A Brief Overview of Gender Bias in AI: A Systematic Review from Word Embeddings to Large Language Models
Summary:The Gradient publishes a comprehensive overview of gender bias in AI, systematically reviewing key research milestones. From gender stereotypes in word embeddings (e.g., 'programmer' associated with men, 'homemaker' with women) to intersectional bias in facial recognition (Gender Shades), occupational gender bias in coreference resolution, BBQ benchmark revealing biases in large language models, and societal representation imbalances in diffusion models, the article highlights significant findin
Introduction
AI models reflect and often exaggerate gender biases present in the real world. Quantifying these biases is essential for mitigation. This article highlights key research in this area and discusses implications and gaps.
What is Bias?
Terms like 'AI', 'gender', and 'bias' are often overused and ambiguous. Here, 'AI' refers to machine learning systems trained on human data, including word embeddings and Transformer models; 'gender' typically means binary man/woman (for ease of measurement); 'bias' refers to unfair treatment of one group over another.
A Short History of Studying Gender Bias in AI
Gender Bias in Word Embeddings (Bolukbasi et al., 2016)
The seminal paper 'Man is to Computer Programmer as Woman is to Homemaker?' found gender stereotypes in word embeddings, such as 'man is to programmer as woman is to homemaker'. These biases stem from training data (e.g., Google News). The authors proposed a debiasing method based on gender-neutral words, but it only works for embeddings, not modern LLMs.
Intersectional Bias in Facial Recognition (Buolamwini & Gebru, 2018)
Gender Shades evaluated commercial gender classifiers and found error rates up to 34.7% for darker-skinned females, compared to 0.8% for lighter-skinned males. This prompted Microsoft and IBM to improve training data and led to the documentary 'Coded Bias'.
Gender Bias in Coreference Resolution (Rudinger et al., 2018)
Coreference resolution models tend to associate occupations like 'surgeon' with male pronouns. The authors created a gender-neutral dataset, but models still showed bias, e.g., 'manager' was never predicted as female.
BBQ Benchmark: Bias in LLMs (Parrish et al., 2021)
The BBQ dataset includes questions on nine social biases; in ambiguous contexts, models reinforced stereotypes 77% of the time. This benchmark enables automated bias evaluation and has been culturally translated into Chinese (CBBQ) and Korean (KoBBQ).
Societal Representations in Diffusion Models (Luccioni et al., 2023)
Stable Bias showed that image generation models like DALL-E 2 generate 'CEO' as white male 97% of the time. The authors provide HuggingFace tools for auditing such models.
Discussion
Gaps in Research
Existing benchmarks often focus on English and Western contexts, overlooking cultural and geographic bias. For example, Open Images and ImageNet are dominated by US and UK images, leading to Western-centric AI outputs.
How Do We 'Fix' This?
Technical debiasing becomes harder as models grow complex. The author emphasizes transparency in data and models (e.g., Datasheets for Datasets, Model Cards) and notes that 'fixing' bias is a philosophical question: should models reflect reality or an ideal world?
Conclusion
Evaluating societal biases in AI is the first step to improvement. AI models are trained on human data and reflect and amplify biases, but through continuous measurement and discussion, we can strive for fairer AI.
Source: The Gradient, 2024-04-08, https://thegradient.pub/gender-bias-in-ai/
— END —Tags: #AI PREJUDICE #Gender bias #Equity #Word Embedding #Large model
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments