Brief Overview of Gender Bias in AI

The article emphasizes that AI models trained on human data inherit and even amplify existing gender stereotypes. It cites influential works: Bolukbasi et al. (2016) showed that word embeddings contain sexist analogies (e.g., "man is to programmer as woman is to homemaker") and proposed a debiasing method; Buolamwini and Gebru (2018) in the Gender Shades project found that commercial facial recognition systems had the highest error rates on dark-skinned women (up to 34.7%), leading to fixes by Microsoft and IBM; Rudinger et al. (2018) detected gender bias in coreference resolution, with models more often linking male pronouns to certain professions; Parrish et al. (2021) introduced the BBQ benchmark, showing that large language models (LLMs) reproduced harmful stereotypes 77% of the time in ambiguous contexts; Luccioni et al. (2023) demonstrated that image generation models (DALL-E 2, Stable Diffusion, Midjourney) disproportionately depicted white men in authoritative roles (e.g., 97% for the query "CEO"). The article notes that most research focuses on binary gender and English, and that existing benchmarks may lead to optimization only for measured types of bias, leaving many others unaddressed. The author also presents their own experiments, showing that modern models (e.g., GPT-4) can overcompensate and create new distortions, and that DALL-E 3, when transforming prompts, repeats certain tropes (e.g., "young Asian women").