AI models do not just reflect gender bias. They amplify it. A 2016 paper by Bolukbasi et al. demonstrated this mathematically using word embeddings trained on Google News: the analogy 'man is to computer programmer as woman is to homemaker' fell out of simple vector arithmetic. A 2018 audit by Buolamwini and Gebru found commercial facial recognition systems misclassified darker-skinned females at error rates up to 34.7%, compared to 0.8% for lighter-skinned males. These are not edge cases. They are documented, reproducible failures in systems that were already deployed.
This article by The Gradient maps the research lineage of gender bias measurement in AI, from early word embedding debiasing methods that no longer apply to modern transformers, to coreference resolution benchmarks, to text-to-image systems that default to male representations of 'doctor' and female representations of 'nurse'. The debiasing fix Bolukbasi proposed worked only on static embeddings. It does not transfer to GPT-scale systems. That gap is the point. Each solution has exposed a new layer of the problem, and the article is honest about where the field has stalled.
The piece is worth reading in full for its treatment of what is not yet measured: non-binary gender representation, bias in languages other than English, and the downstream harms that follow when biased embeddings feed into hiring tools or document ranking systems. The author flags these gaps explicitly. The reading list appended to the original is a functional research primer on its own.
[READ ORIGINAL →]