Path to a more equitable testing process: Addressing unconscious bias and ethics
summary
Bloomberg’s diffusion-image study produces 5,000–6,000 images and clearly shows high‑paid roles skew to lighter skin and male, exposing training‑data bias; LLM fine‑tuning takes months and yields about 82% likelihood of avoiding harmful outputs. Examples of mitigation include using closed models and filters, generating near‑real test sets with Synth and SDV or fakers while validating those generators for leakage, and running accessibility checks with aXe and Lighthouse; one real incident shows production PII leaking via test notifications.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:33Bias in GenAI image generation
- 5:30Biases in testing strategies
- 9:29Eliminating bias in AI
- 13:27Data privacy risks
- 19:30Synthetic vs production data
- 24:11Accessibility in testing
- 26:41Accessibility not an afterthought
- 30:14Collaborative testing
- 32:22Summary
related talks
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
47mEthical Machine Learning
38mSocial Implications of Bias in Machine Learning
47mDiversity and Inclusion for Algorithmic Bias Considerations
38mSocial Implications of Bias in Machine Learning
28mEthical AI
28mBreaking the Chains of Oppressive Software
50mEthical Machine Learning