How does one lead newsrooms in the age of generative AI? 

The European Broadcasting Union (EBU), an alliance of public-service media that counts 113 members in 56 countries, examined answers to this question in a 75-page report.

There is a mismatch between how quickly tech companies are moving and how fast media organisations can determine the risks and opportunities associated with new AI models and tools. And there is still very little cost-benefit evaluation and impact measurement when media companies use AI, the EBU pointed out.

Here are some interesting use cases and learnings I came across in the report. Note that in each case, these news brands (Rappler, Omroep Zwart, Radio Télévision Suisse, BBC, Radio Sweden, and Bayerischer Rundfunk) are trying to use AI to solve a genuine problem — and the motivation behind most of these experiments is that they are trying to serve their audiences more effectively.

Still, the role of humans remains incredibly important.

(And if you are looking for a wider lens on GenAI usage in the media industry, please browse through our archives as well as through the three reports I have written, which are free for INMA members.)

Sonali

GenAI helps news companies respond to their audiences

In the Philippines, investigative news outlet Rappler wanted to hear from a wider, more diverse, range of its audience. But, as a small organisation, found it difficult to scale its focus groups. So, it tested a GenAI-supported solution.

Rappler created a virtual focus-group discussion. An AI system acts as a moderator in this group, asking an initial set of questions. Then it synthesises the participants’ responses — which could be in text and audio — and asks follow-up questions. Finally, it generates summaries of what was discussed.

Rappler found the tool generated many more insights than regular survey answers and was capable of picking up local languages reasonably well. Still, feedback showed more participants found the human-moderated consultations more engaging, meaningful, and trustworthy.  

Dutch public broadcaster Omroep Zwart also struggled with serving and representing viewpoints from its diverse audience, finding focus groups are difficult to organise and expensive to run. Its team created digital twins, which are virtual representations of diverse audience segments. The project used AI-driven digital personas to identify and integrate missing perspectives into the creative process.

“Early findings show that AI-generated feedback can help content makers identify harmful representations, improve inclusivity, and better understand audience sensitivities. The tool fostered more awareness of underrepresented perspectives during ideation and scripting stages,” the report said. 

Some key learnings: Diverse data sources (demographic, behavioural, social media) are essential for creating realistic and representative digital twins that capture a wide range of perspectives, while specific audience data enhances the relevance and usability of feedback from digital twins.

The team at Radio Télévision Suisse wanted to ensure it was meeting its audience’s needs when coming up with different angles to a story. So, RTS developed an AI model trained to operate on a diverse dataset of text, audio, and video while incorporating its journalists’ requirements and the editorial charter. 

This model categorises each piece of content according to the specific audience needs it fulfills, such as updating, diverting, inspiring, connecting, or helping. It automatically analyses incoming content and identifies the primary user need it addresses.

RTS journalists are keen to analyse their content offerings and identify coverage gaps or overproduction in content for meeting specific user needs. “The tool has had a significant impact on the understanding and acceptance of the user needs model within the organisation,” the report said. 

“Tools demonstrate value better than theory: Seeing the framework in action fostered genuine adoption among journalists.

“Workflow integration reduces resistance: By embedding the user needs model directly into existing content analysis processes, journalists naturally incorporate these concepts into their daily work rather than seeing them as a burden.”

The British Broadcasting Corporation realised its audience wanted more live coverage of football matches. Currently, editorial teams listen to and manually transcribe BBC Audio commentaries to feed into live pages that are updated with the latest action in matches through text and stills.  

So, the BBC piloted a tool that transcribes commentaries and delivers quotes and generative summaries based on those transcripts. Its editors look these over before publication.

“While the system-generated outputs contained some errors, with some player names proving particularly challenging, they were broadly accurate and compelling before being subjected to the editing process,” the report said.

Date for the Calendar: Friday, May 23

The INMA World Congress of News Media is almost upon us (May 19-23 in New York). And with it, my GenAI seminar, which will feature insightful speakers on topics that matter to us. I hope to see you there.

GenAI: “It’s of paramount importance to include the newsroom from the start”

Radio Sweden produces and publishes 370 audio news clips every day. Each clip needs a headline, a summary in three bullet points, and an alt-text (for image description). So, SR built a tool that creates these texts, at a high standard of accuracy, from audio transcripts. An editor then looks over them before publication.

