Bayesian experimental design (BED) traditionally aims to collect data that is maximally informative about a target quantity, without considering other properties of the resulting dataset. A new paper introduces FairBED, an extension that explicitly targets datasets that are both informative and fair. The authors define fairness for datasets themselves, measuring how compatible a dataset is with a "fair world" in which the target variable is unrelated to sensitive attributes.

FairBED builds practical objectives that maximize expected information gain about the target while penalizing information that reveals sensitive attributes or their coupling with the target. The framework is also extended to active learning, where data is acquired sequentially. The authors report empirical results showing that models trained on data gathered with FairBED offer improved fairness-accuracy trade-offs compared with randomly acquired data and with conventional BED.

The work reframes fairness as a property of the data-collection process rather than only of the trained model, potentially giving practitioners a principled way to build fairer datasets from the start. As a single-source digest, this summary reflects only the claims made in the arXiv paper; no independent replication or commentary is included.