How Google’s AI-generated overviews shape Swiss democracy

The referendum campaign on the ‘No to a Switzerland of 10 million' was polarised. Algorithm Watch CH and the University of Zurich have examined the AI summaries that frequently appear above search results on Google and Bing. Gian Ehrenzeller / Keystone

Multiple studies show how AI can subtly shape public opinion in democracies. Swiss watchdog AlgorithmWatch CH calls for greater transparency from AI companies and tougher regulation.

AI is quietly reshaping democratic debate by changing how people find and absorb information. “AI-generated overviews are becoming increasingly significant in shaping public opinion in a democracy,” says Corinna Hertweck, a researcher at the civil society organisation AlgorithmWatch CH.

During the referendum campaign for the “No to ten million” initiative, AlgorithmWatch CH, in cooperation with the Department of Informatics at the University of Zurich, analysed more than 170,000 automatically generated search queries. The researchers examined how often the search engines Google and BingExternal link delivered an AI-generated overview in response to a query and which sources were used.

The study found that 43% of all Google queries were answered with AI-generated overviews. Around one third of the sources behind these answers were considered balanced, especially those from news articles and government websites.

The researchers also examined how many of the sources linked in the AI overviews supported or rejected the “No to ten million” initiative. They found that more sources were critical of it. AlgorithmWatch CH, however, stresses that this does not reflect which way the AI-generated answers lean.   

AI-generated overviews usually list arguments both for and against something. “But sometimes, one side gets a longer explanation while the other side only gets a brief mention,” Hertweck explains. But these are only initial impressions. A quantitative analysis of how balanced the responses are has yet to be carried out.

Framing can influence public opinion

For Hertweck this subtle influence is unsettling. “Millions of people use Google and get AI overviews,” she says. “Subtle framing like this can definitely shape public opinion.”

During the polarised referendum campaign over the “No to ten million” initiative launched by the right-wing Swiss People’s Party, Zurich-based advertising agency Feinheit examined chatbots and AI overviews. Their analysis found that ChatGPT and Google Gemini highlighted more arguments against the initiative while, according to Feinheit, Google’s AI OverviewsExternal link were “on the side of the initiative’s supporters”.

Researchers at the universities of Maryland and WashingtonExternal link have found that AI-generated overviews are particularly persuasive for users with “little knowledge of a topic or greater trust in AI”. This shows how important AI models have become as seemingly neutral sources of information in Switzerland’s direct democracy.

“When I ask Google a question, I no longer get several links at the top, but just an AI-generated overview,” says Hertweck. “It usually sounds plausible and carries a certain authority.“ The so-called “zero-click searches” – queries that are answered without users leaving the Google page – are on the rise. “That can make us less critical of the information we are getting,” says Hertweck.

The difference between the official voting booklet and AI-generated overviews

In Switzerland, the debate of whether and if information is neutral is part of the democratic process. Sometimes the debate is sparked by the reporting of public-service media outlets which includes Swissinfo.

The debate can be even fiercer over the so-called voting booklet that Swiss voters receive together with their ballot papers. The information it contains must be neutral and supporters and opponents must be given equal space to make their case. When it comes to campaign advertising, things are a bit different. There are no spending limits, which gives those with a larger budget greater public visibility.

Search engines’ AI-generated overviews and chatbots have become an increasingly important source of information in Swiss democracy without the usual democratic debate about their role. Yet, no one really knows what their seemingly sober answers are based on.

AlgorithmWatch CH sees threat for media diversity

That’s why Herweck demands greater transparency. “The public needs to know how AI overviews are generated.” The fact that “market power lies with very few companies” is, according to Herweck, a threat to media diversity and reliable information.

She also points to an economic cul-de-sac created by “zero-click searches” on Google. “When people don’t leave the Google page, it weakens media diversity. News websites lose clicks and with them the financial resources needed for good journalism.”

In the end, she says, the reliable sources that AI overviews depend on could disappear. Hertweck also believes that companies such as Google should acknowledge that they no longer merely distribute content but create it themselves through AI overviews.

She also points to a grey area in Switzerland’s draft law on communication platforms and search engines. “The law does not explicitly cover the integration of generative AI, even though we are seeing it more and more often, as for example in AI-generated overviews,” Hertweck explains.

‘Things are moving incredibly fast’

Hertweck heads the “RAISD” research project at AlgorithmWatch CH. Generative AI can impact a democracy because it plays a role in shaping individuals’ opinions, including those of people working in politics. “Apart from AI-generated overviews, we also want to look at how chatbots are used in politics, for example when politicians or their staff turn to them. It could have far-reaching consequences if chatbots are involved in drafting legislation,” Hertwick says of future plans.

The RAISD research project will analyse the interaction between AI, democracy and society until 2028. However, Hertweck does not dare predict what generative AI will look like by then. “Things are moving incredibly fast which is a huge challenge for policymakers and researchers,” she concludes.

Oxford study on how AI can “steer collective opinion”

AI also plays a role in the debate from human-to-human communication. A new study by the universities of Oxford and Potsdam shows that users cannot always trust their own written texts –  if they ask AI to edit them.

AI-mediated communication can steer collective opinionExternal link is the title of the study. For their research, computer scientists asked AI models to” improve” opinion posts. The models softened the views or reversed them on topics ranging from abortion and feminism to climate change.

Professor Sandra Wachter and her team then looked at what could happen during these interactions between humans. If, for example, an AI-influenced LinkedIn post receives a response that has also been influenced by AI, the discussion gradually shifts – both for the participants and for outsiders watching the debate.

Pro-Life-Bias in Grok

“Our work has shown that we need a broader discussion on how we feel about language models and their developers because the policymakers may not have been aware of the risks when they drafted the EU AI Act,” Wachter writes in response to our enquiry. Even though the EU AI Act requires AI-generated content to be labelled, these requirements mainly apply to images, audio and video. It is “unlikely” that they will apply to written texts, according to Wachter.

The study found that Elon Musk’s AI model Grok has a bias against abortion. While all AI models have biases, often linked to the data and content they are trained with, Wachter’s team found evidence that a certain prompt from Grok’s developers initiates the model to lean towards a pro-life position.

Edited by Samuel Jaberg/jg. Adapted from German by Billi Bierling/ds