POLICY DRIVEN STRATEGIES
  • Home
  • Services
  • About
  • News
  • Contact





The world needs an AI SALT moment

10/5/2026

0 Comments

 
​The world should move faster to realize the benefits of artificial intelligence—and become far more disciplined about how it develops and releases its most powerful systems.

These are not contradictory ambitions. They should be the foundations of a common international strategy.

AI is already being usefully applied in healthcare, education and scientific research. Yet the same advances that expand its benefits can also increase its potential for misuse. The International AI Safety Report 2026 documents growing evidence of AI-assisted cyberattacks, concerns about biological misuse, and limitations in the ability to evaluate and control increasingly capable systems. It also distinguishes documented harms from more uncertain—but potentially severe—future risks. 

The policy challenge is therefore not to choose between innovation and safety. It is to distinguish the widespread adoption of beneficial AI from the development and deployment of capabilities that could create unacceptable risks.

To accomplish both, the world should draw a lesson from an unlikely source: the Strategic Arms Limitation Talks (SALT) during the Cold War.

When the United States and Soviet Union began formal SALT negotiations in 1969, they remained strategic adversaries. Nevertheless, those negotiations produced the 1972 Anti-Ballistic Missile Treaty and an interim agreement limiting specified strategic offensive arms. Competition continued, but the two countries established negotiated constraints on parts of that competition.
The agreements did not depend solely on assurances of good intentions. They provided for technical verification and a standing consultative mechanism to address implementation and compliance. Nor did they eliminate the arms race: the interim agreement restricted certain missile launchers while explicitly allowing modernization within its terms.

The lesson for AI is not that international agreements solve every problem. It is that rivals can recognize a shared interest in limiting dangers, even while disagreeing profoundly about almost everything else.

A country need not abandon its economic ambitions, national security interests or political values to acknowledge that some AI failures could impose consequences well beyond its borders. Nor should cooperation require confidence that every other participant is benevolent. It should require reciprocal obligations, credible scrutiny and consequences for non-compliance.

Waiting for geopolitical trust before negotiating AI safeguards gets the sequence wrong. Carefully designed cooperation should help build confidence—not depend on its prior existence.

AI is a general-purpose technology, not a class of weapons. Its capabilities depend on how models are trained, modified, connected to tools and deployed. The international safety report also cautions that performance in pre-deployment evaluations does not reliably predict real-world behaviour. There is no simple equivalent of counting missile launchers.

A workable agreement should therefore borrow SALT’s principles—reciprocity, verification, consultation and negotiated restraint.

It should also pursue a broader purpose: ensuring that the benefits of AI spread widely while its most dangerous capabilities remain subject to meaningful oversight.

In parallel, governments should invest in AI literacy, workforce transition, secure infrastructure and practical assistance for smaller businesses and public institutions. Universities and other research organizations should help test applications, train professionals and assess whether promised improvements actually materialize.

The objective should be better outcomes, not simply more AI. Organizations should be encouraged to use models appropriate to the task, preserve human judgment and demonstrate value before expanding deployment.

At the frontier, obligations should become stronger as capabilities and potential consequences increase. “Frontier” refers to highly capable, general-purpose models and systems at the leading edge of performance -- not every automated tool or specialized application.

This distinction must not create a regulatory blind spot. A smaller model used in a consequential medical or employment decision still warrants appropriate safeguards. Application-level protections and frontier-model oversight should complement one another.

The principle should be proportionality: enable beneficial uses while placing the strongest obligations where the potential for severe harm is greatest.

The strongest objection to restraint is also the strongest argument for international negotiation: no country or company wants to slow down while a competitor advances.

A purely unilateral appeal for caution leaves that incentive intact. Reciprocal commitments can change it.

The United States and China should be central participants in frontier-AI negotiations, but they should not be the only ones. Other countries have legitimate interests in the security, accessibility and economic consequences of these systems. An agreement governing a technology with global effects should not become a bargain through which a few powerful states and companies determine everyone else’s future.

