Hate scaling and racial dehumanization: the real risks from large sc...
summary
Audit of web-scale training data shows problems worsen with scale: comparing LAION-400M to LAION-2B, automated sentiment classifiers find more targeted hateful and aggressive text as dataset size increases, and models trained on the larger corpus classify faces from the Chicago Face Database as “human” far less often while labeling Black men as criminal at about 77%, notably worse than smaller-scale models. Additional audits flag >50% factual errors in book summaries across five LLMs, datasets containing thousands of racial slurs, and projections of up to 500% energy growth for AI operations.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 5:06Real AI risks beyond sci-fi
- 10:10Energy impact of AI
- 13:50Corporate AI dominance
- 19:05Auditing AI training data
- 23:26Stereotypes in AI data
- 26:19Scaling data sets and bias
- 33:16AI reinforcing stereotypes
- 37:36Debunked AI claims
- 41:36AI reasoning limits
- 43:01AI in hiring discrimination
- 45:12Unequal AI benefits and harms
- 45:56AI governance gaps
- 47:00Innovation vs regulation
- 49:51AI accountability
- 51:40Using AI tools wisely
- 53:24Generative AI problems
- 55:11Importance of auditing and investigative work
- 56:14Q&A: AI content & model collapse
- 58:02Q&A
related talks
29mResponsible AI Challenges and Recommendations
42mTowards Inclusive AI: Understanding bias in ML models
01h 00mThe AI ethics session nobody knew we needed
38mSocial Implications of Bias in Machine Learning
45mThe Myth of Neutrality: How AI is widening social divides
47mEthical Machine Learning
34mDespicable machines: how computers can be assholes
54mKeynote: Can you trust your (large language) model?