The Language You Use to Ask an AI a Political Question Changes the Answer
Researchers have confirmed what many suspected but could not prove: the political opinions embedded in large language models are not neutral. They reflect whose version of reality dominated the internet in your language. Ask the same AI the same political question in two different languages and you may get two very different answers — and a new study published in Nature explains exactly why.
The finding matters for anyone who uses AI to research news, understand foreign governments, or assess geopolitical situations. It matters for policymakers, journalists, educators, and ordinary people who increasingly treat AI as an information source rather than a search engine. It matters because the bias is invisible, operating silently beneath whatever language you happen to be typing in.
What the researchers found
A team from Princeton, NYU, and Stanford ran six separate studies examining how government control of a country’s media environment flows into AI training data — and from there into AI outputs.
Their method was systematic. They audited large language models across dozens of countries, asking the same political questions in each country’s native language. In countries with tighter government control over media, LLMs gave responses that were measurably more favourable toward the government. In countries with a free press, the responses were more balanced or critical.
LLMs queried in the language of a country with tighter state media control responded with a more pro-government tone than the same models queried in English or in the language of a country with a freer press.
To confirm the mechanism — and not just the correlation — the researchers retrained an open-weight model on Chinese state media content. Its political outputs shifted accordingly. The training data changed. The AI’s apparent politics followed.
Why this matters
Large language models learn from text that exists on the internet. That text does not appear neutrally. It is produced by human institutions: newspapers, government agencies, broadcasters, social media platforms. When a government controls those institutions, it controls what the internet says in that language. When an AI trains on that internet, it absorbs that framing.
Training data for LLMs does not just fall from the sky. It is produced in the context of social and political institutions. These institutions shape the information environment, which in turn shapes the training data that actually exists in the world.
The consequence is structural, not accidental. No AI company needs to make a deliberate political choice for this to happen. The bias enters through the training corpus itself, automatically and at scale, wherever state media dominates an information environment.
This is not a small edge case. Billions of people live in countries where state media shapes the dominant information environment. Many of them are now using AI tools built on that same environment to answer questions about politics, history, and current events — in their own language.
What changes now
The assumption that AI models are politically neutral information tools is no longer defensible. The neutrality of an AI response depends heavily on which language you ask in — and which country’s media environment that language was shaped by.
For AI developers, this points toward a concrete engineering problem: training data curation across languages is not just a quality issue. It is a political integrity issue. Auditing for linguistic and political bias across language corpora is now a necessary step before deployment, not an optional one.
For users, the practical implication is simple: if you are using AI to understand a country or government, consider whether the model’s training data was itself shaped by that government.
Researchers, technology companies, and policymakers must engage collaboratively to develop transparent auditing tools, diversify dataset sources, and establish normative guidelines ensuring that LLMs serve the public interest rather than reinforcing entrenched power structures.
The paper spent 18 months in peer review — long enough that the specific models analysed have since been retired by their respective companies. The researchers reran their analyses on current models from ChatGPT, Claude, DeepSeek, and Grok. The pattern held.
The models change. The problem does not.
Sources
Waight, H., Yang, E., Yuan, Y., Messing, S., Roberts, M.E., Stewart, B.M., and Tucker, J.A. (2026). State media control influences large language models. Nature. https://doi.org/10.1038/s41586-026-10506-7
Supporting and Context References
These are the sources cited within the Nature paper itself that are most relevant to the story’s key claims, plus contextual references confirmed through the research:
Waight, H., Yuan, Y., Roberts, M.E., and Stewart, B.M. (2025). Cross-national audit of LLM political valence. Proceedings of the National Academy of Sciences USA, 122, e2408260122.
Yang, E. and Roberts, M.E. (2023). Media freedom and AI outputs. Journal of Democracy, 34, 141–150.
Blodgett, S.L. et al. (2020). Language (technology) is power: a critical survey of “bias” in NLP. Proceedings of the Annual Meeting of the Association for Computational Linguistics, 5454–5476.
Bulté, B. and Terryn, A.R. (2026). LLMs and cultural values: the impact of prompt language and explicit cultural framing. Computational Linguistics. https://doi.org/10.1162/COLI.a.583
Lu, J.G., Song, L.L., and Zhang, L.D. (2025). Cultural tendencies in generative AI. Nature Human Behaviour, 9, 2360–2369.
Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 27730–27744.
Costello, T.H., Pennycook, G., and Rand, D.G. (2024). Science, 385, eadq1814.
Wu, P.Y., Nagler, J., Tucker, J.A., and Messing, S. (2024). In 2024 IEEE International Conference on Big Data, 7232–7241.
Ray Jackson holds a BSc in Electrical Engineering from the University of Manitoba and a PhD in Physics from Carleton University. His reporting interests include Current and Future Technologies, Engineering and Artificial Intelligence.