instinct of Nordic newsrooms to collaborate and explore homegrown tools is, in part, a response to fears of overreliance on American technology companies.

Offhand references were made to the Trump administration throughout the summit, and to the fact that CEOs of major AI companies like OpenAI, Microsoft, and Meta have been cozying up to the new administration. In Denmark, in particular, geopolitical tensions with the U.S. have been simmering ever since Trump declared his intent to occupy Greenland, which is an autonomous territory in the Kingdom of Denmark.

Rasmus Kleis Nielsen, the former director of the Reuters Institute for the Study of Journalism, and professor at the University of Copenhagen, attempted to take these questions head on in his concluding keynote of the conference.

Nielsen outlined the long road ahead for building viable, large, European-owned technology platforms and spotlighted in practice how resource-intensive these projects would be, especially if the public sector gets involved in their development. He also questioned whether users would actually migrate to alternative “smaller” European platforms, in place of large ones, using the dwindling user figures on Mastodon as an example.

“This is a lot more complicated than ditching your Tesla and buying a Renault,” said Nielsen. Far from offering solutions, Nielsen instead sketched out the textures of the “wicked problem” of European tech sovereignty, which he compared to the climate crisis, welfare collapse, and regional defense.

While pointing to the considerable cost of building alternatives, he offered up a series of questions for the audience, “What do you want this more than? What are you willing to do? What are you willing to pay? What will you sacrifice for this alternative?”