This working paper introduces and establishes priority for AI Meta-Bias, a concept first developed and published by the author in December 2025. AI Meta-Bias is a second-order bias located in the measurement frameworks used to evaluate AI systems, not within the systems themselves. It is bias about bias: the assumptions and blind spots built into the tools we use to find AI bias, made permanently invisible by the decision to treat those tools as objective. The paper presents the formal definition, distinguishes the concept from adjacent work including automation bias and algorithmic bias, provides an evidence base drawn from peer-reviewed research and documented cases, and sets out a research proposal for the full empirical paper in preparation. The related concept of the Recursive Bias Paradox, the argument that unbiased AI is conceptually incoherent, is also originated by this author and covered in this deposit.
Meriel Batterley (Sat,) studied this question.