How well can language models now provide very recent information from the web? We wanted to see whether there had been any progress, using a specific example: Bavaria’s local elections on 8 March and the mayoral run-off elections on 22 March. We tested how accurately Google’s AI Mode and Perplexity could provide election results.
We carried out the tests for both larger cities and smaller municipalities that received relatively little press coverage, and the results varied. Overall, Google’s AI Mode and Perplexity showed very similar strengths and weaknesses.
High-profile information was generally easy to retrieve, for example who the incumbent was, who was no longer standing for election or who had been voted out of office. Even when we asked about the incumbent without mentioning the recent elections, both systems consistently referred to the election or to the fact that the current term of office was about to end. The percentage results of mayoral elections and any subsequent run-offs were also mostly reported correctly. In addition, the AI tools provided plausible general assessments, for example of voter turnout or the strongest political force.
More difficulties arose with detailed questions about municipal or city councils. Here, inaccuracies were more common: percentages were sometimes omitted, seat distributions were not always correct and individual details contained minor errors. In some cases, for example, only a politician’s first name was wrong. Exact voter turnout figures were also not always provided, even though a general assessment was usually possible without difficulty.
Our conclusion: Overall, the AI tools reproduced a large amount of information correctly, and their general assessments were also accurate. However, minor errors continued to appear. This combination of largely correct information and occasional inaccuracies can be particularly problematic because mistakes are easy to overlook and may be accepted uncritically. Careful verification therefore remains essential. For more detailed research requiring exact seat distributions, names and percentages, using AI models does not currently save time compared with consulting the primary sources directly, such as municipal or city websites.
Despite these remaining challenges, we view the overall development positively, particularly in terms of how well very recent events are now incorporated into AI-generated answers.
By Markus Hoffmann
