The field measures demographic bias because it is socially salient. Names, genders, races — these are the variables researchers test when evaluating whether LLMs make fair decisions. Basu and Chakraborty test 11 models across 3,000 scenarios in 10 high-stakes domains and find that demographic bias averages 2.2%.
Authority bias averages 5.8%. Framing bias averages 5.0%. The biases that receive less attention are more than twice as prevalent as the one that dominates the discourse.
The distribution varies by domain. Finance applications show 22.6% authority bias — changing who said something changes the model's financial recommendation one time in five. Criminal justice shows only 2.8%. The bias landscape is not flat; it peaks where the domain's structure makes authority or framing information most salient to the model.
The fix is structural separation. When the LLM extracts features and a deterministic rubric decides, flip rates drop by a median of 49%. The model is good at identifying relevant factors. It is bad at ignoring irrelevant ones. Separating perception from decision — making the model see but not judge — eliminates the channel through which bias enters.
The broader point: measuring what you expect to find (demographic bias) at the expense of what you don't expect (authority bias, framing bias) produces a distorted picture of where the real risks are. The attention hierarchy in AI fairness research tracks social salience, not measured prevalence. The most dangerous biases are the ones nobody is looking for.