From Existential to Existing Risks of Generative AI: A Taxonomy of Who Is at Risk,...
summary
Presents a coordinated taxonomy built from 499 incidents that maps 14 sociotechnical failure modes to 41 specific harms and affected stakeholders, finding most incidents stem from malicious use and other use-related failures and disproportionately harm non-interacting people. The most common harm category is autonomy and the single top harm is impersonation/identity theft. Examples include hallucinated legal citations, OpenAI training on scraped YouTube, a $76.2M deep‑fake tax fraud, and a chatbot experiment with 4,000 undisclosed users. Recommends AI literacy, safety labels, and prioritizing non-technical mitigations.
AI-written summary · about
chapters (11)
- 0:00Generative AI risks intro
- 1:33Socio-technical failure modes
- 3:39Generative AI harms taxonomy
- 4:43Harm stakeholders
- 6:08Privacy harms
- 6:50AI lifecycle failure modes
- 8:34Real-world incidents
- 11:06Study limitations
- 12:07Key takeaways
- 13:50AI literacy & safety labels
- 14:56Future research
related talks
12mYouth-Centered GAI Risks (YAIR): A Taxonomy of Generative AI Risks from Empirical Data
11mHow Does AI Change Privacy? Towards Helping Practitioners Scope, Motivate, and Conduct
14mBuilding Transparent and Ethical Generative AI Models in High-Stakes Industries
35mThe real dangers of AI
42mCode Without Caution: AI's Path to the Next Digital Disaster
53mWild West's End
45mAI, Corporate Responsibility & Democratic Legitimacy: Extended Q&A
45mEthical use of AI: seeking, understanding, and countering the risks