By Cecilia Campbell, WAN-IFRA Innovate Local Program Editor.

Amedia’s AI Sandbox aims to enable more and better journalism by letting reporters use AI as an assistant. This means they can spend additional time in the community covering even more critical local stories. The sandbox environment also has a trust dimension. By setting it up, Amedia guarantees safe AI use for local journalism. As an additional bonus, Amedia wants more than 1,000 journalists to use the new sandbox to help guide their product development organically.

Background and context of the job to be done

Amedia encompasses 120 news titles across the entire length of Norway, with 1000 journalists on the ground in local newsrooms and communities. The principal business model is subscriptions, and the group has over 700,000 subscribers, generating 2 million daily readers in a country of 4.3 million adults.

Markus Rask Jensen has spent many years working in local newsrooms, but as of 2023, he is part of the central Amedia team, which supports the local teams. “The local newsrooms are our core, and that will never change. Everything we do centrally is geared to solving problems that arise locally.”

“When it comes to AI specifically, developing the sandbox is not something we do to gain efficiencies and produce more of the same generic content. This is about freeing up journalists’ time – to support the meeting between humans, presence in local communities, and our local identity.”

With the economic downturn, resources in newsrooms worldwide are getting squeezed. “This is a problem we must solve – how do we serve our readers with fewer resources? At the same time, they expect more from us, as they compare us to other content and entertainment providers.”

On a side note, Amedia has used AI for many years and has had a dedicated central AI team since 2018. The team works on processes such as data analysis, personalization, and functionality like converting audio-to-text.

So one reason for the new sandbox is to help journalists with routine work, allowing them to spend more time with sources and stories in the community. The other reason to create an in-house AI environment is to do with trust. Markus Rask Jensen pointed out that trust is the currency of journalism, and the safe use of AI is vital for trust. Amedia has set up guidelines for generative AI, aimed at helping local editors, who are ultimately responsible for how they want AI to be used in their newsrooms. The sandbox is a physical (albeit digital!) manifestation of the guidelines – this is a safe environment for journalists to experiment and use AI.

“We want to experiment with editorial use of AI in a safe and protected environment. And in terms of the bigger picture here – we need to protect our content; subscriptions drive most of our revenue. We’ve deliberately been careful when feeding content from behind our paywall into LLMs. With our own AI system, we’re in control of input and output,” Markus said.

What is the tool, and how is it used?

This is what Amedia’s AI hub looks like:

The vision

“A set of tools that helps our journalists become better at their jobs, more efficient, and enables them to spend more time meeting real people – to write stories from the real world.”

The team

In terms of organisation, the project is driven by two initiative leads in the central organisation. Eight editorial participants are working as ambassadors in a number of local newsrooms. All 1000 journalists are encouraged to use the tool, but voluntarily. “We’re not saying you have to use it, which I know some of Schibsted’s titles have done. We don’t want to do that. Instead, we’d try to inspire them to use it. But to be fair, I don’t think journalists realistically can avoid using AI tools today.

The tool

Markus Rask Jensen did a short sandbox tool demo– watch the video for this!

It’s plugged into three Large Language Models including ChatGPT4 and Google Gemini. The user interface is reminiscent of the now familiar ChatGPT interface, with space for prompts and answers. There are also a number of shortcuts in the shape of buttons at the top, to functions like summaries, headline suggestions and text-to-audio.

In addition to providing a safe environment enabling input and output control, the sandbox also leverages what Markus called the “Journalistic hive mind of Amedia.” As journalists use the tool, the usage is logged. This allows the organic use by journalists to guide further development of functionality. So far, the usage patterns show that chatting with headlines is the most popular feature.

It’s up to the local editors to communicate to readers about the use of AI in the production of journalism, but the central guidelines encourage complete transparency. “We recommend that whenever the AI tool has been used, you tell readers: ‘This article was produced with the help of AI tools. This is a really important part of the trust I mentioned.”

What’s next?

Markus foresees a rapid development of new functionality. The team had a brainstorming session around new ideas. See below:

The fact-checker is being worked on now. This is to help with the hallucinations prevalent in LLM output. “The reason LLMs hallucinate is that they predict. But they can only predict based on the context you give them. If you anchor the information they are to use, you can use them to fact-check. It’s a different kind of model – basically, instead of asking all the training data of ChatGPT, who is the mayor of Tromsø, you anchor a verified source of information, which the LLM will then use,” said Markus.

The auto-writer is also coming – in short, this tool quickly writes a basic article on a defined story, such as a coming weather front. With the AI tool handling the base article writing, the reporter has time to talk to sources locally to build the story out.

Summary of why Amedia built their own AI sandbox

• Flexibility in terms of third-party providers.
• Ensures data security through the use of secure login.
• Cost efficiency and control.
• Amedia gains knowledge of business needs and potential use cases through use patterns.
• A chance for journalists to organically help guide the direction of product development.
• A way to ensure that they spread expertise throughout the organization.

Markus Rask Jensen emphasised that creating a basic sandbox is not a huge technical project. “The main body of work for our project took 1-2 people about a month to do, and not full-time. I think a newsroom could hire developers to spend 10–15 hours to set up a safe generative AI environment like this.”

Useful links

The case is about Norwegian local media group Amedia

Get a feel for their local journalism (whether you speak Norwegian or not) at one of their titles, such as Glåmdalen Firdaposten or Lofot Tidende.

Markus Rask Jensen’s presentation slides are here.

OTHER RESOURCES AROUND AI / LOCAL

• Report: Practical AI for Local Media. Published 2023 by Media Voices in the UK. Lots of cases and best practices (pre-GenAI) on use cases like personalisation and content automation.
• JournalismAI. Initiative operated out of the London School of Economics since 2019. Lots of resources, including a database of cases.
• The team around Professor Nick Diakopoulos at Northwestern University has put together a very useful overview of newsroom guidelines on generative AI from a couple of dozen newsrooms. They have also set up a webpage as a hub for resources around generative AI in the newsroom.
• The Local Media Association has just published a Strategic Guide for Publishers in the Age of AI Search.

If you have questions, contact Cecilia Campbell at the WAN-IFRA Innovate Local team. If you know of other similar cases, please let us know, and we will add them here. Thanks in advance!