We’ve all been hearing about using AI for personalisation for years now. Today’s newsletter surfaces examples of two news publications in different corners of the world that are actually getting it done.

This is not your standard recommendation-engine stuff. Each of these use cases puts the news consumer in the driver’s seat rather than the editor or the algorithm.

In one case, the publication is producing different versions of the same content and letting the user choose how they would like to consume it. In the other, the news site offers the user a range of prompts as suggestions for questions they may want answered via chat but also accepts other questions — and then answers them based on its own content.

“We want readers to choose the best option” for content 

It’s not often that one meets an editor-in-chief who publicly declares that his aim is an AI-driven newsroom. These are not words that usually fit in the same sentence together. Data-driven, sure. But AI-driven? 

Meet Markus Knall, editor-in-chief at Munich-based Ippen Digital. He likes the idea of having different versions of the same article for different audiences. 

“I think versioning is one of the big strategic fields in AI for newsrooms,” Knall told me. “We have 50 different brands. In the perfect world, we would write the same story in 50 different tones — in brand-specific styles.

“Versioning is huge for us because our network is so huge. That’s what LLMs are good for.”   

Knall also wants to try different versions of an article for Facebook, Twitter, Instagram, Google, a newsletter, and the app. He estimates they have up to 10 versions of an article right now, and this applies to both wire copy as well as enterprise journalism. 

At the moment, Ippen offers readers the option to summarise articles in bullet points or as a longer precis. If you are in their app, a little magic wand appears that lets you choose your summarisation option.

“We want readers to choose the best option,” said Alessandro Alviani, who heads Ippen Digital’s 10-person, cross-functional AI team building these tools. “We want to make sure we can reach everyone.”

Knall wants to experiment with reporters using AI to write: “We try to imitate what an author would sound like even if he just uses AI to complete his text.”

Ippen is also working on English versions of its articles, has created a German transcription and summarisation tool to help its newsroom, and is building a chatbot that will pop up on articles and allow the reader to seek further information on the topics mentioned. 

For images, AI is used in two different ways, Alviani said: “We are using AI to extract the most important information in an article and make sure that we have the opportunity to ping the picture databases from AFP or DPA, etc., to suggest to editors the pictures that fit the article best with a limited number of suggestions. Manual searching is very time consuming.”

Editors are also able to generate pictures with a couple of keywords. Ippen has clear editorial guidelines around this, Alviani said. 

For its videos, ChatGPT is writing scripts, which are combined with a synthetic voice. Ippen now produces four times more video content than it did a year earlier with the same number of people.

There is another important project under way, Alviani said: “Fine-tuning our own LLMs using open source models. The idea is to become less and less dependent on OpenAI. It is really important to us. We see there is huge potential impact.” 

The team is also creating additional tools to support editors and reduce the risk of hallucinations. “We have a project to determine thresholds for a tool to evaluate accuracy — the LLM will reprompt itself if results are below the threshold to reduce hallucination risks,” Alviani said. 

Does the topic of this newsletter resonate with you? Is there something in particular you’d like us to cover here? Please drop me a line or DM me in the INMA Slack workspace. 

“We fully understand the same size won’t fit all”

HT Media is experimenting with AI in personalisation. What could be more personalised than letting the user themself pick the content they want to consume?

So, India’s HT built a chatbot that lets readers ask questions about business and economic news. It recently went live on one of HT’s sites, Mint, which specialises in financial news. 

Yudhveer Mor, chief product and technology officer at HT Media in New Delhi, said it was very important to his team to understand the source of the data. That was a challenge in abstract language models. “We are now using models that were trained on our articles.”

They have limited the use cases for chat to content that is suitable for Mint, he said. “For example, on Mint, I don’t want people to ask about the weather outside.” A user on the site is offered a list of sample questions to pick from, based on the biggest news stories of the moment.

HT has been impressed by the high click-through rate and level of engagement with its chat. 

“It’s a very big step in how people are consuming content,” Mor said. “Our devices tend to be very immersive. In a chatbot, you are nudging people to look at least five options and click on one of them, instead of just passively scrolling as content is streamed to them (as on Instagram). That additional user intervention is pretty hard. A user may spend five minutes on the Mint app and 50 minutes on Instagram. How to bring it naturally to them is my focus area.”

Another AI application is, of course, recommendation engines — but not just for the reader.

“How can we do content recommendations? For the user: What is the next article you should read? And for the content team: What is the next article you should write? This is something we were able to leverage. We fully understand the same size won’t fit all,” Mor said.

Almost 90% of the visitors to HT Media’s sites are first-time users, so personalisation cannot be based on specific information about that person, Mor said: “For first timers, what is the propensity to read further? OpenAI helped us a lot with that. We would put an article in ChatGPT and ask what they will read next. That started driving a lot of engagement for us.

“We haven’t yet achieved our full potential. We are just scratching the surface right now. It has opened our thinking process.” 

HT is also experimenting with text-to-audio and text-to-video features. They tried a feature where users could listen to articles in an Indian accent, but usage turned out to be low.

“The opportunity cost is very low now with all these tools,” Mor said. “This is a great time to do as many experiments as you can.”

What is your organisation doing with GenAI? I would love to hear from you and to connect you with your peers in the news media industry. Please get in touch. 

What I’m hearing

• When we chat about GenAI, news publishers are busy thinking about how audiences will consume their content in the future and how that will play out. Is the solution small, tight channels, like WhatsApp, where you talk directly to your audience without having to go via those unreliable partners, search or social? Let me know if this is a challenge that you’re thinking about as well.

• Most publishers are trying to think of ways to not be dependent on a single tool or platform, preferring instead to draw from a wider pool.

Worthwhile links

• Great examples: We had a fascinating conversation with Ringier Axel Springer and DPG Media, where we heard about intriguing use cases of GenAI, during my first Webinar for INMA’s GenAI Initiative last week. The recording is available free to all INMA members here, and coverage of the article is available here.

• Opinion: Semafor is using GenAI to pull together perspectives from across the world, in different languages, on current affairs. 

• Data journalism: How the Pew Centre used GenAI to shorten “the timeline for doing boring or rote classification work.”

• Be my guest: Politico embraces Web crawlers (about 90% of U.S. news publications have blocked them).

• Fact-checking: Is ChatGPT any good at it? Actually, surprisingly good.

• ROI: A 10%-20% increase in productivity by using GenAI for daily tasks? I read this report by the Boston Consulting Group with interest. How much of an ROI are you seeing on your GenAI efforts? I'd love to hear about how you're evaluating it. 

• ICYMI: Google has new GenAI tools for both writing and advertising, as well as new features for digging through data.

• “Stochastic parrots” no more: Chatbots may be able to actually understand text. 

Date for the diary: February 28  

I know from my conversations with you, fellow GenAI enthusiasts, that many of us are thinking about building chat products. Please join us on a Webinar, free to INMA members, where we will look at specific case studies and talk to those that have built the products about what they have learnt and what they would do differently if they had the chance.

A non-AI diversion

Remember how the Brits voted to name a boat Boaty McBoatface? The good folks in Minnesota name a snow plow every year, and the names that didn’t win are truly as memorable as the winner: Taylor Drift. 

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.

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