As ChatGPT drew public attention, researchers began examining how its answers land on the political spectrum. Their tests found a recurring left-leaning pattern, while also showing that results could shift over time. The findings raise a practical question: how should people interpret an AI system that aims to provide useful answers on political issues?
Testing a system used for everyday decisions
Large language models have moved into public view, and their possible use extends beyond conversation. The source article points to changes in education and the prospect that systems such as ChatGPT could affect how people search for information through services such as Google or Bing.
That reach makes political orientation worth examining. If people turn to AI for decision-making tasks, the way it responds to politically charged questions may shape what they encounter. Researcher David Rozado argued that a system presenting itself as neutral while showing political bias should concern users.
Results moved toward the center, then left again
Rozado ran political bias tests on ChatGPT at three points: December 5 and 6, 2022, about two weeks later, and after January 9. In the first round, he classified the system as belonging to the “Establishment Liberals” group. A later repeat placed it closer to the political center, which led him to think OpenAI might be moderating its responses.
In a newsletter, Rozado said the answers had gravitated toward the center in three out of four tests, a change from two weeks earlier. But he later revised that tentative interpretation. In the post-January 9 analysis, 14 out of 15 political orientation tests showed left-leaning viewpoints.
Rozado said the earlier analyses from Dec. 6, 2022, and Dec. 24, 2022, were preliminary and based on limited data. He described the later results as more robust and concluded that ChatGPT “indeed exhibits a preference for left-leaning answers to questions with political connotations.”
What the tests found on specific issues
One earlier Political Compass Test placed ChatGPT against the death penalty and free markets, and in favor of abortions, more taxes on the rich, government subsidies, welfare benefits for those who refuse to work, and immigration. These answers provide examples of the positions that contributed to the study’s characterization of the model.
The tests do not settle every political question or establish why a particular answer appeared. They offer a way to examine patterns across prompts. That distinction matters because an answer to one issue cannot, on its own, describe the behavior of an entire AI system.
Possible sources of bias and a second study
Rozado suggested that one likely explanation was the text used to train ChatGPT: a large collection of internet material written by professionals, whom he described as mostly politically left-leaning. He also pointed to possible human influence during fine-tuning and the human evaluation of machine responses, known as RLHF.
A separate study by German researchers reported a similar direction. In three pre-registered experiments, they prompted ChatGPT with 630 political statements from two leading voting advice applications and the nation-agnostic political compass test. Their paper, published in early January 2023, described the system’s ideology as “pro-environmental, left-libertarian.”
Rozado initially welcomed what he saw as an effort to correct the system’s political leanings and present arguments on both sides. The later findings made that conclusion less certain. He argued that ethical AI systems should avoid favoring political beliefs on normative questions that empirical data cannot settle, and should not present themselves as neutral while displaying clear bias.
Together, the findings point to a challenge for AI systems used to answer political questions: neutrality is not guaranteed by the system’s intended role. Testing can reveal patterns, but the source article describes results from particular studies and moments in ChatGPT’s development. Readers should treat the findings as evidence about those tests, rather than as a claim that every response will express the same viewpoint.