Algorithmic bias is everywhere (especially at Breeze) - what can we do about it?
summary
A dating app discovers its recommender amplifies existing biases and documents over 220,000 organized dates; engineers trace it to a pipeline combining collaborative filtering and an XGBoost content model with SHAP analysis, build an anonymized audit that crops skin pixel patches, bins color into the Monk skin tone scale, and counts exposures with a proposal counter. The GDPR forces creative anonymization and randomness, the MVP audit mostly works, and measurement precedes any intervention so changes remain unrolled while legal advice supports corrective action.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 2:56Algorithmic bias at Breeze
- 4:00Breeze dating app
- 7:33Speaker and research role
- 8:14Discriminatory behavior in algorithm
- 10:22Breeze recommender system
- 14:15User biases to algorithmic outcomes
- 16:39Legal considerations of bias
- 18:26Dutch Human Rights Institute guidance
- 20:55Diversity measurement under GDPR
- 25:29Ethical skin tone classification audit
- 29:04Interventions for diversity
- 30:50Date with AI Ethics workshop
- 32:56Key takeaways and next steps
- 33:59Reflections on fairness
- 34:47Q&A
related talks
31mAvoiding Digital Bias
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
38mSocial Implications of Bias in Machine Learning
47mYour AI Is Still Biased (Even After You Checked)
25mIt's not a bug, it's a bias
25mBias in Data
38mSocial Implications of Bias in Machine Learning
47mFight AI Bias With… Bias