The Swedish chatbot political-bias study found no clear preference for any parliamentary party in the political writing it tested. The Social Democrats were closest to all six model profiles, but that match was not statistically significant after adjustment for multiple comparisons. The results also varied by model, issue, and task.
The study found no clear party preference
Researchers compared generated writing with positions taken by Sweden’s eight parliamentary parties. Although the Social Democrats were the closest match for all six models, none of the within-model differences in distance to the parties remained statistically significant after the adjustment. That result does not establish that a chatbot favored the Social Democrats.
How researchers designed the test
The study tested six large language models, or LLMs, on Swedish-language writing tasks ahead of Sweden’s 2026 parliamentary election. Its authors crossed policy propositions with writing templates and three kinds of prompt framing:
| Study dimension | Reported details |
| Models | Claude, DeepSeek, Gemini, Mistral, ChatGPT, and Grok |
| Policy propositions | 107 Swedish political propositions |
| Writing templates | 77 templates |
| Prompt framing | Neutral, positive, and negative framings |
| Responses | 24,717 prompts per model; 148,302 responses across six models |
| Party comparison | Model responses compared with the positions of Sweden’s eight parliamentary parties |
The three framings describe how prompts were worded. They are separate from the patterns researchers found in the generated text, such as ChatGPT’s more frequent neutral or ambivalent responses.
Model profiles depended on the issue and task
Claude, DeepSeek, Gemini, and Mistral had similar profiles in the test. ChatGPT more often produced neutral or ambivalent text, while Grok differed most on migration, crime, and gender. The researchers also reported that models largely followed instructions when prompted to argue for or against a proposal.
The study does not identify a universally least-biased chatbot. It examined generated political writing, not how users respond to it: the experiment did not include prior conversation histories or measure whether chatbot answers affected voting.
