AI systems cite state-controlled media more than traffic-matched commercial outlets when prompted in a country's own language

state-mediaai-searchgenerative-aimedia-concentration

Using the AI Search Arena dataset's non-English prompts (roughly 40% of its prompt set), Carragher and Carley (arXiv:2610.40164) restricted their comparison to a specific slice: users located in a given country, prompting in that country's own language, about that country's own outlets. Within that slice, state-controlled media outlets were cited 1.65x [95% CI 1.46–1.87] more often than commercial outlets matched to the same traffic share — in absolute terms, 1.01x their traffic share for state outlets versus 0.61x for the matched commercial group, across 2,961 citations spanning 16 countries. The effect is attributed primarily to Russian- and Mandarin-language prompts, which drew markedly more citations to Russian and Chinese state media respectively than English-language prompts about the same outlets did (5.3x and 2.5x more, per the paper's Table 8). The same pattern did not appear in the paper's separate English-only Ahrefs panel, which the authors attribute to that panel's English-only prompt construction rather than treating it as evidence against the effect.

This result is presented as building directly on Waight, Yang, Yuan, Messing, Roberts, Stewart & Tucker's 2026 Nature finding that large language models "exhibit pro-government valences when prompted in the languages of countries with lower media freedom" — the present paper's native-language test is this vault's first encounter with that finding, read only as Carragher and Carley cite it, not independently verified here. The authors themselves caveat that the Arena dataset is underpowered for some sub-tests and that owner-level statistical claims specifically are "difficult to defend" on it; the 1.65x headline figure rests on the better-powered 16-country/2,961-citation cut rather than the thinnest slice.