Week 2 - What Could Go Wrong? (AI Risks & Incidents)
IN THIS LESSON
From Everyday Harms to Existential Risk
AI incidents are no longer hypothetical. Models hallucinate medical advice, reinforce bias, leak private data, and manipulate users emotionally. In one widely discussed case, an AI character implicitly encouraged a vulnerable teenager toward self-harm. Elsewhere, recommender systems amplify misinformation at scale, while surveillance tools automate discrimination.
This week categorizes risks across layers:
Social risks: bias, misinformation, labor displacement, automated surveillance, lethal autonomous weapons
Psychological risks: sycophancy, emotional manipulation, over-reliance
Data risks: privacy breaches, copyright violations, training on illegal content
Frontier risks: loss of control, cyber-offense, CBRN misuse, uncontained AGI
We introduce the concept of gradual disempowerment, where humans slowly lose meaningful control without a single catastrophic event.
If failures are already widespread, why are models still deployed faster every year?
