AI Safety Researcher Argues Pausing at Human-Level Harder Than ASAP
Why waiting for human-level AI makes a pause nearly impossible.
In a cross-posted essay on LessWrong, AI researcher MichaelDickens counters the popular notion that we should pause AI only after reaching human-level intelligence. He argues that waiting makes a pause harder, not easier. Human-level AI will generate enormous economic value, giving companies and lobbyists powerful reasons to resist any halt. Moreover, AI systems that can accelerate safety research will also—and likely more effectively—accelerate capabilities improvements, because capabilities are easier to measure and optimize. This asymmetry means the gap between dangerous and safe AI could widen faster than we can manage.
Dickens also points out that advanced AI lowers the barrier to entry, enabling clandestine development in garages rather than requiring teams of PhDs, making enforcement nearly impossible. He critiques the strategy of “burning the lead” at the last minute as psychologically unrealistic. Finally, he acknowledges one counterpoint: public backlash against AI-driven unemployment could create political will to pause. But he concludes that relying on that future disruption is too risky, and urges action now rather than waiting for a human-level milestone.
- Human-level AI's massive economic value creates strong incentives for industry to resist a pause.
- AI accelerates capabilities faster than safety because measurable tasks (like coding) improve more readily than abstract reasoning.
- Clandestine AI research becomes easier with advanced models, making enforcement of any international pause far more difficult.
Why It Matters
The timing of an AI pause affects global regulation strategies — delaying until human-level may be too late.