The AI ethics session nobody knew we needed
summary
Search engines and LLMs routinely return a well-known falsehood, naming Cristiano Ronaldo vs Christine Sinclair as the top international scorer when Sinclair actually has more goals; image models like DALL·E poor vs rich couple produce race and class stereotypes; face-recognition systems show markedly worse accuracy on darker-skinned and female faces; code assistants train on public repos without filtering so GitHub Copilot unfiltered training pulls from unintended sources. The session argues LLM output perpetuates bias into future models, averages 10–15 personal biases per person, and demands human editing and fact‑checking not blind acceptance.
AI-written summary · about
chapters (10)
- 0:00Welcome
- 1:27Why AI ethics needed
- 3:55Systemic biases in training data
- 8:09Five stages of AI ethics grief
- 9:55Denial stage challenges
- 14:53Anger stage limitations
- 24:54Bargaining stage statistics
- 34:06Depression stage flawed outputs
- 48:38Acceptance stage responsibility
- 57:54Final thoughts
related talks
38mSocial Implications of Bias in Machine Learning
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML
29mResponsible AI Challenges and Recommendations
45mThe Myth of Neutrality: How AI is widening social divides
11mMachines of Love & Grace – Building Truly Inclusive AI
47mEthics & AI: Identifying & preventing bias in predictive models
38mSocial Implications of Bias in Machine Learning
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....