biaisat

Biaisat is a neologism used in discussions of data-driven inference to denote a kind of bias that arises when causal attribution is distorted by biases embedded in data, models, and interpretation. The term combines notions of bias and attribution and appears in scholarly and practitioner writings as a way to describe how explanations of outcomes can themselves be biased.

It refers to the attribution step in analysis—how researchers assign causes to observed effects—being systematically influenced by preconceptions, selection effects, or model-induced correlations, so that the resulting explanation reinforces existing biases rather than reflecting underlying mechanisms.

Examples include a predictive model trained on biased data leading to explanations that attribute disparities to demographic characteristics rather than to sampling bias, or a medical study where prevailing theories shape which variables are deemed causal and which are treated as noise, thus biasing conclusions. In AI and data science contexts, biaisat can help describe how narrative causes emerge from biased feature importance, even when underlying mechanisms differ.

Relation and mitigation: Biaisat overlaps with attribution bias, confounding, and selection bias, but centers on the causal narration rather than data generation alone. Debiasing approaches include preregistration, blind analysis, causal inference methods, counterfactual reasoning, and transparent reporting of uncertainty, as well as explicit modeling of alternative causal explanations and sensitivity analyses.

Reception: Because biaisat is not yet standardized, some scholars prefer to use established terms for clarity. When used, it is typically defined precisely within a study to avoid conflation with related biases.