As more news media companies become aware of the many ways they can harness generative AI, they may face a challenge getting internal teams on board with using those tools.

During Wednesday’s Webinar, GenAI audio and internal tools: experiments from two innovative news brands, members heard how Expressen, the Swedish media company owned by Bonnier, created a useful, relevant tool for its journalists — and then got them to embrace it.

Jakob Wagner, head of digital and editorial development at Expressen, walked INMA members through the company’s journey to integrate AI into its newsroom operations.

Like many companies, Expressen began looking into the possibilities presented by AI last year, quickly determining it was something that warranted a closer look. Before long, the news company initiated a comprehensive training programme to help the staff understand the possibilities and challenges of AI services.

That was vital to promoting widespread adoption, Wagner said: “It’s important for everyone to have a basic understanding of different AI services, otherwise they won't really understand the possibilities and the challenges.”

Task forces lead to AI tools

In addition to the training, Expressen formed five AI task forces to tackle different opportunities and challenges, and started experimenting with creating various internal tools using OpenAI. An early discovery was that while GenAI offered good results and provided creative ideas, it was inefficient to have to use separate Web pages for different AI services.

“It takes a lot of time going back and forth and copying and pasting text,” Wagner explained. “So to be able to use it more efficiently, we thought it would be good to integrate different kinds of prompt use cases closer to where they actually produce the content.”

That led to a goal of developing an in-house AI tool that enhanced creativity and efficiency for journalists, freeing them up to do other tasks. In addition to providing creative shortcuts, Expressen was committed to ensuring the tone of AI’s voice was consistent with the brand, which would be essential when creating headlines and other elements.

 Through a series of workshops, Expressed identified four key areas where AI could generate suggestions: headlines and subheadings, social media texts, article summaries, and push notifications. Then it conducted a sprint to “both understand the impact of such tool and involve a lot of people to create engagement.”

It was also a way of getting journalists, machine learning experts, and developers to work together on the development to ensure the tool offers value to the journalists: “So it’s not just fancy tools but also usable tools.”

A new tool of the trade

The initial proof of concept tool that was developed allowed journalists to select different control parameters, such as tone of voice, temperature, and length of the text generated.

It was first tested with a static version using 200 articles and the feedback from the test version was positive enough to be declared a success. After more improvements, it was launched in late December and has been updated continuously since then.

Now, the Chrome plugin allows the user to take an article from the CMS and create a prompt based on user input. Journalists request a suggestion for a story or headline, including what tone of voice it should have and the target audience it’s being written for. Then, the text field offers a few suggestions. Wagner said in many cases, the journalist will combine two or three suggestions to create their perfect headline or phrases.  

Initial testing showed headlines written with assistance from the tool often outperformed those written by editors without assistance. The newsroom has found it helpful and provides ongoing input on how to improve it, such as having the tool provide shorter suggestions or remove words that wouldn’t be used in headlines.

“But it was a great start and something we could put up for everyone to use,” Wagner said.

It has since been launched at other newspapers within the group, demonstrating its scalability and adaptability: “They can use the same kind of prompts and customise them or they can do other prompts with other use cases and use the same technique,” Wagner noted.

Experimentation with the tool continues to improve its output quality, enhance prompts, and develop new prompts, such as SEO prompts. Behind the scenes, machine learning engineers and developers have looked at the underlying architecture and what model to use:

“We have done a lot of comparing of different AI models from different companies to understand what gives us the best results.”

Moving into the future, Wagner said Expressen will continue to experiment, evaluate results, and enhance the tool to provide better and more efficient assistance to its journalists. For example, the current process of regenerating suggestions will become a chat feature.

“In the first version, you could press one time and get suggestions and then you press [again] and select, ‘I want new suggestions.’ That started a new thread in the model, so even though it wasn’t the same suggestion, the model didn't know the previous headlines,” he explained.

“In the upcoming version we wanted the editor to be able to write, ‘OK, I like this, but could you do something more like this?’ And we’re working on different ways to identify and fix for suggestions so that the suggestions that are shown are better and better.”

This is just the beginning and the work is ongoing, with many potential opportunities for use, Wagner said: “It’s used every day in the newsroom, and we work on getting the suggestions better. We are still working on other prompts as well that will be released during the fall. And we also work on getting more people to use it.

“We’ll also diversify what the tool can help with.”

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