The first paper session at the SEASON 2026 conference in Hamburg that I’m attending is on search engines and political information, and starts with Íris Damião, whose focus is on political bias in search engines and Large Language Models. Search engines have evolved a lot since they emerged in the 1990s, of course, but continue to be very widely trusted by their users – even when searching for some very sensitive information.
This is the case for instance when searching for election information, where people also search for information on election procedures, party programmes, and even whom to vote for; some of this has now also shifted to conversational AI chatbot interactions. But how do search and AI algorithms select the information they provide? What is the impact of any visibility biases (which may also reflect media coverage biases, of course) in such search results on voting intentions and outcomes? Do search engines and LLMs prioritise content of a specific political leaning?
This study focussed on two elections; the 2024 EU parliament election and the 2024 US presidential election. These represent two very different political systems, of course: a massively multi-party, transnational system and a highly polarised, two-party national environment. The study approached this by using the OpenWPM Webcrawler, which automates the Firefox browser, to make it appear like a human users; using proxy systems, it made many such bots appear as if they came from five different EU countries or from various US counties that leaned left, centre, or right.
These bots would simultaneously query several search engines and AI chatbots with a range of neutral political questions (in the EU) and questions about candidates, policies, and voting suggestions (in the US). Search results were then saved, and headlines and URLs gathered; chatbot responses were filtered for political parties, issues, and leanings. In the EU, results were grouped by the major European Parliament party groupings; in the US, by Republican- and Democrat-leaning issues.
In the EU election, Google and Bing rarely mentioned specific political entities in the results (12% and 7%, respectively), but such mentions focussed prominently on the radical right; chatbots CoPilot and ChatGPT mentioned specific entities much more often (62% and 41%), and the radical right again leads these mentions. In the US, search engines referred to political entities much more (59% and 36%), but here Google mentioned Republican-aligned issues much more, while Bing focussed more on Democrat-leaning issues. Chatbots named entities in 100% of responses, with a reasonable balance across both sides of politics, though strongly Democrat-leaning issues were rare.
So, search engine results and LLM responses do not represent all sides of politics evenly – but should this be the aim in the first place? These leanings may be also influenced by past election results, polling intentions, media coverage, and other contextual factors that influence the attention paid by search engines and chatbot discussions; these factors can serve as external benchmarks to compare search and chat results against. For media coverage, this benchmark was established via MediaCloud.
Such comparisons show that in search, the EU radical right remains highly overrepresented; in chatbot results, the balance is more aligned with these benchmarks. In the US, there remains a slight overrepresentation of moderately Republican-aligned issues at least in Google. These results point to a need for more independent oversight.
