coordination that only works when everyone agrees is not coordination

the phrase global coordination can conceal an astonishing amount of unfinished work. who is coordinating whom, over which decision, with what information, under whose authority, and with what recourse for the people who object? a room full of institutions announcing a shared commitment may be the beginning of an answer. it is not the mechanism. the mechanism is what happens when one participant believes everyone else is exaggerating the risk, another fears losing its advantage, and a third has no reason to trust the people setting the terms.

cooperation research offers a more useful starting point than demanding that disagreement evaporate. Dafoe and colleagues’ Open Problems in Cooperative AI frames cooperation as a set of research problems involving understanding, communication, commitment and institutions. that agenda helps make the problem tractable: actors may have different interests and incomplete information, and the system still needs to produce better joint outcomes. none of this amounts to a proof that frontier-AI competition can be resolved by a clever protocol. it identifies questions worth working on. Open Problems in Cooperative AI ↗.

the first distinction is between an agreement and a dependable arrangement. an agreement says what participants intend. a dependable arrangement gives them reasons to follow through, ways to check relevant behavior, channels to contest an apparent violation, and responses that do not require improvising an entirely new institution during a crisis. it also acknowledges what cannot be verified. promising total transparency about a sensitive system is often neither feasible nor desirable. pretending that opaque assurances are equivalent to verification is no better.

for ASI Readiness, the necessary threshold is cooperation under disagreement and asymmetric incentives. the impossible part is trying to get there without quietly converting safety into a claim that one group deserves permanent control. smaller organizations, affected communities and people outside the countries building the most capable systems have legitimate interests in the decisions. legitimacy cannot be reverse-engineered from technical competence. the people who can build a system do not automatically possess the authority to choose every risk everyone else must live with.

this is why values cannot be treated as an annoying input error that more intelligence will clean up. some disputes concern facts and can be reduced through evidence. others concern distributions of power, acceptable risk and the kinds of lives people want to lead. a model may help clarify consequences or surface inconsistencies. its competence does not manufacture public consent. our proposed work should preserve room for disagreement, appeal and revision, including the possibility that different communities reasonably choose different limits.

start with something narrower than a world treaty. our proposed first coordination study concerns confidential incident reporting among a small group of consenting organizations. the practical question is whether participants can share evidence of a common failure early enough to improve defenses without creating an incentive to conceal incidents, expose sensitive material, or punish the first organization to report. define the information that is necessary, who may see it, how it is checked, when it can be disclosed, and how a false or disputed report is handled. each design choice creates a different incentive.

then test the arrangement in simulation. introduce unequal costs, delayed disclosures, ambiguous evidence, a participant that withholds information, and a smaller participant with less bargaining power. look for the conditions under which cooperation collapses or produces an unfair outcome. measure the cost of participation as well as the information gained. a protocol that looks successful only because everyone obediently follows the script has failed to study the problem. a result showing that a mechanism does not work under a realistic condition would be valuable if it prevents a ceremonial agreement from being mistaken for protection.

simulation has limits. a laboratory game cannot reproduce geopolitical conflict, establish the legitimacy of an institution, or prove that real organizations will behave as modeled. its role is to expose assumptions and compare mechanisms before a carefully bounded pilot. any claim beyond that needs further evidence. we should be just as hostile to inflated governance experiments as to inflated model benchmarks. an elegant simulation is not civilization agreeing to its parameters.

the forming lab should bring mechanism designers, security researchers, political scientists, negotiation practitioners and people affected by deployment decisions into this work. nobody has to pretend that a common research program requires a common worldview. the invitation is to build arrangements that remain useful when our preferences diverge, our information is incomplete, and the easy consensus ends. if coordination only works after every difficult human problem has already been solved, we have described a fantasy. readiness begins by making cooperation possible before that threshold.

Proposed first work

Design and simulate a bounded incident-sharing protocol with consent, confidentiality limits, verification, dispute resolution and unequal participant incentives. Publish failure conditions before proposing a pilot. Distinguish simulated performance from real institutional adoption.