References
Numbered as they are on the slides. If you came here from number seven, number seven is still number seven.
Breaking the Pattern
Open-source approaches to quantifying bias in generative AI — the works cited, and the tools you can actually run. The earlier version of this argument is in the speaking list.
The first four are the problem. The rest are the instruments.
- AI-assisted recruitment is biased. Here's how to beat it — World Economic Forum, 2019. Hiring as the first place most people met algorithmic bias.
- Lensa AI portraits and misogyny — The Guardian, 2022. What a consumer app returned to women who fed it their own photographs.
- AI police sketches — Vice, 2022. Generative models pointed at forensic use, with predictable results.
- It's Not Fair: Detecting Algorithmic Bias with Open Source Tools — RSA Conference, 2022. The earlier talk this one builds on.
- BOLD — dataset and metrics for measuring bias in open-ended language generation.
- BBQ — the Bias Benchmark for QA, which tests what a model assumes when the question does not say.
- More human than human: measuring ChatGPT political bias — Public Choice, 2023.
- StableBias — a running demonstration of how generated faces shift with the prompt's profession.
- T2IAT — measuring valence and stereotypical bias in text-to-image generation.
- A Sign That Spells: DALL-E 2, Invisual Images and The Racial Politics of Feature Space — arXiv, 2022.
If a link here has rotted, or the tooling has moved on, tell me and I will fix it.