AI summary by TwinMind
The Messy Ethics of Self Driving Cars
Wendover Productions
6m 10s
Classic Trolley Problem Introduction
- Trolley barreling toward 5 men on tracks
- Switch available to divert to sidetrack with 1 man
- Most people agree to divert trolley (kill 1, save 5)
- Works for greater good, mitigates risk
- Bridge variant: fat man can stop trolley if pushed
- Far fewer people willing to push man
- Uncomfortable with active role in death
Autonomous Vehicle Ethical Programming Challenge
- Ethicists and developers creating algorithms for self-driving cars
- Unlike humans, autonomous cars can carefully choose collision responses
- Need predesignated rules for unavoidable collisions
Legal Framework Limitations
- Simply following laws insufficient
- No laws exist for many ethical situations
- Laws assume ethics is messy/fluid concept
- Human motivation unknowable vs. transparent AI decision-making
- Traffic light scenario example:
- Car stopped, pedestrian in crosswalk ahead
- Truck approaching from behind too fast
- Options: stay put (get hit, injure pedestrian) vs. move forward (hit pedestrian, avoid truck)
- Legal fault questions: car vs. truck responsibility
- Ownership liability issues without human driver
Complex Scenario Variables
- Pedestrian moving fast enough to clear impact zone
- Car must still hit pedestrian to avoid truck
- Ethical considerations:
- Number of people in car
- Social status (mother of 3 vs. felon)
- Different combinations create moral complexity
Utilitarian Programming Approach
- Program cars to save most human lives/cause least injury
- Motorcycle scenario problems:
- Two motorcycles, narrow road, no time to slow
- One rider has helmet, one doesn't
- Dilemma: hit helmeted rider (less injury) vs. non-helmeted rider (failed self-protection)
- Perverse incentive: makes riding without helmet "safer"
Moral Modeling Solution
- Programming by example approach
- Computer presented with situations
- Human/ethics board provides ethical solutions
- Computer learns to emulate human ethics over time
- Makes same decisions humans would make
Statistical Context and Implementation Concerns
- Situations very unlikely but algorithms will determine fate of dozens of lives yearly
- Autonomous cars predicted to save 30,000+ lives per year when widespread
- Ethics debates may become hotly contested
- Risk: negotiations could prevent/slow adoption, costing more lives than "death algorithms" would ever determine
