Researchers from Karlstad University explored how generative AI search engines, such as Microsoft Copilot, are transforming political news retrieval - and with it, the global flow of information.
The study "Sourcing behavior and the role of news media in AI-powered search engines in the digital media ecosystem: Comparing political news retrieval across five languages" was published in the journal Telecommunications Policy and is co-authored by Professor Cornelia Brantner, Professor Michael Karlsson, and Joanne Kuai.
Focusing on the 2024 Taiwan presidential election, the research analyzed how Microsoft's generative AI-powered Copilot responded to political news prompts in five languages - English, Traditional Chinese, Simplified Chinese, German, and Swedish. While the chatbot efficiently synthesized election information and favored professional journalistic sources, the study found critical issues with misinformation, overreliance on Anglophone sources, and lack of cultural representation.
– Our findings show that although AI search engines promise enhanced access to news, they risk marginalizing regional perspectives and replicating systemic biases, says lead author Cornelia Brantner. Moreover, almost 40 procent of citations to professional news media were either misattributed, included factual errors, or both.
The research also reveals that Copilot disproportionately draws from UK and US-based news media - such as The Economist and BBC - even when queried in other languages and from non-Anglophone locations. This sourcing pattern raises concerns about the homogenization of political discourse and underrepresentation of local voices in democratic processes.
The study raises red flags regarding source transparency and factual accuracy. For example, 20 procent of citations to professional news outlets included factual errors due to misrepresentation by the AI system itself - errors that could have serious implications in sensitive political contexts.
The Swedish case was particularly striking. Despite all prompting being conducted from a Swedish IP address, Copilot’s responses in Swedish relied heavily on English-language sources, with limited inclusion of professional Swedish media. Most notably, none of Sweden’s major commercial news outlets appeared in the chatbot’s responses, despite their documented coverage of the election. The researchers attribute this absence to Swedish publishers’ choice to opt out of AI crawling via robots.txt, a platform counterbalancing strategy aimed at protecting their content from being used without compensation.
– The result is that Swedish users are primarily served international or generic content - even in their native language, says Professor Michael Karlsson. This raises important questions about media visibility, democratic relevance, and the potential cost of opting out of AI ecosystems.
Instead, Copilot frequently referenced Swedish Wikipedia, foreign think tanks, and US-based policy organizations - sources that, while accessible, do not reflect Sweden’s media landscape or editorial values. The findings suggest that news organizations in smaller media markets may face an uneasy trade-off between autonomy and visibility in the age of AI.
– Search engines powered by large language models are now acting as algorithmic gatekeepers, says Professor Karlsson. This has implications for the autonomy of journalism, especially in smaller or paywalled markets like Sweden, where visibility is at risk if publishers opt out of being crawled by AI.
The authors urge policymakers and technology developers to implement stronger safeguards for generative AI, including requirements for accurate source attribution and curbing the creation and spread of misinformation, but also measures to support journalism and the inclusion of diverse, regionally grounded sources.
– AI chatbots should not become a substitute for quality journalism, says Joanne Kuai. We need systems that amplify - not erase - media pluralism in a digital world.
