BERT-based disinformation classifier (AA:2023:01)

Summary advice

The audit commission believes there is a low risk of (higher-dimensional) proxy discrimination by the BERT-based disinformation classifier and that the particular difference in treatment identified by the quantitative bias scan can be justified, if certain conditions apply.

Source of case

Applying our self-build unsupervised bias detection tool on a self-trained BERT-based disinformation classifier on the Twitter1516 dataset. Learn more on Github.

Stanford's AI Audit Challenge 2023
This case study, in combination with our bias scan tool, has been selected as a finalist for Stanford’s AI Audit Challenge 2023.

A visual presentation of this case study can be found in this slide deck.


Dowload the full report and problem statement here.

Normative advice commission

  • Anne Meuwese, Professor in Public Law & AI at Leiden University
  • Hinda Haned, Professor in Responsible Data Science at University of Amsterdam
  • Raphaële Xenidis, Associate Professor in EU law at Sciences Po Paris
  • Aileen Nielsen, Fellow Law&Tech at ETH Zürich
  • Carlos Hernández-Echevarría, Assistant Director and Head of Public Policy at the anti-disinformation nonprofit fact-checker
  • Ellen Judson, Head of CASM and Sophia Knight, Researcher, CASM at Britain’s leading cross-party think tank Demos

Funded by