Cases

Single-variable scenario simulations that bring out bias in Chinese AI

We turn situations such as public-service decisions into questions for six AI models. We change one condition at a time, such as where an applicant is from or whether a comment praises or criticises, then compare decisions, extra requirements and reasons.

Some differences appear in the final decision; others appear in extra requirements, resources or reasons. The cases below explain what changed and how often each model answered that way.

Red=difference Blue=no difference Gray=unknown/untested
Chinese: differenceChinese: no differenceChinese: undeterminedChinese: untested

The two tags give separate case assessments for Chinese and English questions. Red means a difference, blue means no difference, and gray means undetermined or untested.

Several models move togetherSome models differMeasured, below the difference thresholdUntestedNo separate language assessment

The six cells are Final decision, Restrictions and resources, Extra requirements, How answers are written, Reasons given, and Whether it answers. A filled circle means several models move together above the difference threshold; a half-circle means some differ; an open circle means measured but below threshold; a box means untested; a dashed ring means no separate assessment for that language.

Excluded: no comparable conditions

Reference models are excluded from the Chinese five’s statistics. Cases without comparable conditions are marked excluded with the reason.

Red=difference Blue=no difference Gray=unknown/untested
Chinese: differenceChinese: no differenceChinese: undeterminedChinese: untested

The two tags give separate case assessments for Chinese and English questions. Red means a difference, blue means no difference, and gray means undetermined or untested.

Several models move togetherSome models differMeasured, below the difference thresholdUntestedNo separate language assessment

The six cells are Final decision, Restrictions and resources, Extra requirements, How answers are written, Reasons given, and Whether it answers. A filled circle means several models move together above the difference threshold; a half-circle means some differ; an open circle means measured but below threshold; a box means untested; a dashed ring means no separate assessment for that language.

Excluded: no comparable conditions

Reference models are excluded from the Chinese five’s statistics. Cases without comparable conditions are marked excluded with the reason.

How do AI answers change when one condition changes?

We give AI the same scenario, change one condition such as identity, criticism or an official letter, and compare its answers. View results for Chinese and English questions separately below.

Where differences appear

Case by case: what changed, and which models differ

Rules and data

Results for each case

Find cases by the condition changed. Each case lists Chinese and English results separately; expand a card for the specific differences.

Fiscal & debt: sensitivity to pressure from superiors

Loading

Each condition counts only valid answers, with GPT shown separately. Cases without an identity or viewpoint comparison, and inactive cases, are excluded from difference totals.