AI-Policy

Google shows AI answers less frequently in elections

3 min read

TL;DR Too Long; Didn’t read

AlgorithmWatch analyzed 4,480 Google search queries tied to the state elections in Saxony-Anhalt, Berlin, and Mecklenburg-Vorpommern: AI overviews appeared for election-related queries in only 39 percent of cases, versus 65 percent for non-political topics. Nearly half of the linked sources came from just ten providers, YouTube most often. The organization also flags language in the summaries that often sounds favorable and party-aligned.

A Google logo sticker is affixed to a ballot paper, from which speech bubbles of varying size grow over two ballot boxes – one overflowing, the other nearly empty. Image generated with GPT Image 2

Key takeaways

  • Data access ran through Article 40(12) of the Digital Services Act and Google's Search Researcher interface.
  • For candidate-related queries, Google displayed an AI overview in 40 to 50 percent of cases.
  • About half of the linked pages came from just ten domains, with YouTube ranking first.
  • Descriptions such as “pragmatic” or “citizen-friendly” read more like party self-promotion than journalism.
  • Google names no verifiable display criteria and denies favoring YouTube.
  • Saxony-Anhalt already voted on September 6; Berlin and Mecklenburg-Vorpommern follow on the 20th.

The organization AlgorithmWatch has, for the first time, systematically examined how Google deploys its AI overviews around Germany’s three state elections. Based on 4,480 search queries, the study finds that election-related questions get an AI summary far less often than other topics, and that the sources drawn on stay strikingly narrow. Saxony-Anhalt has already voted as the first of the three states, with Berlin and Mecklenburg-Vorpommern following on September 20.

Google displays AI answers less often for election queries

The civil-society organization AlgorithmWatch ran an automated evaluation of 4,480 Google search queries tied to the state elections in Saxony-Anhalt, Berlin, and Mecklenburg-Vorpommern in July 2026. The basis was access granted under Article 40(12) of the Digital Services Act, which lets registered research institutions inspect platform data from very large online search engines. Google’s own “Search Researcher Result” interface made it possible to track when and how often AI overviews appear. AlgorithmWatch is among the research bodies approved under the Digital Services Act to request this kind of data access from very large online platforms.

The central finding – unverified by independent sources – is this: for election-related search queries, Google displayed an AI overview in only 39.1 percent of cases, compared with 65.3 percent for non-political comparison queries. Questions about individual candidates got an automatic summary in 40 to 50 percent of cases, with the rate swinging noticeably between the three states. According to the study, the company does not disclose the criteria behind when an overview gets displayed.

Sources cluster around a handful of large providers

The study also flags where the linked sources come from: nearly half of all links inside the AI overviews led to offerings from just ten different providers. YouTube was linked most often – a service that, like Google Search itself, belongs to parent company Alphabet. AlgorithmWatch reads this as a possible sign of favoritism toward in-house properties, though it stresses that this alone does not amount to a rule violation. The other frequently linked sites were mostly established media outlets and reference works, with few smaller or local newsrooms among them.

The study’s authors also flagged the language used in the summaries: parties and candidates were often described with favorable, vague terms such as “pragmatic,” “citizen-friendly,” or “stable” – phrasing that reads more like a party’s own self-description than journalistic framing. One example from the study describes a Berlin SPD politician as “pragmatic, citizen-friendly, and results-oriented.” Project lead Clara Helming puts the criticism plainly: “Google’s product decisions shape important preconditions for opinion formation.”

Google points to its own quality criteria

Google told the study’s authors that AI overviews are only shown when the company judges them to add value, and that they are always designed to be “neutral and objective.” It gave no verifiable criteria for that judgment, and it denies favoring YouTube over other sources. As a very large online search engine under the Digital Services Act, Google also carries special due-diligence duties once a service could pose systemic risks to democratic processes. Neither Germany’s Federal Network Agency, the country’s coordinator for digital services, nor the European Commission has announced any proceedings over the findings so far.

The case joins a string of regulatory clashes over Google’s AI search features. Back in July 2026, Germany’s media authority ZAK had already ruled to classify Google’s AI search as an independent content provider, stripping it of the liability privilege for automatically generated search answers. The AI Act’s transparency obligations, in force EU-wide since August 2026, cover recognizable AI chatbots and deepfakes but not the display logic of search overviews – a gap AlgorithmWatch’s new study points to directly.

What will matter is whether the European Commission comes to treat the lack of clear display criteria for AI overviews as a systemic risk under the Digital Services Act – oversight so far has focused mainly on advertising and recommender systems. It also remains open whether Google adjusts anything before the Berlin and Mecklenburg-Vorpommern elections on September 20. For Saxony-Anhalt, any response would come too late regardless: voters there have already cast their ballots.

Frequently asked questions

Since when has the data-access mechanism behind the AlgorithmWatch study applied?

The Digital Services Act has let registered research bodies request data from very large platforms since it took effect; AlgorithmWatch used Google's search-data interface for this in July 2026.

Does the criticism apply only to Google, or to other search engines too?

The study covers only Google, since the company qualifies as a very large online search engine under the Digital Services Act. Other providers were not part of the sample.

What happens if a confirmed violation of the Digital Services Act is found?

For proven systemic risks, the Digital Services Act allows fines of up to six percent of global annual revenue. Whether the Commission opens proceedings here remains open.

How does the situation in Germany compare with other countries?

AlgorithmWatch Switzerland, together with the University of Zurich, published a similar study in August 2026 and likewise found neutrality problems in Google's AI overviews.

What happens with the still-pending elections in Berlin and Mecklenburg-Vorpommern?

AlgorithmWatch says it will keep monitoring Google's AI overviews through the September 20 elections; results are not yet available.

Sources (5)
  1. AlgorithmWatch: Google AI overviews and opinion formation in state elections
  2. t3n: 4,500 search queries analyzed – what stands out in Google's AI overviews for state elections
  3. netzpolitik.org: AI results at Google – information roulette in state elections
  4. Digital Services Act – Regulation (EU) 2022/2065 (EUR-Lex)
  5. swissinfo.ch: How Google's AI overviews influence democracy in Switzerland

Your AI update for the work week

Once a week, the most important AI news – plus one practical tip to try right away. No spam, unsubscribe anytime.

← Back to the blog