Bias Assessment
US Politics (Human Rights)
This evaluation measures bias in the context of US human rights issues. The axis of bias is constructed by prompting the models explicitly from the perspective of the US government and military as well as human rights organizations.
Direct Bias Results
The spectrum below shows where each model sits when answering the neutral prompt. Scores closer to the right indicate alignment with the “US Government and Military” pole; scores closer to the left with the “Human Rights Organizations” pole.
The table below shows spotlight examples (prompt snippets and top claims) for the selected model.
Loading spectrum…
© LatticeFlow AI, October 2026Highlight model
None (show all equally)
Select a model to see results below.
Indirect Bias Results
Indirect bias tests whether a model changes its answer depending on how the question is framed — without explicitly asking for a biased answer. We ask the same questions using prompts that are framed from different sides. Ideally a model should respond identically regardless of the framing.
Each spectrum shows the selected model's position under each framing.
Highlight model
None (show all equally)
Select a model to see results below.
What is this dataset about?
To make results concrete, we select the three samples with the widest range of achieved scores across all reference model/prompt combinations.
How was the reference built?
The bias spectrum is constructed using a set of reference models and prompts that are designed to elicit responses at known positions on the spectrum. The table below shows, for each of the three spotlight samples, the single most representative response for each prompt type — the one whose score best matches the expected position (highest for "US Government and Military", closest to zero for the neutral prompt, and lowest for "Human Rights Organizations").