Although Aftonbladet has been working with AI for a few years, the “AI boom” that began in late 2022 prompted the company to start experimenting with and implementing generative AI into its journalism.

During the recent INMA Webinar Insights and inspirations for building GenAI chat products, Aftonbladet’s Martin Schori shared the company’s experiences creating a chatbot for the Swedish elections in June.

Schori, deputy managing editor and associate publisher, said Aftonbladet had enjoyed some success with its GenAI features, the most successful of which was the AI-generated article summaries.

But, while the newsroom had plenty of ideas about how to use GenAI, Schori said there wasn’t time to prioritise working on those ideas.

“That’s where the idea of the AI Hub was born,” he said. “We felt that if we want to be really fast and experimental, we needed to build something on top of the organisation.”

The idea was to create a team and disconnect them from their regular jobs for six months, allowing them to focus on GenAI projects. Launched in October 2023, the AI Hub was extended to a one-year programme powered by a cross-functional team.

Education foundation   

A critical starting point for the AI Hub was to ensure that everyone in the company was educated about using prompts correctly. The Prompt Queen Class trained employees on the best ways to use ChatGPT and get the best results.

“After the educational part, we developed lots of editorial tools,” Schori said.

One tool allows users to input the link to a published story and receive five suggestions for follow-up stories. Another project scoured the newspaper’s content to see if all voices were being represented. That research revealed a significant gender gap in coverage, with about 80% of articles and photos featuring men.

“We can’t control everything: our prime minister is a man; we have to write about him. Men shoot each other and start wars, and we have to cover that as well,” he explained. “But some of our content we can control, for example, which experts we are using. So one of the first things we did after this big investigation was to create a GPT that will find us female experts.”

The tool allows users to input the type of story they are writing and it will find relevant female experts.

Bring on the bot

After experimenting with smaller in-house tools, the AI Hub embarked on its most ambitious project: creating a bot that could answer questions about the June 2024 elections. Schori said the project was risky because “the election’s not a very engaging news event for Swedish people,” and it was uncertain whether users would be interested in using it.

Named Valkompisen, which means “election buddy” in Swedish, the friendly, informal robot was given access to Aftonbladet’s database of interviews, surveys, and more than 100 different sources. It was also instructed not to hallucinate and only to get answers from the database. It had many pre-prompted questions, but users could ask any question.

“Of course, we had the people trying to break the bot or who were just trolling,” he said. “But a vast majority of the questions were legitimate questions about the elections.”

While the AI Hub was hopeful that Valkompisen would get some use, Schori wasn’t holding his breath: “I told one of my colleagues that maybe we only get 10 questions for the whole period,” he recalled. Instead, more than 17,000 questions poured in during the first 24 hours.

“That was pretty amazing.”

The project was a lot of work for the journalists, who changed some of the pre-prompted questions, examined the questions being asked, and fact-checked the answers provided by the bot.

“Last time I checked, we have fact-checked thousands of questions out of the 180,000 that we got. Since we got so many questions, we haven’t been able to fact-check every one, but so far, no hallucinations.”

Valkompisen remained active for about a week after the election so users could ask about election results. Behind the scenes, the AI Hub was removing everything that was no longer relevant so that information wouldn’t be included in the results.

While it took a lot of work from the team, it had impressive results. At the end of the elections, only 40% of the questions were pre-written prompts; users wrote the other 60%. To submit questions, a user had to be logged in (although a subscription wasn’t required), and Schori said this yielded more than 11,000 new log-ins.

“That’s a pretty crazy conversion rate there. Our initiatives usually have around 5% [conversions], and this one landed on 40%. So that was great.”

Insights for the future

Since Valkompisen was considered an experiment, Schori said the team didn’t set any goals in terms of revenue, log-ins, or the number of questions asked. However, the results exceeded expectations and provided ideas of what to cover in future elections and what people are most interested in learning about.

One of the biggest takeaways, he noted, was creating these tools requires a lot of work.

“We have been working with the efficiency part,” Schori said. “All of you guys have been to conferences and hearing about how much time and money we’re going to save with generative AI. We have tried. We haven’t found that yet. I’m not sure that it’ll come.” 

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