The final speaker ion this session at the 2026 International Communication Association conference in Cape Town is Jana Peters, whose paper compares opinion poll predictions related to the unexpected 2025 German federal election with predictions generated by ChatGPT. How reliable are AI-generated election polls?
This relates to concerns about the role of horse-race journalism coverage in shaping public opinion during election campaigns, as citizens’ encounters of such polls might lead them to shape their own voting intentions accordingly (variously through bandwagon or underdog effects, for instance). And election polls are themselves not necessarily reliable, although citizens tend to trust these regardless of the reliability of the polling institute conducting them.
Some 44% of German Internet users now use AI, and LLMs are increasingly replacing conventional search engines; such systems, however, remain opaque, and their results vary considerably depending on how exactly they are prompted. How reliable, then, was AI-generated poll information in the final phase of the 2025 German election campaign?
The project prompted ChatGPT from new sessions each day between 29 January and 23 February, from two different devices; it consistently asked for the expected election result percentages for the major parties, from several polling institutes, and for an average of these results, and compared these results to the actual polls.
Overall, ChatGPT broadly reported accurate results, though for some parties (CDU/CSU, Left Party, AfD) there were significant deviations on some days. Results from both ChatGPT sessions were somewhat inconsistent, too. Some days showed notable divergences across multiple parties.
This points to potential problems, both with the consistency between separate ChatGPT sessions and the reliability of results provided to users.
