Google's election AI Overviews are opaque, rely on few sources, and sometimes take sides
decoder ·
AlgorithmWatch, a German advocacy group, used research access under the EU's Digital Services Act to analyze Google's AI Overviews for election-related searches in three eastern German states. They found that Google displayed AI Overviews much less often for political queries (39.1%) than for non-political ones (65.3%), with inconsistent patterns and particularly fewer overviews for questions about the far-right AfD party. The study also revealed that the sources cited were heavily concentrated, with nearly half of all links coming from just ten domains and Google's own YouTube being the most cited source. Politicians and parties were often described in positive, flattering terms that mirrored their own self-descriptions, and the tone varied by party, with the CDU receiving clearly positive coverage in over 80% of overviews while the AfD received none. The AI also assigned contradictory political labels to the same candidate without any discernible pattern. Although clear factual errors were rare, AlgorithmWatch attributes the biased wording to language models' tendency toward sycophancy and calls for transparency and liability rules.
decoder