Nor should a framework allow governments to avoid obligations by delegating development to private firms -- or allow firms to escape scrutiny by relocating. Participating states should implement common minimum requirements through domestic law, with responsibilities extending to relevant public and private developers.

There is already a starting point. The 2024 Seoul Frontier AI Safety Commitments called on participating companies to establish severe-risk thresholds, evaluate their systems and, in extreme circumstances, refrain from development or deployment when adequate mitigations could not keep risks below those thresholds. Those commitments were voluntary.

The next step should be to negotiate which obligations must become independently assessed, enforceable and internationally consistent.

An initial agreement need not attempt to regulate every aspect of AI. It should concentrate on a limited set of obligations that are consequential enough to matter and practical enough to implement.

First, shared thresholds for severe risk. Countries should agree on capabilities and deployment conditions that trigger additional scrutiny, including the potential to substantially enable high-impact cyberattacks, biological misuse or dangerous autonomous actions. Computing power could help identify systems requiring review, but it should not substitute for assessing what those systems can actually do.

Second, independent evaluation and conditional permission to proceed. Developers of systems crossing agreed thresholds should provide qualified evaluators with secure access and evidence that risks are adequately managed. Where that evidence is insufficient, regulators should be able to require additional safeguards, restrict access or pause a specific high-risk development or deployment step. This would not be a blanket moratorium on AI. It would be a requirement to demonstrate sufficient safety before proceeding.

Third, independent verification, incident reporting and crisis communication. Negotiators should develop a combination of confidential disclosures, auditable records, protected evaluations and appropriate infrastructure reporting. They should also establish channels for rapidly sharing serious incidents and resolving suspected violations. Independent verification would necessarily be imperfect; its purpose should be to make significant non-compliance harder to conceal and easier to address, without exposing sensitive information indiscriminately.
​
These obligations should be backed by domestic enforcement and agreed procedures for disputes, corrective action and proportionate penalties. They should also be reviewed regularly.
The international network of AI safety institutes already provides a basis for shared research, testing practices and technical cooperation. That infrastructure should be strengthened to support—not replace—public accountability.

An international safety framework will lack legitimacy if it appears designed mainly to preserve the advantages of today’s leading developers.

Negotiated safeguards should therefore be accompanied by commitments to broaden access to useful AI: support for local-language applications, affordable computing, workforce development and research partnerships. Regulation should preserve room for smaller firms, universities and responsible open-model development rather than making compliance manageable only for the largest companies.

Countries should not be asked to accept constraints while being excluded from the benefits.
Universities have a particularly important role to play. Governments should support their participation in independent evaluation, safety research and the assessment of social and economic impacts. That role should come with secure access to appropriate research infrastructure, disclosure of conflicts of interest and the freedom to publish critical findings.

The UN’s Global Dialogue on AI Governance provides a forum for involving governments, industry, academia and civil society in these wider questions. Such dialogue should help connect technical safeguards with the concerns of countries and communities that might otherwise have little influence over AI’s direction.

A comprehensive treaty will be difficult. That is an argument for starting with achievable agreements, not for postponing negotiations.

Countries could begin with common evaluation methods, serious-incident reporting, reciprocal testing arrangements and direct communication channels. They could then negotiate binding obligations for the highest-risk systems as the technical basis for verification improves. Domestic protections should advance alongside that process, not wait for its completion.

The aim should not be to freeze technological progress or demand that rivals become friends. It should be to prevent competition from making everyone less secure.

We should want AI to help more people learn, discover, work and solve difficult problems. We should also insist that those developing its most powerful capabilities carry obligations commensurate with the risks they create.
​
The world does not need to choose between an AI future and a safer future. It needs the international cooperation to make them the same future.
0 Comments



Leave a Reply.

  • Home
  • Services
  • About
  • News
  • Contact