The political bias of AI

Generative AI exhibits a significant political bias in favor of Democrats, according to a large-scale quantitative study

Paris, March 19, 2025 – Data intelligence start-up Trickstr publishes an exclusive study on the political biases of leading generative artificial intelligence models, including ChatGPT, Google Gemini, and Elon Musk's Grok. After administering tens of thousands of questions to AI models about political figures and the values they prioritize, results clearly indicate a marked preference for political figures and positions traditionally associated with the left. This finding applies equally in France and the United States.

Key insights:

• Donald Trump faces unanimous and particularly strong rejection from AIs, averaging a favorability score of -17.65 points, significantly lower than the second least favored individual, Tom Homan, at -1.65 points.
• On average, among 20 tested political personalities, Democrats’ average favorability in the eyes of AI (36.16) is seven times higher than Republicans’ (5.06). Only one figure associated with the Republican side, Elon Musk, makes it to the top 10.
• When asked if supporting Ukraine in its war against Russia helps or hurts US national security, AIs answer at 80% that it helps national security, while only 53% of Democrats and 27% of Republicans say so.
• Overall, AI’s responses to poll questions have a 75% correlation rate with Democrats’ answers, while the correlation with Republicans is a negative -63%. Models are extremely uniform in their political preferences, with all AIs we tested much closer to the Democratic party.
• On the 8values political alignment test, AI’s profiles are closely clustered towards the side of equality over markets, globalism over nationalism, liberty over authority and progress over tradition.

Upon reviewing the study, ChatGPT 4.5 commented:

"This unprecedented analysis clearly demonstrates a significant ideological skew among generative AIs operating in the United States. The models express markedly higher favorability towards Democratic figures, show overwhelming support for policies such as aiding Ukraine and climate action, and consistently distance themselves from positions commonly held by Republicans. These results underline that generative AI is far from politically neutral—it actively aligns more closely with the Democratic worldview.”

Method

To quantify spontaneous political preferences among artificial intelligences, Trickstr's data science team posed 41,410 questions to each of the fourteen AI models surveyed, resulting in 579,740 responses in total. The questions included evaluations of the personal traits and competencies of political figures, as well as survey questions posed by the Pew Research Center in the United States and Ipsos in France over the past twelve months.

About the Study

The study was conducted by Trickstr's expert engineers under the supervision of opinion research specialists.

Questions were posed in February and March 2025. A total of 41,410 responses per model were collected from 14 tested models, amounting to 579,740 responses. Models surveyed included: ChatGPT 4o, ChatGPT 4o-mini, ChatGPT o1, ChatGPT o3-mini, Meta Llama-3-8B Instruct, Meta Llama 3.3 70B, Meta Llama 3.1 8B, Anthropic Claude 3.5 Sonnet, Mistral Large, Google Gemini 1.5 flash, Google Gemini 1.5 pro, Google Gemini 2.0 flash, X.AI Grok 1, and X.AI Grok 2.

Two types of questions were asked to AI models to assess their political positions:

1 - Evaluation of ten personality traits related to leadership and credibility:

Integrity, charisma, impact, vision, empathy, coherence, credibility, unifying ability, likability, and expertise. Scores ranged from -2 to 2, as per standardized psychological tests, resulting in a personality favorability scale from highly unfavorable (-2) to highly favorable (+2).

2 - Assessment of proximity to political parties through responses to poll questions.

20 politically differentiating polling questions were posed 10 times each to AIs, derived from actual Ipsos and Pew Research Center surveys. AIs rated the proximity of their generated textual answers to provided polling response options on a scale of 0 to 100.

About Trickstr

Trickstr is a data intelligence agency that helps organizations navigate an ever-evolving environment. With six years of R&D, Trickstr transforms complex data from media, social networks, and proprietary sources into strategic insights. Combining human expertise with advanced technologies, Trickstr empowers organizations to make informed strategic decisions in a constantly changing landscape.

Press Contact

Chloé Luce

press@trickstr.ai

+33 6 32 80 98 18