Governments are racing to regulate AI faster than they understand it. The approaches differ sharply by region, and the stakes are high: regulate too hard and you stall benefits; too soft and you invite harm.
Major frameworks
| Region | Approach |
|---|---|
| EU AI Act | Risk-tiered: bans "unacceptable" uses, heavy rules for "high-risk," lighter for the rest |
| US | Sectoral + executive orders; lighter-touch, innovation-focused |
| China | State-controlled, content/alignment to government values |
| UK | "Pro-innovation," principles over hard rules |
The risk-tier idea (EU)
Unacceptable (banned) → social scoring, manipulative AI
High-risk (strict) → hiring, lending, medical, biometric
Limited (transparency) → chatbots must disclose they're AI
Minimal (free) → spam filters, gamesThe hard problems
- Pacing — laws take years; models ship monthly.
- Definition — "AI" is a moving target; rules risk being obsolete on arrival.
- Jurisdiction — a model trained anywhere serves everywhere (Moloch and Coordination Failures).
- Frontier risk — most rules target applications, not the catastrophic-risk frontier.
Regulation is one proposed answer to the vulnerable world — trading some freedom and speed for safety.
Related: The Vulnerable World Hypothesis · Autonomous Weapons · Algorithmic Bias