Brief overview and discussion of gender bias in AI
For International Women’s Day, I wanted to write a short article about gender bias in AI.
AI models reflect and often exaggerate existing gender biases in the real world. To properly address and mitigate these biases, it is important to quantify the biases present in your model.
This article introduces some of the important work that has been done (and is currently being done) to discover, evaluate, and measure various aspects of gender bias in AI models. I also discuss the implications of this work and highlight some of the differences I found.
All of these terms (‘gender’, ‘bias’ and ‘AI’) can be a bit overused or ambiguous.
In the context of AI research, ‘gender’ typically includes the binary male/female (because it is easier for computer scientists to measure) and sometimes also includes a ‘neutral’ category. “AI” refers to machine learning systems trained on human-generated data and includes both statistical models such as word embeddings and newer Transformer-based models such as ChatGPT.
In the context of this article, I use ‘prejudice’ to broadly refer to the unequal, disadvantageous and unfair treatment of one group over another.
There are many ways to classify, define, and quantify bias, stereotypes, and harm, but these are beyond the scope of this article. I’ve included a reading list at the end of the article, so if you’re curious, take a look.
here minuscule We sought a sample of influential papers studying gender bias in AI. This list is not comprehensive by any means, but it is intended to illustrate the diversity of research studying gender bias (and other kinds of social bias) in AI.