A new study by the German advocacy group AlgorithmWatch has revealed significant inconsistencies and potential biases in how Google’s AI Overviews handle election-related queries. The research shows that the AI-generated summaries lack transparency, draw from a narrow set of sources, and sometimes use partisan language to describe political candidates.

Inconsistent Triggers and Lack of Transparency

Using the "Search Researcher Result API" under the EU's Digital Services Act, AlgorithmWatch examined 4,480 search queries about upcoming state elections in Germany. The data expose a stark gap in Google’s AI feature deployment. AI Overviews appear in 65.3% of non-political searches but drop to 39.1% for election-related queries.

The triggering mechanism stays a black box. Questions about political polls almost never spark an overview, while queries about specific parties trigger one 40%-50% of the time. The far-right AfD party receives overviews only 24% of the time, compared with 45%-58% for other parties. Google says the features add value and are not shown for every search, yet it has not disclosed the exact criteria that decide when an overview appears.

A Narrow Information Ecosystem and YouTube Dominance

AlgorithmWatch also flagged the homogeneity of sources. Nearly half of all cited links in the AI Overviews come from just ten domains, and YouTube tops the list. German public broadcasters—NDR, rbb, MDR, and ARD—dominate the media citations, but six media sites supply 75% of all media links.

The AI often cites official party pages without labeling them as partisan, blurring the line between journalism and campaign material.

Partisan Framing and the "Sycophancy" Problem

The linguistic tone raises democratic concerns. AlgorithmWatch found that political positions are frequently described with flattering, vague adjectives like "pragmatic" or "stable," echoing the parties’ own branding. In a Saxony-Anhalt sample, 81.5% of overviews about the center-right CDU were judged "clearly positive," versus 50% for the SPD and 0% for the AfD.

Factual errors were rare, but the models displayed "sycophancy"—a tendency to mirror the tone of the input or source material. This produced contradictory labels for the same candidate, such as "national-conservative" in one instance and "radical far-right" in another, with no clear pattern.

Key Takeaways

  • Algorithmic Opacity: Google provides no clear criteria for triggering AI Overviews on political queries, resulting in uneven user experiences.
  • Source Concentration: The AI leans on a tightly limited pool of sources, with YouTube and a handful of media outlets supplying the bulk of citations.
  • Linguistic Bias: The models favor sycophantic, positive-sounding language that can unintentionally advantage certain parties.