Nordic AI in Media Summit 2025: Five takeaways from this annual event on the future of news

Journalists, academics and media executives gathered this week to discuss AI in a room formerly used to house a printing press. It was the third edition of the Nordic AI in Media Summit, hosted in Copenhagen by the Nordic AI Journalism Network. The initiative is led by Kasper Lindskow and Sara Inkeri Vardar from JP/Politikens Media Group, Agnes Stenbom from Schibsted and Olle Zachrison, until now head of AI and news strategy at Swedish Radio and soon to join the BBC as senior news editor for AI. 

The summit included keynote lectures by experts such as Ezra Eeman from NPO, Nikita Roy from Newsroom Robots Labs, Naja Nielsen, outgoing digital director at BBC News, and our former director, Rasmus Kleis Nielsen, now at the University of Copenhagen. There were also presentations of AI projects and tools from many organisations, and talks targeting specific issues, including how to incorporate AI in newsroom workflows, as well as in strategies and values.

In this fast-evolving environment, the practical ways in which journalists are using AI are more relevant than ever. At the same time, the conference also hosted broad discussions on emerging issues related to AI. Here are five key takeaways from the discussions. You will be able to catch up with the conference programme in full here. 

1. Scanning and interrogating documents

Elin Stueland from Norwegian regional newspaper Stavanger Aftenblad and Rune Ytreberg from local tabloid iTromsø shared a tool that can help local journalists find stories from public documents without having to take the time to scan through all the information available manually. 

The tool, called ‘Djinn’, identifies potential stories and produces summaries in simple language for journalists to then verify and add context to. Djinn can be accessed through Slack, and journalists are training it to recognise ‘news judgement’ by evaluating its output through a thumbs-up or down button similar to the one found on Netflix.

Ytreberg said that Djinn cuts the daily time journalists spend searching public documents for stories from 2 hours to 10 minutes. “How do you make a journalist happy? Make the boring work go away,” Stueland said.

Erja Yläjärvi, senior editor-in-chief at Finnish newspaper Helsingin Sanomat, presented a similar initiative by her newsroom, ‘Watchdog’, a tool to find news and alert editors. This also scans official documents and generates summaries, with editors notified of potential news. Helsingin Sanomat has gone even further with “Document rake”, a tool that allows journalists to chat with large sets of documents in a way that mimics Notebook LLM. 

Tore Rich and Jonas Pröschold from JP/Politikens discussed some of the capabilities of the publisher’s multitasking API tool Magna, which has a similar capability in its ‘DocChat’ function.

“Writing is not a core skill for most journalists in the very near future,” said Yläjärvi. “These AI tools already show us we have to be thinking about what our core skills are as journalists. Within change management, these discussions come about very quickly.” 

2. Partnering with librarians to improve AI models in Swedish

Mikaela Åstrand and Leonora Vesterbacka presented their project to improve Swedish speech recognition in automated captions. Åstrand, an AI/ML engineer at Swedish broadcaster SVT, discussed how she had been noticing consistent errors in existing captioning models in Swedish. Despite being the most spoken Nordic language, Swedish is spoken by a comparatively small population, creating problems in collecting a large enough dataset to train AI models to recognise and transcribe speech accurately. 

SVT’s own data wasn’t enough, leading Åstrand to turn to the National Library of Sweden, which collects all domestic printed and audio-visual materials in Swedish. She teamed up with Vesterbacka, a senior data scientist at the library, to adapt and improve Openai’s 2022 Whisper model. 

The model was only trained on 2,000 hours of Swedish out of 680,000 total hours of audio, meaning it wasn’t very accurate in the language, particularly in the smaller version of the model. Åstrand and Vesterbacka fine-tuned it with 50,000 additional hours of Swedish from the National Library and other sources, including the Swedish Parliament, which is responsible for recording everything said in the chamber and providing accurate transcription. Their resulting model, KB-Whisper, managed to bring the word error rate down, with a 47% improvement in the large model.

3. A more comprehensive approach

While a large part of the conference was once again taken up with presentations of AI tools, projects and uses in the newsroom, an emerging theme this year was widespread adoption of AI by news organisations, and a growing focus on an extensive approach that penetrates every stage of the news production process, in contrast to individual projects tailored to specific use cases.

