Neutral by Design: More Than a Century of Hidden Bias in Classification Systems

Classification systems often present themselves as neutral tools while carrying the hierarchies of the people who design them. This paper follows that pattern across three settings. In nineteenth-century ethnology, Frederick Douglass and W.E.B. Du Bois exposed race science as prejudice dressed in method, years before Franz Boas made similar arguments with the authority of the academy behind him. In late-twentieth-century American English, Jane Hill showed how Mock Spanish passed off a racist stereotype as friendly humor. In current AI practice, cultural labels in persona schemas draw stereotyped behavior out of language models, a finding supported by independent studies and by counterfactual audits of the MatrAIx framework. The same mechanism recurs in all three: the category is built by those with power, framed as neutral, completed by its audience, denied by its users, and cited as its own proof. Because corrections layered on top of a label have repeatedly failed, the paper argues for removing cultural labels and describing situations instead, and offers six recommendations for the designers, engineers, researchers, and product managers who build and use persona systems.