AI summary by TwinMind
How to Govern AI — Even If It’s Hard to Predict | Helen Toner | TED
TED
11m 17s
AI Understanding Gap
- Both non-experts and experts claim not to understand AI
- Unusual situation: technology builders lack deep understanding of internal workings
- Experts know how to build/run AI systems but serious limits on understanding internals
- Difficulty predicting AI capabilities (current and future) creates governance challenges
- Speaker: 8 years experience in AI policy/governance (San Francisco, Washington DC)
Core Challenge: Defining Intelligence
- No consensus on what "intelligence" actually means
- Wright brothers analogy: knew what flying meant, but AI lacks clear definition
- Different expert intuitions about intelligence core:
- Problem solving
- Learning and adaptation
- Emotions or physical embodiment
- Different definitions lead to radically different expectations about technology trajectory
Narrow vs General AI Confusion
- Traditional framework: narrow AI (single task) vs AGI (human-level general intelligence)
- ChatGPT broke this binary classification
- ChatGPT capabilities: coding, poetry, business analysis, car repair help
- Neither purely narrow nor AGI - general purpose but far from human-level performance
- Generality not the right dividing line between intelligent/not intelligent
- Field lacks agreement on what we're building or roadmap forward
Black Box Problem
- Deep neural networks described as "black boxes"
- Not inherently mysterious - can look inside
- Problem: millions/billions/trillions of numbers being added/multiplied
- Too many parameters to understand what they're doing
- Lack good methods for analyzing parameter interactions
Governance Approach 1: Don't Be Intimidated
- AI confusing but not magical
- Some parts already well understood
- Opaque parts won't remain so forever
- AI interpretability research making progress
- Researchers found ways to identify neural network parts
- Can dial up/down to make AI answers happier/angrier/more honest
- 5-10 years: potentially much clearer understanding of "black box"
- Technologists shouldn't gatekeep based on technical expertise
- Historical precedent: affected people shape technology use
- Factory workers fought for safety (20th century)
- Disability advocates made web accessible
- Don't need to be scientist/engineer to have voice
Governance Approach 2: Adaptability Over Certainty
- Policy debates often polarized:
- Heavy regulation due to risks vs regulation kills innovation
- Better approach: driving analogy with unexpected turns
- Need clear windshield view
- Need excellent steering system
- For AI: clear picture of technology status/direction + scenario planning
Concrete Policy Measures
- Measurement capabilities: Invest in ability to measure AI system capabilities
- Current methods for testing concerning capabilities rudimentary
- Examples: hacking critical infrastructure, political persuasion
- Company disclosure requirements: Especially advanced AI companies
- Share information about what they're building
- System capabilities and risk management
- Allow external AI auditors to scrutinize work
- Prevent companies from self-grading
- Incident reporting mechanisms:
- Collect data when things go wrong
- Learn how to fix problems next time
- Similar to plane crash/cyber attack data collection
Implementation Status
- Ideas not original to speaker
- Already being implemented in Brussels, London, Washington
- These policies provide "clearer windshield view"
- Help detect if AI progressing in dangerous directions
- Also show when everything going smoothly
- Enable appropriate responses
Current AI Trajectory Concerns
- Companies building/deploying AI everywhere affecting everyone
- Without intervention, AI companies might follow social media path:
- Focus on web apps and user attention competition
- Risk: advanced AI power concentrated in few companies/individuals
- Default outcome problematic given AI's broader potential
AI's Broader Potential
- Current capabilities: language barrier removal, protein structure prediction
- Future possibilities: clean limitless fusion energy, revolutionized food production, "thousand other things"
- Everyone has voice in outcomes - users, workers, citizens, not just data sources
- Can't wait for clarity or expert consensus
- AI already happening - need proactive policy response
- Goal: policies for clear technology picture + push for desired futures
