Hello, everyone. How’s your year going so far? 

As we settle into 2025, we’re looking at two trends in today’s GenAI newsletter. One relates to the bigger-picture view that many news organisations are taking when it comes to GenAI applications. The other is a use case of multimodality, which promises to transform both journalistic workflows and the way audiences consume news. 

I hope they are both of interest.

Sonali

Big bets for 2025 with GenAI

After a couple of years of experimenting with various applications of GenAI, here’s an interesting trend we’re seeing: news media businesses creating end-to-end content solutions using GenAI across the content chain.

These are news organisations that have decided to focus on longer-term GenAI projects that can move the needle for them rather than simply seeking quick wins or flashy consumer-facing products.

They are not thinking about disconnected “point solutions” that might solve immediate problems — they are instead building systems that create capabilities for future solutions. 

“I am putting my bets on multimodal retrieval augmented generation (RAG),” said Lyn-Yi Chung, deputy chief editor at Mediacorp’s CNA Digital team and lead for its AI strategy and solutions team in the newsroom. “I don’t see video as just media but as text — and the same for audio and images. We’ve heard a lot about how content should be ‘liquid’ and easily transformed to suit a consumer’s preferences.”

A strong focus on RAG could take pressure off journalists, she said.

“Everything coming in should be meta-tagged, indexable, and searchable. It is the foundation for smarter copy testing and for true story-centric coverage. Imagine being able to quickly pull together resources from every corner across the newsroom — the digital team, television, radio and more,” she said.

The key challenge is overcoming friction between newsroom software and systems. “You can’t trigger RAG across these platforms if you don’t build out data flows and connectors,” Chung said.

“We’ve seen a rush of AI features being unveiled in newsroom software and services. I suggest vendors do a lot more listening this year and work on making their offerings more open and adaptable to AI plug-ins to sharpen their value for media companies.”

Media companies will increasingly be thinking about building a “Swiss army knife for AI features,” despite it being “a crazy ambitious project,” she said.

“It’s like we’re looking at parts of a Ferrari. It’s easy to get distracted by shiny little things. But here’s our chance to build something more powerful.”

This is very much in keeping with the advice that Uli Koppen, who runs the AI + Automation Lab at Bayerischer Rundfunk, shared with the INMA community in our last GenAI Master Class: Build “trojan unicorns,” strong AI use cases that have the solution to a bigger problem “in their belly” — and which help build data infrastructure that can be used over and over again — rather than simply building products.

Similarly, Alessandro Alviani, who leads GenAI at Süddeutsche Zeitung Digitale Medien, explains it this way: “Summaries is something we have been working on. They are now a commodity. I see summaries as a building block. What do you do with the summaries once you are able to summarise? For example, you can match summaries with audio or additional distribution formats.” 

Bauer Media, which publishes more than 200 titles in Germany, the United Kingdom, Poland, and France, is on a similar journey. After launching its five-year-long Imagine programme, which looked at the best ways to use AI tools at the company, the publisher is now embarking on its Embrace programme, aimed at getting its staff to wholeheartedly lean into AI usage.

Bauer has run a slew of experiments, looking at 71 different use cases, and the point of the programme is to bring together different use cases and scale the best ideas.

Embrace is entirely content focused, said Rob Aherne, who runs the programmes: “Embrace spans the whole content-creation process, from ideation to distribution. We could do seven AI projects. But if we’re going to put our chips on something, it is content, because that is the heart of the business.”

The programme is largely internally focused, although some components could end up powering consumer-facing products, such as text to speech projects, which Bauer has used to create podcasts out of archival content.

Did you catch our last GenAI webinar, presenting a playbook for media businesses at different stages of GenAI maturity? You can watch a replay here. 

GenAI goes multimodal at Politico

Here’s an interesting multimodal GenAI use case, this one from Axel Springer-owned Politico, in Washington, D.C.  

The political news site often needs to cover live events, such as debates, where many presenters are speaking in quick succession on a range of topics. Politico wants to provide its audience with a live blog so they can keep abreast of interesting developments, particularly on topics that interest them.

“One of the challenges we’ve always had to solve is: What do we do in the moment of a live event, a debate, or a speech?” said Andrew Briz, editorial director of newsroom engineering

But existing live chat tools provide a “disjointed” experience, particularly if a reader comes in after the event has started and is trying to keep up with the live blog while scrolling through what has already been said, Briz pointed out. He also wanted to give readers the ability to filter the conversation to focus on particular topics.

“Now, with AI, it is realistic to do a constantly updated, up-to-the-moment summary,” which can be categorised by topic so readers can quickly find information relevant to their interests — and so journalists can focus on analysing the content rather than simply transcribing or reporting it, he said.

The tool ingests an audio feed, which could be either pure speech or video, transcribes it, cuts it up, and identifies which speakers are speaking (editors need to give it that information once). The editors create four topics they think are important. The tool undertakes a retrieval augmented generation (RAG) search on those topics, finds the parts of the transcript that apply to that topic, and creates a summary.

The next step is to improve the tool so the quality of the content generated is strong and clear. At the moment, an editor needs to go over it carefully to ensure that there is no gibberish published. 

How does the audience feel about this product? Politico added “thumbs up” and “thumbs down” buttons and asked its readers: Is this helpful?

“My thought originally was that these were tied to summaries themselves, that people will ‘thumbs up’ good summaries and ‘thumbs down’ bad summaries, and then we’d be able to see the good ones from the bad ones without our having to go in and rank them all,” Briz said. 

“What happened instead was whatever was at the front of the line got the most engagement, which makes sense — most people saw those buttons and they didn’t necessarily click to the other summaries. But what was surprising was almost every summary had exactly a 50-50 thumbs-up to thumbs-down ratio, whether it was the first one, which had tons of engagement, or one of the later ones that had much less engagement. 

“My read on that is that they are reacting to the technology itself, not to the actual question: Was this helpful. I think we’re suffering a little bit here from a backlash of some people really don’t like this technology and they’re just going to hit that no button every time. I was expecting to see some sort of distribution and it’s non-existent.”

Worthwhile links 

• GenAI and slop: Quartz is producing AI-generated copy that is based on other AI-generated copy.
• GenAI and labour unrest: The Guardian may be using AI to help with copy during a strike.
• GenAI and copyright: Many Indian news publishers team up to sue OpenAI. This includes The Indian Express, Hindustan Times, NDTV, and the Digital News Publishers Association, which represents roughly 20 companies, including Mukesh Ambani Network18, Dainik Bhaskar, Zee News, India Today Group, and The Hindu.
• GenAI and copyright II: OpenAI asks for publishers to be stopped from joining the lawsuit.
• GenAI and copyright III: Also, oh, the irony: OpenAI says DeepSeek stole its intellectual property.
• GenAI and copyright IV: LinkedIn sued for training models on user-generated content.
• GenAI and copyright V: “We don’t want to create a situation where all the risk is on the content producers and all the profit is on Big Tech.” 
• GenAI and translations: A handy playbook put together by the Centro de Periodismo Investigativo for anyone who is thinking of using LLMs for translations.
• GenAI and customer service: Lyft cuts average customer service resolution time by 87%.
• An AI diversion: Deutsche Bank uses memes to tell its clients about AI trends.

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.

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