This shift was highlighted by Ezra Eeman, strategy and innovation director at Dutch public broadcaster NPO, in his opening keynote, where he also mentioned a move towards hybrid flows where human journalists and AI tools work together at every stage of news production.

This approach was shown by Gard Steiro, news editor at Norwegian tabloid VG, and Fabian Heckenberger, senior editor for AI at Süddeutsche Zeitung. 

Steiro presented VG’s 2028 strategy, with AI present in every step. Similarly, Heckenberger illustrated how generative AI can reshape every step of an institutional plan for change, including affecting mission aims and culture. However, the issue remains that comprehensive change requires time to think and plan, something that many journalists don’t have amid the daily news cycle. “How do we not overtax our newsroom? How do we steer through multiple waves of innovation?” Heckenberger asked.

4. Automating writing and AI agents

Tina Rogers, head of delivery for editorial AI and automation at digital sports media group Better Collective, spoke about how the publisher is experimenting and creating content with generative AI. 

Better Collective allows journalists to use AI to translate and rewrite articles on its content hub in the style of another one of its publications to allow for sharing content between its titles. They also use AI to produce original pieces focusing on production, particularly guides and data-heavy content. One example is batch-creating short articles announcing the results of minor-league games using data and LLMs. These can even be published automatically, Rogers shared.

Another experiment Better Collective is running is creating content with AI agents. 

Agents go beyond LLMs and even AI ‘characters’, AI expert David Caswell explained in a panel with Hugging Face press lead Florent Daudens and Irene Jay Liu, director of AI for emerging tech and regulation at the International Fund for Public Interest Media (IFPIM). AI agents are a combination of reasoning and action in a specific environment, Caswell said. They use reasoning to deconstruct a set of tasks aiming towards a goal, and are different from software in that they’re not as brittle and can address unpredictability and recognise when guardrails are violated. 

Better Collective started by applying AI agents to transfer news, which often begins with a rumour on social media. Specialised AI agents perform specific tasks like social media curator, researcher, sports writer, fact-checker, internal link builder and publisher, starting with a post on social media and ending with a published article. Rogers said Better Collective isn’t advertising its automated transfer site, which is publicly accessible on the web, but is using it as a playground to test these tools.

Daudens predicted that AI agents could be the next user interface and the death of the Internet as we know it. In this potential scenario, people wouldn’t browse websites themselves: their AI agents would do this on their behalf to achieve any task required. This might have deep implications for journalism, including the possibility of news having to be packaged for an audience of AI agents rather than directly for humans.

5. What will it take to build a European alternative to Big Tech?

In the final keynote address of the conference, our Senior Research Associate Rasmus Kleis Nielsen asked tough questions about what it would mean, in practice, to build European alternatives to the popular US-based platforms many people use. 

This has been a topic of conversation, especially in recent months, following concern in Europe after Twitter’s sale to Elon Musk, Meta’s shift in policy regarding fact-checking, and platform executives getting closer to Donald Trump.

If we want to move away from American platforms and towards European alternatives, we have to acknowledge the significant practical issues this entails, Nielsen said. He outlined three possible models for European alternatives, each aiming to solve for a different analysis of what the problem with the existing ones actually is: whether it’s that they’re American, or that they’re for-profit, or that they are big.

Each of these potential models, he said, brings its own issues: people seem not to want smaller alternatives (for example, Mastodon) and a large European platform would require a significant regulatory and financial commitment from governments that are already struggling with complex, expensive issues like climate change, defence and aging populations.

“Just as we think of issues about climate, defence, welfare, it’s good to want to discuss them, but it’s not enough. We want more than just a declaration; we want details. We particularly want details if we’re journalists or scientists. We want more than just rhetoric. We need to ask not just, ‘Do you want this?’ but ‘What do you want this more than? What are you willing to sacrifice, or to pay for, to have this?’” Nielsen said.

If you want to know more…

• Catch up with last year’s NAMS coverage here
• Read our interviews with NAMS cofounders Agnes Stenbom, Kasper Lindskow and Olle Zachrison
• We have published a few of David Caswell’s essays on AI here 
• Read our interview with Florent Daudens and review his NAMS presentation

Explore the Reuters Institute’s work on AI here, including research and more original pieces