Norway’s VG finding success with six AI-related newsroom tools – WAN-IFRA

5 minutes, 2 seconds Read

Established in 1945, and online since 1995, VG is one of the country’s most popular news sources.

“In 2024, we are in a very strong position in Norway. Our traffic comes directly rather than from social media platforms,” Jari Bakken, Newsroom Engineer, VG, told participants at our Digital Media India conference in New Delhi.

See also: Mastering frontpage engagement: Norway’s VG aligns strategy and analytics

‘ChatGPT changed the way we use AI’

ChatGPT has opened up the space for developers to use AI in new ways, Bakken said.

For example, today, developers are exploring different language models and APIs (Application Programming Interface) to build customised applications.

Bakken is part of a task force exploring the possibilities of using generative AI in VG’s newsroom. Although the team has built many tools internally, there are six AI-related ones that stand out, he said.

1- Article Summaries that appeal to young readers

Article summaries are being generated using OpenAI’s GPT-4, an improved version of the AI model.

The full article text is fed into ChatGPT, which returns a summary in bullet points that is placed inside an expandable box below the headline.

Young readers click the summary box more often

The expandable box helps VG track the number of readers clicking on it and how often it is used.

“This has been quite a great success,” Bakken said.

Use of the summary box has been growing, and it is especially popular with young readers.

VG’s internal analysis of click-through rates for February-March shows that 28 percent of readers aged 15-34 who see the box click it, compared to 18 percent of other readers.

However, there have been concerns. For example, there was some concern whether providing AI-generated content directly to the reader was a good idea.

To address this, a journalist is always involved in reviewing and rating the AI-generated summary. A disclaimer indicating the summary is AI-generated is mentioned inside the box to ensure transparency. Additionally, the final output is compared to the AI-generated text to further train the AI models.

The other concern was whether people would only read the summaries, potentially reducing overall reading time. In fact, people who clicked on the box tend to read deeper into the article, spending more time on it.

“This engagement rate is consistently about 20 percent higher than those who did not expand the article. This was a bit surprising, but interesting,” Bakken said.

2- Jojo app: A transcription tool for journalists

VG’s editorial development team built a speech-to-text app based on OpenAI Whisper – an open-source transcription model.

Jojo was developed after studying reporters in the newsroom to identify ways to make the reporting process more efficient. Interview and press conference recordings can be uploaded to the Jojo app to get transcribed text of the audio.

Over 10,000 hours of audio have been transcribed so far

For journalists worried about uploading sensitive interview recordings, the team developed a desktop version of the app where everything operates locally on the machine.

“We’ve also integrated Jojo into Slack for live transcriptions. This is beneficial for press conferences, where manual transcription used to be labour-intensive,” Bakken said.

3- Statistics Agent: Crunch tables, create charts

The Statistics Agent tool was developed to create charts and graphics for visualising statistics, one of the daily tasks for the developer team at VG. It is based on a paper published by researchers at Google, called the ReAct Agent.

The team used to locate the correct table in the official statistical database of Norway and select the right figures and then convert them into a graphical chart. The new tool can do the same process using AI. The generated chart is double-checked before publishing.

“Sometimes, the tool will find the wrong table, or it will crash if the table is too big,” Bakken said, adding, “But this shows the potential and the direction in which we might be headed with tools that can perform advanced tasks.”

4- FOIA bot: AI understands law

This bot was created to write appeals or complaints for Freedom of Information Act (FOIA) application denials. The bot has a user interface where the denial letter can be pasted, and it generates complaints based on the input.

Although ChatGPT can write a decent complaint, it doesn’t have an understanding of Norwegian FOIA laws.

Therefore, the team developed a Retrieval Augmented Generation System with access to some legal resources. It contained templates for drafting complaints from Norway’s press agency.

This enabled the AI to generate better complaints, provided that the relevant sources were given with the prompt.

“That worked a lot better because it is able to understand quite detailed things in the law. It’s a much better tool than the raw AI,” Bakken said. “But it can still hallucinate, so everything must be double-checked.”

5- Sequelizer: Ask AI to create story ideas

Sequelizer was created to generate story ideas for journalists based on their articles. The tool can suggest follow-up ideas for a particular story just by pasting the link to the article.

“It’s not great, but some ideas are good, some bad, but it’s interesting. It gives a kick to creativity,” Bakken said.

The tool is also built into VG’s Slack channel where journalists can generate a list of follow-up ideas with a click of a button.

6- Text-to-data: A tool for investigative journalists

This tool is more towards investigative journalism, Bakken said. It can read through an article and extract data about people, places and organisations that are mentioned in the article, and try to find the relationships between and among them.

“It can be really interesting if a journalist can run this across a thousand articles. You will get an interesting database about relationships between all the people in the articles you mentioned,” Bakken said

This shows the potential of having AI do some parts of the investigative work and maybe tasks that we wouldn’t have done ourselves manually, Bakken said.

Also during Digital Media India, the Hindustan Times Digital Streams received the Gold award for Best Use of AI in the Newsroom.

This post was originally published on this site

Similar Posts