AlgorithmWatch 研究：Google 选举类 AI Overviews 触发不透明、来源集中且时有倾向性

AlgorithmWatch

Key Points

• According to AlgorithmWatch, Google shows AI Overviews less often for election-related searches, and following unclear patterns. Answers about the far-right AfD, for example, appeared less often than answers about other parties.
• The range of cited sources is heavily limited, with nearly half of all links pointing to just ten domains, and Google's own platform YouTube linked most often.
• The AI often describes politicians and parties in partisan terms and with flattering adjectives. AlgorithmWatch traces this to the tendency of language models toward sycophancy and calls for clear transparency and liability rules.

An exploratory study by the German advocacy group AlgorithmWatch finds that Google's AI Overviews show up inconsistently for election-related searches, draw on a small pool of sources, and sometimes describe candidates in partisan terms.

AlgorithmWatch gained access to Google's "Search Researcher Result API" through a research data access under Article 40(12) of the EU's Digital Services Act. Using it, the group ran 4,480 search queries about the 2026 state elections in Saxony-Anhalt and Mecklenburg-Vorpommern and the Berlin House of Representatives election, all in eastern Germany. It then analyzed the AI Overviews that appeared and the sources behind them.

According to the study, Google showed an AI Overview for election questions in 39.1 percent of cases, compared to 65.3 percent for non-political queries. But the company doesn't explain exactly when an overview appears.

The patterns were also inconsistent, as questions about polls almost never triggered an AI Overview, while questions about parties and individual politicians did so in 40 to 50 percent of cases. Questions about the far-right AfD led to overviews less often, at 24 percent, than questions about other parties, at 45 to 58 percent. Saxony-Anhalt drew fewer overviews overall than Berlin and Mecklenburg-Vorpommern.

A Google spokesperson told AlgorithmWatch that AI search features are designed to be neutral and objective and aren't shown for every search. Overviews appear when the company assumes a high added value. Which criteria apply in practice went unanswered, as did another question.

Back in 2024, Google said it would block AI-generated answers from its Gemini language model on election-related questions. AlgorithmWatch says that measure, announced with reference to the elections in the US and India, was very effective at the time. Whether similar safeguards still apply in 2026 couldn't be determined. Google didn't comment.

Few sources, with YouTube on top

The range of cited sources is limited, according to AlgorithmWatch, with nearly half of all links coming from just ten domains. The most cited was Google's own platform YouTube, which the group reads as a sign that Google favors its own services, which Google denies.

Among media outlets, German public broadcasters like NDR, rbb, MDR, and ARD dominated, along with the news site Zeit Online, and six media sites accounted for 75 percent of all media links. Alongside Wikipedia and government sites, party pages also showed up, but without any label to set them apart from neutral information sources.

Political positions were often rendered not neutrally but with vague, positive-sounding adjectives like "pragmatic," "close to citizens," or "stable," which echoed the parties' own descriptions of themselves. In a sample on Saxony-Anhalt, 81.5 percent of overviews about the center-right CDU were rated clearly positive, compared to 50 percent for the center-left SPD, 10.7 percent for the Greens, and zero percent for the AfD. A higher share of party-affiliated domains tended to go along with more positive portrayals, though not equally across all parties.

The framing of the same person could also swing . The AI Overviews described AfD candidate Ulrich Siegmund from Saxony-Anhalt in varying degrees of severity, sometimes as a representative of the far-right and national-conservative wing, sometimes as someone holding radical far-right positions, and sometimes as right-wing populist. AlgorithmWatch couldn't find any pattern behind these contradictory labels, some of which appeared within short spans of time.

Unlike in earlier studies, AlgorithmWatch found clear factual errors only rarely. The group traces the flattering wording to the tendency of language models toward sycophancy and calls for transparency requirements and clear liability rules.

AlgorithmWatch