Historical Analogues That Can Inform AI Governance
Bibliographic Info
Vermeer, Michael J. D. Historical Analogues That Can Inform AI Governance. RAND Corporation, RR-A3408-1, 2024.
Core Question
“What is AI like?” — and how the answer determines which governance model fits.
Key Argument
None of the historical governance models is a perfect analogue for AI, but each holds lessons depending on how AI develops. The effectiveness of governance depends on matching mechanisms to technology characteristics.
Four Historical Analogues Examined
- Nuclear Technology: High resources, controllable physical assets, consensus on catastrophic risk
- The Internet: Government as funder-facilitator, private sector leads, minimal safety concerns at origin
- Encryption Products: Cautionary tale of governance failure when consensus erodes
- Genetic Engineering: Voluntary moratoria, scientific community self-governance, trusted coordinating bodies
Three Governance Themes
- Sustained consensus on norms: When it erodes, governance fails
- Physical vs. nonphysical assets: Governing nonphysical assets (code, knowledge, model weights) is fundamentally harder
- Public-private partnerships: Essential but take different forms depending on who leads
Key Observations
- The Collingridge Dilemma: When tech is new, easy to change but hard to see impacts; when mature, impacts clear but change is costly
- Pursuit of monopoly on frontier AI development is likely unrealistic (mirrors early nuclear thinking)
- Speculative or ambiguous risks won’t form the basis of strong consensus
- Controls substantially dependent on governing nonphysical assets will likely fail
- Economic security should be viewed as part of national security, not in tension with it