How AI search summaries are entering Switzerland’s democratic debate

43% of Google queries produced an AI-generated overview during research into Switzerland’s “No to a Switzerland of 10 million” referendum campaign. The figure places an automated summary directly inside the information flow that many voters use to understand an initiative, its arguments and its likely consequences.

AlgorithmWatch CH and the University of Zurich analysed more than 170,000 automatically generated search queries across Google and Bing. Their work examined when an overview appeared and which sources it used. The study arrived as the campaign, launched by the right-wing Swiss People’s Party, exposed sharp divisions over population growth, immigration and the country’s future capacity.

Google’s summary can appear above the familiar list of links. That changes the first encounter with a political issue. A user may read a compact answer, accept its tone and stop there. Hertweck of AlgorithmWatch CH says AI overviews are becoming increasingly significant in shaping public opinion in a democracy.

Switzerland’s direct democratic system gives that shift particular weight. Voters regularly decide on initiatives and referendums, often after short campaigns packed with competing claims. The wording and selection inside a search summary can therefore enter a debate that once depended more visibly on campaign material, journalism, official documents and conversations.

Around one third of the sources behind Google’s answers were judged balanced. Researchers identified news articles and government websites among the material that offered relatively even treatment of the issue. That finding gives the summaries a connection to established information sources, while leaving open how the system selects, combines and ranks them.

The researchers also found more cited sources critical of the initiative than supportive of it. AlgorithmWatch CH cautioned that this result does not show which way the generated answers themselves leaned. An overview can cite material from one side and still give greater space, prominence or clarity to the other.

That distinction matters because AI summaries compress a political dispute. They can combine several documents into a few sentences, remove the original context and present a smooth answer without making the editorial choices easy to see. Users may see links, yet lack a clear account of why particular sources appeared or how the system resolved disagreement between them.

The study did not provide a quantitative assessment of the balance of the responses. Hertweck said some overviews gave one side a longer explanation while mentioning the other only briefly. Researchers therefore describe their findings as an important first picture, rather than a final verdict on the political content of Google’s system.

Framing can influence a voter before the campaign’s evidence is opened. AI-generated overviews often list arguments for and against an initiative, according to the source article. Their influence can emerge through less visible choices: the order of points, the length of an explanation, the adjectives used and the sources placed closest to the claim.

Zurich advertising agency Feinheit examined chatbots and AI overviews during the same campaign. Its analysis found that ChatGPT and Google Gemini highlighted more arguments against the initiative. Feinheit described Google’s AI Overviews as more favourable to the initiative’s supporters. The findings came from a separate analysis and do not replace the AlgorithmWatch CH and University of Zurich study.

External research from the universities of Maryland and Washington found that AI-generated overviews were particularly persuasive for users with little knowledge of a subject or strong trust in artificial intelligence. That finding points to a practical challenge for political communication. A first-time voter looking up a complex initiative may rely heavily on the first coherent explanation that appears.

Hertweck says the summary’s plausible tone can carry authority. The risk grows when users accept the answer without checking the underlying links, campaign documents or official material. A concise explanation may save time, while also reducing the effort required to test its assumptions.

Zero-click searches are changing the route from question to conclusion. When Google answers a query on its own page, users may never open the articles, government pages or campaign documents listed below. Hertweck says people receive fewer links at the top and more AI-generated overviews, a format that can make them less critical of the information presented.

That behaviour matters in Switzerland, where voters receive official information with their ballot papers and where the neutrality of political explanations remains part of public debate. The official voting booklet follows rules intended to present the arguments around an issue fairly. A commercial search system operates through a different design, with its own training data, ranking methods and product decisions.

The distinction does not automatically make one source reliable and another unreliable. It does make their responsibilities and transparency different. Voters can identify the authorship and formal status of an official booklet. An AI overview may provide citations while offering little explanation of how it selected, summarised or weighted the material.

AlgorithmWatch CH is calling for greater transparency from AI companies and tougher regulation. Those demands target the information infrastructure around campaigns, where private technology firms can shape visibility without standing as identifiable political actors. The referendum debate has made that influence easier to see.

Switzerland’s next democratic safeguard will need to include the search box. The evidence from the 10 million campaign does not establish that Google’s overviews consistently favour one political position. It shows that the summaries appeared frequently, drew on mixed sources and could present unequal explanations of competing arguments.

That creates several immediate tasks for researchers, regulators and voters. Independent studies need access to representative queries and archived outputs so that systems can be evaluated during live campaigns. Technology companies should explain the sources, ranking signals and safeguards behind political summaries. Regulators must decide what transparency should require when a private platform becomes a major gateway to public information.

Media literacy also needs to become more practical. Voters should learn to identify an AI-generated answer, open its citations, compare the wording with the original source and consult the official voting booklet. News organisations and civic groups can help by publishing clear explainers that distinguish verified facts from campaign claims.

The Swiss debate will continue through initiatives, referendums and elections. Search engines will remain part of how people prepare to vote. The standard for democratic information will increasingly depend on whether citizens can see how an answer was assembled, assess its limits and decide what deserves their trust.