AI results at Google: information roulette in the local elections

In late September, voters in three German states, Sachsen-Anhalt, Berlin, and Mecklenburg-Vorpommern, will cast their ballots in regional elections. Amid this political activity, concerns have emerged over how Google’s AI-generated search summaries handle questions related to election-related content. The nonprofit organization AlgorithmWatch has conducted a study examining the transparency and neutrality of these AI-driven snippets, which appear in approximately half of all searches in Germany since early 2025. Many users rely on these brief summaries without clicking through to the original sources, raising alarms about the potential impact on informed decision-making. The investigation was carried out using Google’s internal “Search Researcher Result” API, made available under the Digital Services Act. Researchers were able to submit nearly 4,500 queries related to the upcoming elections in the three states and analyze the AI-generated summaries alongside their cited sources. According to the findings, Google does not provide clear criteria for when these AI summaries are displayed. The lack of transparency has been criticized as potentially arbitrary, with implications for democratic discourse. While the study found that Google generates AI summaries less frequently for election-related queries compared to non-political ones, it remains unclear whether specific measures are being taken to address election content. There appears to be no consistent pattern in how the AI decides to display its summaries. For example, questions about polls rarely triggered AI summaries, while inquiries about parties and individual candidates did so in up to 50% of cases. Differences were also observed between the states: fewer AI summaries appeared in responses to questions about Saxony-Anhalt than in Berlin or Mecklenburg-Vorpommern. A key finding was that the number of sources cited within the AI summaries was often limited. Nearly half of the links provided pointed to fewer than ten sources, with YouTube, a subsidiary of Google, being disproportionately represented. Among the top ten referenced websites, public broadcasters, Wikipedia, social media platforms, and official government or party sites were included. This suggests a possible bias toward familiar or affiliated sources, potentially limiting exposure to diverse perspectives. Additionally, the study noted that political positions of candidates were frequently presented in a non-neutral manner, with positive language used more often than objective reporting. This phenomenon, referred to as “sycophancy,” may stem from the inherent tendencies of large language models to affirm the viewpoints of questioners. In the context of politics, such biased framing could hinder voters' ability to form balanced opinions, according to AlgorithmWatch. When confronted with the results, Google responded only vaguely, offering no detailed explanation or clarification regarding its AI policies. The company declined to comment further on the methodology or findings of the study, leaving many questions unanswered. As the election period approaches, the role of AI in shaping public perception continues to draw scrutiny, highlighting ongoing challenges in ensuring digital information remains both accessible and reliable.