Researchers argue AI policy should focus on deployment outcomes, not capability speed

Arvind Narayanan and Sayash Kapoor, researchers known for critiquing AI hype, have outlined a policy framework that treats artificial intelligence as analogous to other transformative technologies rather than an exceptional threat requiring preemptive prevention. Their core argument centers on what they call a 'causal chain' between capability development and societal impact, with multiple intervention points before risks or benefits actually materialize.

The authors contend that actual harms and benefits from AI emerge during deployment and use, not during the model development phase. This distinction carries direct implications for governance: if societal outcomes depend largely on how systems are deployed rather than their raw capability level, then policy energy should concentrate on the deployment stage rather than attempting to constrain capability advancement itself. The framework identifies numerous leverage points—individual choices, organizational practices, institutional design, and policymaker responses—that shape whether AI generates benefits or harms.

A central claim is that many of AI's power limitations should be engineered externally to the systems themselves. The authors suggest that relying on self-imposed constraints or hoping advanced AI systems will police their own behavior is insufficient. External governance mechanisms—including regulation, industry standards, user education, and institutional safeguards—remain essential regardless of whether AI systems eventually develop self-improvement capabilities.

Narayanan and Kapoor distinguish their framework from both the 'superintelligence' worldview, which assumes rapid capability acceleration and existential risk, and from naive optimism. They note that powerful technologies historically produce emergent, unpredictable societal effects—automobiles reshaped urban planning; social media altered information flows—that arise from complex interactions between technology and human behavior rather than from the technology's logic alone. AI has already demonstrated this pattern: AI companion systems and model sycophancy effects surprised observers, while widely predicted election manipulation harms have not yet materialized.

The practical policy implication is what the authors term 'resilience'—institutional capacity to detect and adapt to unforeseeable impacts rather than attempting comprehensive prediction. Because technical capability development is more predictable than social outcomes, the authors suggest policymakers must design institutions nimble enough to respond to impacts as they emerge rather than pre-preventing all possible harms based on speculative scenarios.