International regulation of artificial intelligence

By Matthew Parish, Associate Editor

Tuesday 16 June 2026

A recurrent mistake in discussions about artificial intelligence is to assume that regulation is principally a technical problem. It is not. It is a political, economic and legal problem that happens to involve technology. The challenge is not merely how to control algorithms but how to govern the organisations, incentives and power structures that increasingly deploy them. Any serious international regulatory framework for artificial intelligence must therefore begin with institutions rather than software.

The history of international regulation offers some guidance. Aviation, telecommunications, maritime navigation, nuclear energy and civil aviation all evolved from periods of technological chaos into systems of predictable governance. None of these systems eliminated competition or innovation. Instead they established common standards that reduced risks while allowing technological development to continue. Artificial intelligence requires a comparable approach.

The first principle should be that artificial intelligence is not a single technology. The phrase encompasses everything from recommendation algorithms and medical diagnostic systems to autonomous weapons and large language models. Attempting to regulate all forms of artificial intelligence through a single set of rules would be as irrational as regulating bicycles and nuclear submarines under the same legal framework.

Accordingly an international regulatory architecture should classify AI systems according to risk.

Low-risk systems would include routine commercial applications such as customer service chatbots, inventory management software and recommendation engines. These should face minimal regulatory burdens beyond transparency requirements and basic consumer protection standards.

Medium-risk systems would include applications affecting employment, finance, healthcare, education and legal decision-making. These systems should be subject to mandatory auditing, explainability requirements and regular independent review.

High-risk systems would include military applications, critical infrastructure control systems, advanced autonomous agents capable of independent decision-making and frontier foundation models exceeding specified computational thresholds. These systems should be subject to international licensing and oversight.

At the centre of such a framework there should be an International Artificial Intelligence Agency, modelled partly upon the International Atomic Energy Agency.

The purpose of such an institution would not be to inspect every algorithm in the world. That would be impossible. Instead it would establish international standards, accredit national regulators, conduct inspections of frontier AI facilities, maintain registries of advanced models and investigate major incidents.

Membership would be universal. States would retain sovereignty over domestic regulation but would agree to minimum international standards. Much as aircraft certified in one jurisdiction can generally operate internationally, AI systems meeting internationally recognised standards could move more freely across borders.

The agency’s governance structure would require careful design. One of the principal weaknesses of contemporary international institutions is the concentration of influence amongst a small number of states. Artificial intelligence development, however, is occurring across multiple jurisdictions simultaneously. The regulatory system must therefore balance the interests of major technological powers with those of smaller states that may be affected by AI systems but lack the resources to develop them independently.

A bicameral governance structure may therefore be appropriate. One chamber could represent states on a sovereign equality basis. Another could allocate voting power according to measurable factors such as economic contribution, computational capacity and scientific research output. Major decisions would require approval from both chambers.

A second pillar of regulation should concern transparency.

At present, many frontier AI systems operate as opaque black boxes. Even their developers often struggle to explain precisely why particular outputs are generated. This creates obvious problems when such systems are used in medicine, law enforcement, finance or military planning.

International standards should therefore require developers of advanced systems to maintain detailed records concerning training data sources, model architecture, computational resources utilised and known limitations. These records need not always be publicly disclosed. Commercial confidentiality remains important. However regulators should possess the authority to inspect them when necessary.

A third pillar should address computational resources.

The most advanced AI models increasingly require enormous computational infrastructure. Unlike software code, which can be copied infinitely at negligible cost, frontier AI development remains constrained by access to specialised semiconductors, electricity and data centres.

This creates a valuable regulatory opportunity. Rather than attempting to monitor every researcher, regulators can focus upon major computational facilities. International licensing regimes could require operators of data centres above specified thresholds to register advanced AI training activities and comply with safety requirements.

This mirrors existing approaches in sectors such as aviation and nuclear energy, where oversight focuses upon critical infrastructure rather than every individual participant.

Military applications require separate treatment.

The development of autonomous weapons represents perhaps the most dangerous aspect of artificial intelligence. Human history demonstrates that military innovation almost invariably precedes civilian regulation. Artificial intelligence is unlikely to be different.

An international convention should therefore establish clear principles regarding autonomous weapons. Systems capable of selecting and engaging human targets without meaningful human oversight should be prohibited. Human accountability must remain present throughout the lethal decision-making chain.

Such restrictions would not eliminate military AI. States will continue to employ artificial intelligence for intelligence analysis, logistics, cyber defence and battlefield coordination. The objective is not prohibition but preservation of human responsibility.

A fourth pillar should concern liability.

One of the most significant deficiencies in current AI governance is uncertainty regarding responsibility. When an autonomous system causes harm, determining who bears legal liability can become extraordinarily difficult.

The principle should be straightforward. Responsibility must always trace back to a legal person.

Developers should bear responsibility for design defects. Deploying organisations should bear responsibility for operational misuse. Corporate officers should bear responsibility where negligence or recklessness is demonstrated. Artificial intelligence systems themselves should never possess independent legal personhood.

This principle preserves one of the central achievements of modern law: accountability ultimately rests with human beings and the institutions they control.

The framework must also address economic concentration.

Artificial intelligence development increasingly favours organisations possessing vast financial resources, enormous datasets and access to specialised computing infrastructure. Left entirely unchecked, this dynamic may produce unprecedented concentrations of economic power.

International competition authorities should therefore cooperate closely regarding mergers, acquisitions and anti-competitive conduct involving frontier AI firms. Regulatory policy should encourage interoperability, open standards and portability of data wherever practical.

Perhaps the most difficult issue concerns enforcement.

International law functions imperfectly because sovereign states retain ultimate authority within their own territories. No global police force exists. Artificial intelligence regulation will encounter the same reality.

The answer lies in incentives rather than coercion.

States complying with international standards should enjoy preferential access to global AI markets, research partnerships and computational resources. Non-compliant jurisdictions would face restrictions on technology transfers, licensing arrangements and participation in international research programmes.

Such mechanisms already exist in various forms within financial regulation, civil aviation and nuclear governance. Their effectiveness derives from mutual dependence rather than force.

Ultimately the objective of international AI regulation should not be to slow innovation. Excessive caution would merely shift development into less transparent jurisdictions. Equally, unrestrained competition would create incentives to ignore safety considerations altogether.

The proper balance lies between these extremes.

Artificial intelligence is neither a miraculous solution to humanity’s problems nor an existential threat destined to destroy civilisation. It is a powerful new industrial technology whose consequences will depend largely upon the institutions governing its use. The challenge facing the international community is therefore not technological but constitutional. Humanity must determine how power created by intelligent machines will be distributed, supervised and constrained.

The sooner that discussion begins in earnest, the more likely it is that artificial intelligence will become an instrument of human flourishing rather than a source of geopolitical instability, economic concentration and regulatory fragmentation. The task is formidable. Yet history suggests that when transformative technologies emerge, societies eventually construct the institutions necessary to govern them. Artificial intelligence should be no exception.