Opening up the box: the complexity of human interaction with AI systems
summary
Argues that every AI design choice matters: image prompts in stable diffusion default to whiteness unless explicitly corrected; a chatbot (ChatGPT) fabricates six fake precedents that sound plausible but are false; a US hospital triage proxy equals cost leads to racial discrimination because spending replaces clinical labels; an NLP breakthrough was driven by an NER metric switch not a true model advance. Presents two working uses — AI spotting overlooked pain and cutting CO2 in manufacturing — and shows which choices made them work.
AI-written summary · about
chapters (13)
- 0:00AI systems and human decisions
- 1:10Who builds AI and why
- 2:58Bias in AI outputs
- 4:30Bigger vs better in AI
- 7:25Defining 'better' in AI
- 8:26Garbage in, garbage out
- 10:15Toxicity and labor in AI
- 11:57Mathematical choices in AI
- 14:41Evaluating AI performance
- 17:55Blindly following AI
- 19:51Call for responsible AI
- 20:56Beneficial AI examples
- 22:45Q&A
related talks
45mEthical use of AI: seeking, understanding, and countering the risks
24mA Call for Human-Centered AI
01h 01mKeynote: It AI-n't What You Think!
57mDesigning Human-AI Experience
44mToxic Trust: The Psychology of Automation Bias in Generative AI
42mBrand Humanizing: A humane turn on AI
28mEthical AI?!
10mMastering the Art and Science of AI Communication