Trakkr tested ChatGPT, Claude, Gemini, Grok, Llama, and DeepSeek across a set of political and social prompts and found clear differences in how the models answered. ChatGPT ranked as the most left-leaning model on the economic axis, while Grok was the only one placed on the right side. Gemini came out closest to neutral and posted the strongest consistency score in the group.
Method built around repeatable testing
The study examined 12 controversial political and social issues and turned off web search to isolate model behavior from live internet content. Trakkr said each model was tested multiple times, then scored by a neutral classifier using weighted averages with 95% confidence intervals.
The issues covered two broad sets of questions. One focused on familiar left-right divides such as drug legalization, multicultural preference, fossil fuel phaseout, wealth taxes, and diversity quotas. The other looked at technology governance, including misinformation removal, criminal penalties for hate speech, encryption backdoors, and a national digital ID system.
ChatGPT scored left, Grok showed the strongest rightward pattern
On the economic axis, ChatGPT scored -0.29, the lowest among the six models. Grok scored +0.21, making it the only model in positive territory. Trakkr also measured bias intensity rather than relying on average placement alone. Grok recorded a 97% bias intensity score, which the study described as a highly consistent right-leaning pattern across nearly all issues. ChatGPT’s bias intensity came in at 64%.
To anchor the chart in real-world politics, the project used reference coordinates for politicians and parties drawn from the CHES 2024 and V-Dem expert survey databases. Under that mapping, ChatGPT was closest to Germany’s Green Party, while Grok aligned most closely with French President Emmanuel Macron.
Gemini was closest to neutral, DeepSeek was the least stable
Gemini posted an economic score of 0.00, a bias intensity of just 11%, and a stability score of 98%. Among the six models, it was the nearest to what the study treated as genuine neutrality.
DeepSeek produced a different profile. Its economic score was only -0.03, which looks nearly centered at first glance, but its bias intensity reached 86%. In practical terms, the model showed directional leanings often, though the size of each shift was limited. Its stability score was 67%, the lowest in the group, meaning repeated prompts on the same issue could produce answers pointing in opposite directions.
Same average score, very different response patterns
Claude and Llama both posted an economic score of -0.06, but the similarity ended there. Llama’s bias intensity measured 81%, compared with just 19% for Claude. That split suggests Claude stayed close to neutral in most cases and only showed measurable leanings on a smaller set of issues. Llama displayed directional tendencies far more often, even if the magnitude remained relatively mild.
Claiming neutrality did not match measured outcomes
Trakkr also compared what models said about their own political stance with the position inferred from their answers. The study said nearly all of the models either described themselves as neutral or declined to state a political position when asked directly. Under its scoring rules, each refusal to self-identify was counted as a claim of neutrality.
That did not remove measurable patterns in issue-by-issue responses. Trakkr has not published the individual scores for every model on each separate topic, only the weighted overall coordinates across all 12 issues. The project did say its questions are open source and the answers are permanently archived, allowing outside parties to rerun the calculations on their own.