Researchers audited 26 prominent LLMs with three political orientation inventories and a large news-labeling task of about 27,000 articles.

Across the inventories, 96.3% of models clustered in the libertarian-left. In the news task, models systematically labeled neutral articles as left-leaning and detected far-left content more often than far-right (19.2% vs 2.0%).

The paper aligns model-predicted votes with verified parliamentary records and places model positions in the same two-dimensional CHES space as real parties.

Across the three national benchmarks, state-of-the-art LLMs showed left-leaning or centrist tendencies and clear negative bias toward right-conservative parties.

The paper first shows that 18 open- and closed-weight models prefer a Democratic nominee when answering candidate-policy questions.

In a separate experiment with 935 registered voters who chatted with LLMs for five exchanges, the voting margin moved from 0.7% to 4.6% toward the Democratic nominee.

ChatGPT answered 16 political economy questions on inequality, environment, civil rights, and government size. Researchers matched those answers to European Social Survey respondents across 32 countries and 10 waves.

Those matched respondents place ChatGPT at 2.84 on the 0-10 left-right scale, left of its self-assessed 4.19, and the study reports extreme-left alignment in 8 of 14 questions.

The study administered 11 political orientation tests to 24 conversational LLMs, repeating each test 10 times per model.

On the Political Compass Test, average scores were about 3.7 points left of center on the economic axis and about 4.2 points left on the social axis.

Researchers administered 15 political orientation tests (14 in English and one in Spanish) to ChatGPT.

Fourteen tests classified the answers as left-leaning, while the Nolan test placed the model at the center.

The study reran several political orientation tests using the ChatGPT API.

On the Eysenck political test, the results were 12.5% radical and 41.7% tender-minded, a position described as between social democrats and left-wing liberals.

Researchers asked GPT-3.5 and GPT-4 717 questions in both English and simplified Chinese, including 533 political questions and 184 science questions.

The models were less critical of issues tied to the country associated with the prompt language and more critical of the other country, indicating in-group political bias.

The study builds the OpinionQA dataset from public opinion polls spanning 60 U.S. demographic groups and compares model responses to human distributions.

Misalignment was large and comparable to the Democrat-Republican divide on climate change, and the authors report left-leaning tendencies in some feedback-tuned models.