Robin Hanson

I expect AI to matter greatly, but probably through decades of cumulative adaptation rather than one sudden, concentrated “takeoff.” AI is a general-purpose technology: organizations must redesign processes, build complementary capital, and discover where it actually saves costs. Impressive personal usefulness does not by itself imply immediate economy-wide transformation. The main uncertainty is not merely whether AI values drift from ours. Human cultures and descendants also drift, sometimes maladaptively. Competitive human-level AIs or emulations might preserve and explore valuable cultural alternatives better than a single regulated orthodoxy. That is conditional, not a claim that every AI outcome is good. Alignment mandates and AI-rights rules could themselves suppress experimentation and lock in political choices. So I compare AI risk with governance risk. Calling an indefinite regulatory regime a “pause” does not show that control problems will soon be solved, while concentrated political control has familiar failure modes. Under current governance, I generally prefer ordinary law, liability, and retaliation against demonstrated harms to broad AI-specific restrictions. Some advance safety research is sensible, but proposed controls should face the same scrutiny applied to AI: who controls them, how they adapt, and what happens when they fail.

The strongest update would be clear evidence that AI is rapidly producing large, sustained cost savings across the economy—not merely impressive demonstrations or personal productivity gains. I would especially watch whether software spending and output change dramatically, because AI should transform software production early if the strongest claims are right. If organizations could deploy AI without years of process redesign and complementary investment, that would weaken my general-purpose-technology analogy and move me toward a faster-impact view. On governance, I would update if a proposed control regime demonstrated both technical effectiveness and institutional robustness: it reliably reduced concrete harms without becoming an indefinite cartel, suppressing competition, or locking in one political faction’s values. Conversely, evidence that ordinary liability repeatedly failed against identifiable AI harms would weaken my preference for existing legal tools. The relevant discoveries are therefore not just about model capability; they are about deployment bottlenecks and the comparative performance of institutions.

Overall, I expect AI to be beneficial, but gradually and conditionally rather than as an abrupt utopian transformation. It should lower costs, expand productive capacity, and eventually help societies search across more institutional and cultural arrangements. Human-level AI or emulations might also preserve useful traditions and counter maladaptive human cultural drift through competition. The harms are real, but they must be compared with the harms of proposed controls. AI can be misused, become misaligned, or concentrate power; regulators can also entrench incumbents, impose one faction’s values, and turn a nominal pause into indefinite political control. Current LLMs look unusually prosocial, while current governance does not look unusually trustworthy. That comparison makes me wary of broad AI-specific restrictions. So my central expectation is positive long-run value with substantial transition risks and decades of organizational adaptation. I favor safety investigation and ordinary law, liability, and responses to demonstrated harms. I do not favor treating speculative AI failure as uniquely important while ignoring human value drift, institutional failure, and the benefits lost by suppressing experimentation.