Confidence-building measures, or CBMs, are the unglamorous machinery of adversarial coexistence. They are steps that reduce the risk of misperception, accident, and escalation between rivals who do not trust each other and are not ready to agree on anything binding. They do not resolve the underlying competition. They make it survivable — and, over time, they build the working relationships and habits of verification that a real agreement requires.
The Cold War toolkit
The US and Soviet Union, at the height of mutual hostility, built a surprising amount of this infrastructure. The 1963 hotline gave leaders a direct line to prevent misread signals from spiralling. The 1972 Incidents at Sea Agreement set rules to keep naval encounters from escalating. Later measures required advance notification of missile tests and large military exercises, and the establishment of nuclear risk reduction centres staffed around the clock. None of these limited a single weapon. All of them reduced the chance that fear and misunderstanding would trigger the catastrophe neither side wanted.
What CBMs could look like for AI
Incident notification and shared reporting
Agreements to promptly inform one another — and a shared body — of serious AI safety incidents, near-misses, or dangerous capability discoveries, so that a frightening event does not have to be interpreted through worst-case assumptions in the dark.
Crisis communication channels
A dedicated, tested line between the relevant authorities in rival states for AI-related emergencies, so that a suspected loss-of-control event or an ambiguous capability jump can be discussed directly rather than misread.
Transparency about intentions and red lines
Voluntary declarations of national AI strategies, safety practices, and the capability lines a state considers unacceptable — reducing the misperception that drives arms races, and surfacing where red lines might overlap.
Advance notification of major training runs
A commitment to notify a monitoring body before the largest-scale development efforts, echoing missile-test and exercise notifications, to build the habit of declaration that verification would later formalise.
Scientist-to-scientist dialogue
Sustained technical exchange between safety researchers across borders, keeping open the channel that has historically seeded arms-control breakthroughs even when official relations froze.
Why they matter more than they look
CBMs do three things that make an eventual treaty possible. They reduce immediate danger during the pre-treaty window, when accident and misperception are the acute risks. They build the practical experience of cooperation — shared reporting formats, tested channels, working relationships — that binding verification depends on. And they generate small, reciprocal demonstrations of good faith that make larger commitments negotiable. A state that has honoured incident notifications for years is a more credible treaty partner than one starting cold.
They are also politically cheap. Because CBMs do not require a country to give up a capability, they avoid the sovereignty and competitive objections that sink binding agreements. A government can agree to a crisis hotline or an incident-reporting channel without conceding anything about its right to develop AI. That low cost is exactly why they are the realistic first moves while the harder negotiation is still out of reach.
You do not build trust by demanding it. You build it through small, verifiable steps that cost little and prove reliability. Every arms-control treaty in history rested on years of confidence-building that came first.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
Starting where agreement is possible
The strategic value of CBMs is that they can begin now, among rivals who are nowhere near a treaty, precisely because they ask for so little. They are the natural content of the current US-China dialogue on AI, the summit process, and the safety institute network. Treating them as a serious track — not a consolation prize — is how the pre-treaty period becomes productive rather than merely dangerous. The comprehensive agreement is the destination. Confidence-building measures are the road, and they are open for travel today. The task is to start walking it before an avoidable accident makes the case for cooperation the hard way.
Common questions.
Steps that reduce the risk of misperception, accident, and escalation between rivals who do not trust each other and are not ready for binding agreements. Cold War examples include the US-Soviet hotline, the Incidents at Sea Agreement, and advance notification of missile tests. They do not limit capabilities or resolve the underlying competition; they make it safer and build the working relationships a real treaty later requires.
Practical, low-cost steps such as: notifying rivals and a shared body of serious AI safety incidents and dangerous capability discoveries; a tested crisis-communication channel for AI emergencies; voluntary transparency about national AI strategies and red lines; advance notification of the largest training runs; and sustained scientist-to-scientist dialogue across borders. None require giving up a capability, which is what makes them achievable now.
Because a comprehensive AI treaty is years away, and the interval before it exists is the most dangerous period. CBMs can begin immediately, reduce the near-term risk of accident and misperception, build the practical cooperation and verification habits a treaty depends on, and generate reciprocal good faith that makes larger commitments negotiable. They are politically cheap because they do not require surrendering any capability.
Yes. Throughout the Cold War, the US and Soviet Union built hotlines, incident-prevention agreements, notification regimes, and nuclear risk reduction centres despite deep hostility. These measures did not limit weapons, but they repeatedly reduced the chance that fear and misunderstanding would trigger catastrophe, and they built the relationships and verification experience that made later arms-control treaties possible.