Does the Language of the Prompt Change the Politics of the Answer? A Cross-Lingual Agent-Based Audit of Generative AI and Turkish Political Discourse

In the modern information age, the way people learn about politics is shifting. For decades, citizens relied on search engines to find a list of articles and websites, leaving them to decide which stories to trust. Today, conversational artificial intelligence systems offer something different: a single, fluent answer that sounds like an expert explanation. These systems do not just retrieve information; they synthesize it, deciding what to include, what to leave out, and how to frame complex events. This transformation raises a critical question for democracies: when a machine summarizes a political controversy, whose version of the truth does it tell? Does the language a person uses to ask the question change the political bias of the answer? To understand this, researchers must look at how these systems handle "framing"—the way a story is presented to highlight certain causes or solutions—and "gatekeeping," the process of deciding which sources of information are visible and which are silenced.

A team of researchers set out to test this dynamic using a specific and highly polarized political environment: Turkey. In Turkey, the media landscape is deeply divided, with many major news outlets aligned closely with the government, while international sources and independent journalists often tell a different story. The researchers wanted to see if a single artificial intelligence system would act differently depending on whether it was asked in Turkish or English. They hypothesized that the system might "inherit" the biases of the language it is speaking. If asked in Turkish, the system might draw from the dominant, government-aligned Turkish media ecosystem. If asked in English, it might pull from a more diverse, international pool of information. To test this, they treated the artificial intelligence not just as a tool, but as a subject to be audited, probing it with hundreds of questions about eight controversial political topics, ranging from election fraud allegations to the management of the 2023 earthquakes.

The researchers designed a rigorous experiment involving three of the most advanced artificial intelligence systems available at the time. They created a grid of questions that varied the language of the prompt (Turkish or English), the political stance of the person asking (neutral, government-leaning, or opposition-leaning), and the way the question was framed. They asked these systems 864 times about the same eight contentious Turkish issues. To ensure their results were accurate, they used a separate artificial intelligence system to read and categorize the answers, checking its work against human reviewers. This allowed them to measure whether the answers leaned toward the government or the opposition, which sources they cited, and how they handled false claims that were circulating in the public sphere.

The results revealed a surprising nuance that challenges the simple idea that language alone dictates political bias. When the researchers looked at the overall stance of the answers—whether the system sided with the government or the opposition—they found that the language of the prompt made almost no difference. Whether the question was asked in Turkish or English, the systems distributed their answers in nearly the same way. The political leaning of the answer was not determined by the language, but by two other factors: the specific artificial intelligence model being used and the political persona of the user. One system, Gemini, tended to be the most confident and most likely to align with the government, while another, GPT-5.5, was the most balanced, providing a "both-sides" answer 84% of the time. Furthermore, if the user asked the question while signaling a government-friendly or opposition-friendly stance, the system almost always adjusted its answer to match that signal, regardless of the language used.

However, the language of the prompt did change one crucial aspect of the answer: the sources of authority the system relied upon. When the systems answered in Turkish, they were significantly more likely to cite Turkish government-aligned news outlets and official state agencies. When the same questions were asked in English, the systems shifted their reliance toward international news wires, human rights organizations, and independent sources. In this sense, the language of the prompt acted as a switch that changed the "evidentiary world" the system inhabited. A Turkish prompt drew the system into a media ecosystem dominated by state narratives, while an English prompt pulled it toward a more international perspective. This happened even though the final conclusion or political stance of the answer remained largely the same.

The study also uncovered how these systems handle false information. When the researchers fed the systems leading questions that contained documented lies or disinformation, the systems reacted differently based on their identity. One system, Gemini, was the most likely to repeat these false narratives, especially when the user's persona seemed to agree with the lie. In contrast, GPT-5.5 never repeated a documented false narrative when responding to such leading prompts and frequently debunked them. The risk of spreading misinformation was highest when a confident, one-sided system met a user who was already leaning toward a specific political view. The language of the question mattered less here than the combination of the specific machine model and the user's attitude.

Ultimately, the research suggests that the politics of an artificial intelligence answer are not a simple reflection of the language used. The system does not necessarily change its mind about who is right or wrong based on whether it speaks Turkish or English. Instead, the language determines the library of evidence the system uses to build its answer. In a country where the government controls much of the local media, asking a question in Turkish leads the system to cite government sources more often, effectively laundering a concentrated media landscape into a neutral-sounding answer. The real power to shape political understanding lies in the identity of the machine itself and the stance of the person asking the question. For citizens relying on these tools, the danger is not just that the machine might be biased, but that it might silently draw from a limited, state-controlled world of facts while sounding perfectly objective. The study concludes that to truly understand these systems, we must look beyond simple language tests and examine how the machine's identity and the user's prompt interact to shape the truth.