The goals formulated at the time were largely around using AI for a richer news experience and improved technical efficiency at Ekstra Bladet, noted Kasper Lindskow, Head of AI, JP/Politikens Media Group, at our recent World News Media Congress in Copenhagen. 

In January 2024, JP/Politikens Hus shifted the focus from Ekstra Bladet, to a new centralised model for the entire group, including Politiken and Jyllands-Posten – the other two national flagship brands.

“We believe we are part of driving a paradigm shift in making journalism more efficient with AI, and envisioning the state of things 5-7 years from now,” said Lindskow, adding, “And we want to do it in a way that enhances the missions and business strategies of each of our brands.”

To achieve this, the newsroom focuses on three product domains – GenAI, metadata and recommender systems. At our Congress, Lindskow deep-dived into the products the organisation has built using GenAI in the past year.

Leveraging early AI experience

A new shared AI unit was launched in February, and its first creation was a ChatGPT clone. This multi-tasking API chatbot tool, called MAGNA Basic, enabled journalists to interact with a large language model (LLM) and was rolled out company-wide.

The decision to clone the technology had two motivations: to prevent tech vendors from accessing their data and, more critically, to ensure the company retained control of its data.

“We’re learning a lot from observing how our employees use MAGNA Basic, and are taking notes to finetune the system, making it better equipped to perform those tasks,” Lindskow said. 

However, more importantly, this experience provided the knowledge to develop another ongoing product – MAGNA Custom.

Workflow optimisation and future innovations

MAGNA Custom connects MAGNA Basic to the newsrooms’ local data sources and adds an intuitive UI, allowing journalists to prompt the model and simply push buttons to carry out tasks. 

“Our goal is not just to build tools for the 10-15 percent frontrunners but really for the 70-80 percent of the newsroom that might not be ready to engage directly with an LLM,” he said.

To this end, JP/Politikens is developing three groups of tools:

Live-blogging support: This tool assists journalists in live-blogging events such as court cases or football games. It offers functionalities for creating live summaries, sorting through unpublished user comments, and providing an overview of user interests.
Article drafting assistance: This set of tools helps journalists draft articles, from correcting small spelling errors to generating drafts from just a few basic facts.
Archive integration: A third tool connects newsrooms to the news archive, aiding journalists to search the archive and perform generative tasks using archived content.

An early version of MAGNA Custom was rolled out at Ekstra Bladet six months ago. 

Moving forward, the group will be merging the two MAGNA UIs: the classical one and the new interface emerging in the GenAI space.

“This means we’ll offer both button options and the ability to continue working with the model’s output through dialog,” Lindskow said.

The team is refining the tool and will be rolling it out to all its newsrooms.

“Since MAGNA has no expiry date, we’ve already begun thinking about its next iteration, and will continue to develop and improve it,” he added.

So far, MAGNA has primarily focused on streamlining the existing workflow. In the next phase, they aim to introduce features that enrich journalism and news coverage, allowing journalists to explore new possibilities with automatic versioning and text generation.

Grounding GenAI in facts a must: Learnings

Lindskow noted that a clear vision and analysis around where publishers picture themselves five years from now is crucial.

“It’s easy to work with GenAI to assist and automate, but a lot harder to find the right cases for using it to enrich our journalism. We need experimentation and an openness to it,” he said. 

He shared a few more learnings:

GenAl must be grounded in external facts when used for news.
Gravitation towards the median is a key challenge for alignment with editorial profile. 

“We go to great lengths to prompt the model to get it away from the mean into the writing style of each of our individual publications. We have task definitions, style and language guides, grounded in external output and that’s really as much as we can push it with prompting,” Lindskow said.

Ease of use and integration into existing workflow is essential for mainstream use. The group is working on integrating MAGNA into the CMS itself.

The post How JP/Politikens leveraged early AI expertise to build centralised news products appeared first on WAN-IFRA.