Artificial intelligence produces relatively fluent text from Finnish-language discussions and can broadly distinguish factual claims from other content. However, journalists are still needed to verify the accuracy of those claims.
Nordis member Faktabaari tested an AI application for the first time in Finland to monitor claims made during an election debate on Wednesday, April 2. The application, developed by Nordis partner Factiverse, collected a total of 128 factual claims from MTV’s regional and municipal election debate featuring party leaders. The debate topics typically revolved around the financial situation of healthcare regions, healthcare cuts and education.
The application generates a near real-time transcription of the televised debate, converting spoken words into text. In addition, it filters out factual claims from the speech and distinguishes them from, for example, political opinions that cannot be fact-checked. The app also searches for available sources to verify the claims and suggests fact-checks. Factiverse AI rates the accuracy of claims on a scale of 0 to 100 (disputed–accurate).
All claims compiled by Factiverse from the party leaders’ debate can be found here.
Of the claims made in MTV’s party leader debate, the application gave the most assessments as “accurate” (støttet). According to Faktabaari’s editorial team, the debate also did not contain a significant number of incorrect or misleading claims, indicating that the party leaders had done their homework. For instance, ahead of last summer’s European Parliament elections, the editorial team fact-checked several inaccurate or misleading statements from individual candidates.
The application identified four claims as questionable. In these cases, a closer look showed that the AI either failed to correctly interpret the claim or to properly source its fact-check.
For example, the AI flagged as questionable a claim by Finland’s Deputy Prime Minister, Riikka Purra: “My party has proposed that doctors should be required to work in the public sector, but we’ve understood that this is not possible under the constitution.” This claim related to a discussion about medical students and newly graduated doctors moving into highly paid temporary positions, which increase healthcare costs.
To verify the claim, the app cited, among other things, a statement from the Finnish Medical Association, which opposes the proposal. However, the AI is not capable of interpreting the Finnish constitution in this matter, so the question of whether Purra’s claim is true remains a subject for journalistic verification. The AI services have so far not had access to Finland’s up-to-date legal database, as some of the associated digital material is copyright-protected. Faktabaari intends to fact-check Purra’s claim.
Overall, in about half of the claims reviewed, the application either succeeded in fact-checking on its own or at least added useful context through sourcing. The application mainly used texts published by Finnish media as sources but also referred to party publications, studies, and reports. Factiverse AI assesses and prioritizes the reliability of media sources based on the Reuters Institute’s Digital News Report, which reflects how trustworthy Finnish citizens consider various outlets.
According to the Faktabaari team, the application already performs well in producing Finnish-language transcriptions and fairly well in compiling claims, although it still sometimes confuses speakers or combines unrelated fragments into garbled text.
Even in its current form, the list of claims and background information provided by the application is useful to journalists conducting fact-checking. However, journalists are still needed for the actual verification. The AI can be seen as a “journalistic support intelligence” that assists and speeds up the fact-checking process.
The tool has already been successfully used in EU election debates in Denmark in spring 2024 and in the U.S. presidential election debates in September 2024. In Norway, the fact-checking media outlet and Nordis partner Faktisk – founded by six major media outlets – has used the tool for verifying facts in TV programs and video content for a couple of years. Norway’s public broadcaster NRK is also considering adopting the service.
Factiverse is a Norwegian startup that provides this verification tool for use by journalists and other stakeholders. As mentioned, both Factiverse and Faktabaari are members of Nordis, which also evaluates the use of AI as journalistic support.
The foundation of Factiverse AI was laid at the University of Stavanger in the latter half of the previous decade. “It is a specialized AI model built by feeding it 50,000 certified fact-checks compiled by the International Fact-Checking Network (IFCN),” says product manager Gaute Kokkvoll. Large language models are only used to summarize the content of the sources for the benefit of fact-checkers.
Development of the model continues in collaboration with fact-checkers from various countries. For example, fact-checkers from the German news agency DPA have participated in the project. Unlike large language models that are trained on massive datasets, Factiverse AI’s development requires fact-checkers to review the AI’s conclusions point by point. This helps the AI learn to recognize claims better in the specific language.