In the lead up to Sweden's parliamentary elections this autumn, a research team led by Chalmers University of Technology, Sweden, investigated how six large language models engage with political issues. A review of nearly 150,000 AI-generated texts showed that the chatbots' responses did not benefit any single parliamentary party, but usually placed themselves in the middle of Swedish politics. Only one language model stood out significantly: Grok.
The 2026 election was the first parliamentary election in Sweden since AI chatbots became a serious part of everyday life - not least for young people and first-time voters. Today, it is possible to turn to a number of services, such as ChatGPT, to get information about politics, reason about political issues or get help to polish up arguments. Discussions have also arisen in society about whether political values are expressed in AI-generated texts, and if so, which ones.
To investigate this, researchers used six large language models - Claude, DeepSeek, Gemini, Mistral, GPT and Grok - to take on 107 political issues. The models, which form the basis for the chatbots we use in our everyday lives, were given tasks that are similar to what humans can ask an AI for help with; from writing analyses and producing arguments to formulating debate articles and posts for social media. With 77 different writing tasks, there were just over 148,000 texts to analyse, on issues ranging from whether schools should be nationalised to whether taxes should be reduced. In the first stage, the texts were classified as either neutral, positive or negative based on the wording of the issue. In the next step, this was compared with how the parties had taken a position on the same issues in the autumn's election campaign.
No clear party-political affiliation
The results do not show any clear preference for any individual parliamentary party.
"It is not possible to say that the models support any party. Rather, we see that the models often end up somewhere towards the middle of the Swedish political landscape," says Bastiaan Bruinsma, senior lecturer at the Department of Computer Science and Engineering at Chalmers University of Technology and the University of Gothenburg.
The researchers could still see some differences between the compared models. GPT, which is the basis for ChatGPT, distinguished itself by mostly delivering neutral and balanced texts, while Grok, which is developed by Elon Musk's AI company xAI, deviated significantly on issues of migration, crime and gender.
"Most of the companies behind the models seem to strive for fairly neutral and inoffensive texts. Grok is different. Elon Musk has stated that he wants the model to be more in line with his own political views, and we can also see more of those values in texts generated by Grok," says Moa Johansson, Professor at the Department of Computer Science and Engineering at Chalmers University of Technology and the University of Gothenburg.
The researchers' new report also shows that the language models' responses are affected by the instruction they receive. When the models were explicitly tasked with arguing for or against a political proposal, they largely followed that direction.
What happens when AI gets to know the user?
The researchers point out that there is an important difference between the study's experiments and how ordinary users encounter AI in everyday life. When someone uses a chatbot, there is often an ongoing conversation, where previous questions and answers can become part of the context. This was not included in the study, but is something the researchers plan to explore further.
"We have examined the language models in isolation, but it is different when you regularly use a chatbot. Then your history is also included. A next step is to investigate what happens then, but also to look at how the answers can actually affect how people vote," says Bruinsma.
