AI summary by TwinMind
AI Is Dangerous, but Not for the Reasons You Think | Sasha Luccioni | TED
TED
10m 19s
Introduction and Context
- AI researcher with 10+ years experience received email claiming AI work will end humanity
- AI headlines range from positive (new medicine molecules, Pope images) to negative (chatbot divorce advice, chlorine gas recipe suggestions)
- Background discussions about doomsday scenarios, existential risk, singularity
- Focus should be on current tangible impacts vs future speculation
Environmental Impact and Sustainability
- AI models exist in physical infrastructure (metal, plastic, energy consumption)
- Each AI query has planetary cost
- BigScience initiative: 1,000 researchers created Bloom (open large language model)
- Emphasis on ethics, transparency, consent
- Training used energy equivalent to 30 homes for 1 year
- Emitted 25 tons CO2 (equivalent to driving car 5x around planet)
- GPT-3 emits 20x more carbon than Bloom
- Large language models grew 2,000x in size over 5 years
- Switching smaller to larger model = 14x more carbon for same task
- AI integration into phones, search engines, fridges, speakers multiplying environmental costs
- Tech companies not measuring/disclosing environmental impacts
- CodeCarbon tool: runs parallel to AI training, estimates energy consumption
- Enables informed choices between models
- Supports deployment on renewable energy
Copyright and Consent Issues
- Artists/authors struggle to prove their work used in AI training without consent
- Spawning.ai (founded by artists) created "Have I Been Trained?" tool
- Searches massive datasets to find personal content
- LAION-5B dataset example: contains images of speaker and other women named Sasha
- Image generation models for "woman named Sasha" often produce bikini models
- Karla Ortiz case: used tool evidence for class action lawsuit against AI companies for copyright infringement
- Spawning.ai partnership with Hugging Face: opt-in/opt-out mechanisms for dataset creation
- Human artwork shouldn't be "all-you-can-eat buffet" for AI training
Bias and Discrimination
- AI models encode stereotypes, racism, sexism
- Dr. Joy Buolamwini's research: facial recognition systems failed to detect her face without white mask
- Systems vastly worse for women of color vs white men
- Deployment in law enforcement = false accusations, wrongful imprisonment
- Porcha Woodruff case: wrongfully accused of carjacking at 8 months pregnant due to AI misidentification
- AI systems are "black boxes" - creators can't explain exact functionality
- Image generation bias in forensic sketches for terms like "dangerous criminal," "terrorists," "gang member"
- Stable Bias Explorer tool: explores bias in image generation models through professions
- Scientists generated as men in glasses/lab coats
- 150 professions analyzed across multiple models
- Significant overrepresentation of whiteness/masculinity vs US Labor Bureau statistics
- Lawyers, CEOs shown as men nearly 100% of time
- Tool presented at UN event on gender bias
- Designed for non-coders to understand AI bias
Solutions and Call to Action
- No single solution for complex issues (bias, copyright, climate change)
- Tools for measuring AI impact enable:
- Companies: choose sustainable, copyright-respecting models
- Legislators: develop regulation/governance mechanisms
- Users: select trustworthy AI models
- AI development is ongoing, not predetermined
- Collective decision-making possible for AI direction
- Current tangible impacts more important than future existential risks