Still, the headlines and bullet points “are very standardised and lack the more elegant style added by a skilled editor,” the report points out, adding: “It’s of paramount importance to include the newsroom from the start — to get their acceptance for using AI in the sensitive editorial workflow.”

German Public broadcaster Bayerischer Rundfunk is experimenting with personalisation, including an interactive audio news brief that focuses on a certain geolocation, which the user can either type in or which could automatically be located. 

The listener can then customise their news brief — for example, asking for news within a 50-kilometre radius, no older than 24 hours, and an alert each time a new item arrives. The AI system then creates a podcast for that listener, who gets an alert and can listen to it, like any radio programme. 

“We have very good user feedback on this product because people are really interested in regional news, and they also want to customise and personalise the news for their needs,” said Uli Köppen, who heads the AI and Automation Lab at BR.

BR has also rolled out an AI writing assistant that helps journalists be more effective, for example, by creating versions of a story for different radio programmes, its Web site, and for TV. 

“And we are debating on how to use this tool. Like, is it just an assistant? What does ‘just an assistant’ really mean? How much text do we produce with it? And do people still have to work on those texts? We’re still saying yes. 

“There are cases where we do have direct automatisation. But in most cases, people should use it for getting ideas, supporting the creative process, for versioning and then reworking the version. We usually don’t want to replace any kind of decision-making process with AI. We carefully review results generated by generative AI because of hallucinations. We can’t afford to publish any mistakes.”

How should newsroom leaders best formulate an AI strategy? 

The EBU makes a strong point:

“They need to own two things: their journalism and their audiences. They must be driven by the desire to invest in, produce, and resurface engaging and original journalism, and the desire to connect with audiences and communities. These two needs should drive any strategy that is supposed to contribute to a healthy media ecosystem, no matter which latest technology is used as a means for that.”

This journalism and audience strategy must be complemented by a tech strategy and a talent strategy:

“While investing in tech talent is important, investing in journalism talent could become even more so in a future where originality of content and human skills of delivery and connection with audiences are key success factors. News organisations increasingly bet on journalists as personal brands that convey authenticity, competence, and, well, humanness  crucially important when automation, cloned voices and avatars are the alternative,” the EBU said.

News brands must invest in reporting — and in critical thinking. 

“Journalists of all trades and fields of expertise need to sharpen their abilities to question the output AI generates. While the old reporting advice ‘follow the money’ is still valid, it will become increasingly important to ‘follow the data’ to uncover which input has shaped AI-generated output and which parts are missing.” 

Worthwhile links

• GenAI trends: A very accessible summary of four AI trends in the media industry.
• GenAI and search: AI Overviews correlated with a 34.5% lower average click-through rate for the top-ranking page compared to similar keywords without an AI Overview.
• GenAI and search II: Google’s AI Overviews now appear in search results for 1.5 billion users per month.
• GenAI and search III: Looks like Google’s AI Overviews are being expanded to YouTube as well.
• GenAI and agents: Visa is now embedding its payments system into chatbots and agents, letting  companies spend less on search advertising.
• GenAI and transparency: When do you disclose AI use to your audience? Some thoughts from a researcher.
• GenAI and transparency II: Another perspective on this from INMA member and Schibsted AI lab leader Martin Schori here.
• GenAI and editors: Wikipedia will use AI to support its editors.
• GenAI and lawsuits: Ziff Davis, which publishes PC Mag and Mashable, is suing OpenAI.
• GenAI and audio: Argentina’s Clarin is using voice cloning for its Opinion section and also features a curated audio news playlist.
• GenAI and comments: Researchers ran a secret experiment with AI writing comments on Reddit, sparking a backlash and possibly a lawsuit.
• GenAI and music: AI-generated music now accounts for 18% of all tracks uploaded to Deezer, signaling major shifts in the industry.
• GenAI and prompts: Prompt engineering is no longer a hot job, as AI models get better at figuring out what users are asking for.
• An AI diversion: Google Gemini falls for an April Fool’s Day joke.

About this newsletter

Today’s newsletter is written by Sonali Verma, based in Toronto, and lead for the INMA Generative AI Initiative. Sonali will share research, case studies, and thought leadership on the topic of generative AI and how it relates to all areas of news media.

This newsletter is a public face of the Generative AI Initiative by INMA, outlined here. E-mail Sonali at sonali.verma@inma.org or connect with her on INMA’s Slack channel with thoughts, suggestions, and questions.

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