This is a genuine strategic disagreement among people who share the same goal, and it echoes a long-running debate in international relations about how regimes are best built. Both roads have delivered major treaties in the past; both have also failed. Understanding the trade-off clarifies what each approach buys and what it risks — and points toward a strategy that uses both.
The case for incrementalism
The incremental argument is grounded in political realism. Comprehensive treaties are extraordinarily hard to negotiate, slow to ratify, and easy for any major party to block. Meanwhile, the technology is advancing now. Incrementalism says: take the achievable steps immediately — voluntary commitments, safety institutes, confidence-building measures, compute controls, a framework convention — and build momentum, trust, and institutions that later support stronger agreement. It reduces near-term risk while the harder deal is still out of reach, and it avoids staking everything on a single negotiation that might fail.
The incremental path also lets governance learn. Early, modest measures reveal what works, build the technical and institutional capacity a treaty would need, and establish norms that make binding commitments easier to reach. Most successful regimes — ozone, climate, trade — grew this way, from framework to protocol to tightened protocol, rather than arriving fully formed.
The case for a comprehensive treaty
The comprehensive argument answers with a warning: incrementalism can become a permanent substitute for the real thing. A world content with voluntary commitments and dialogues may never summon the will for binding constraint, especially when each small step lets everyone claim progress while the fundamental problem — the race to build uncontrollable systems — goes unaddressed. Worse, the timeline may not allow a leisurely accretion. If dangerous capabilities could arrive within years, a strategy that produces binding limits only after decades of incremental trust-building is a strategy that arrives too late.
Comprehensiveness also has a coherence advantage. The pieces of AI governance are interdependent: thresholds are meaningless without verification, verification without enforcement, enforcement without broad participation. A comprehensive treaty can design these to fit together, whereas an incremental patchwork risks gaps, inconsistencies, and a false sense of security from measures that do not actually bind the dangerous behaviour.
The real trade-off
- Speed vs sufficiency. Incrementalism acts sooner but may never reach binding constraint; comprehensiveness aims at what is actually required but may arrive too late or not at all.
- Achievability vs coherence. Small steps are easier to agree but risk a fragmented regime with gaps; a single treaty is coherent but far harder to conclude.
- Momentum vs complacency. Incremental wins build trust — or they let everyone declare victory while the core problem festers.
The danger of incrementalism is that it becomes an excuse never to do the hard thing. The danger of comprehensiveness is that you wait for the perfect treaty while the window closes. The way through is to do both — and to be honest about which one actually stops the race.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
Why the answer is both — with a clear destination
Framed as a binary, the debate is a false choice. The coherent strategy is to pursue incremental measures as the means and a comprehensive binding agreement as the end — using the near-term steps deliberately as building blocks toward the treaty, not as substitutes for it. Safety institutes, confidence-building measures, and compute governance are worth doing now precisely because they reduce immediate risk and construct the capacity a treaty requires. What matters is holding onto the destination: keeping the comprehensive, binding, verified agreement as the explicit goal, so that incremental progress is measured by how much closer it brings that agreement, not by how convincingly it lets the world postpone it. The failure mode to guard against is not taking small steps; it is mistaking them for arrival.
Common questions.
Incrementalism builds governance piece by piece — voluntary commitments, safety institutes, confidence-building measures, compute controls, a framework convention — letting a regime accrete over time. The comprehensive approach aims directly for a single, ambitious treaty that settles the core questions of thresholds, verification, and enforcement at once. They are two strategies for reaching binding AI governance, each with a serious case and a serious weakness.
That comprehensive treaties are slow, hard to ratify, and easy to block, while the technology advances now. Incrementalism takes achievable steps immediately, reducing near-term risk and building the trust, norms, and institutions that stronger agreement later requires. It also lets governance learn what works, and it avoids staking everything on a single negotiation that might fail. Most successful regimes, like the ozone and climate treaties, grew this way.
That incrementalism can become a permanent substitute for real constraint, letting everyone claim progress while the core problem — the race to build uncontrollable systems — goes unaddressed, and that the timeline may not allow slow accretion if dangerous capabilities arrive within years. A comprehensive treaty can also design interdependent elements (thresholds, verification, enforcement, participation) to fit together, avoiding the gaps of a patchwork.
The framing is a false binary. The coherent strategy is to pursue incremental measures as the means and a comprehensive binding agreement as the end — using near-term steps deliberately as building blocks toward the treaty rather than substitutes for it. The key is to keep the binding, verified agreement as the explicit goal, so incremental progress is judged by how much closer it brings that agreement, not by how well it lets the world postpone it.