Experts have unveiled a new proposal for the governance of artificial intelligence as they argue the debate has moved beyond simply whether to slow the speed of AI development.
The new framework, published in the journal AI and Ethics, introduces the "Four As" – Action, Access, Authority and Allocation – and is the work of Dr Kirti Prakash and Dr Ian van der Linde of the Advanced Computing Research Centre at Anglia Ruskin University (ARU) in Cambridge, England.
The "Four As" framework is intended to preserve competition and innovation in frontier AI development while establishing safeguards around capabilities and uses for AI that could cause severe or irreversible harm.
The framework is based around the key ethical problems and draws lessons from previous attempts to govern powerful technologies, including nuclear arms control, CRISPR governance around genome editing and the 1975 Asilomar conference on recombinant DNA technology.
The researchers describe a slowdown on AI development as an "incomplete" governance objective, as AI capability is not confined to a single company or country, and existing models can be copied and improved without training new frontier models. Geopolitics and commercial incentives add further barriers.
Instead, the researchers suggest proportionate safeguards, meaningful human control, independent assessment, clear responsibility and regular reassessment as AI capabilities change.
Through the "Four As", their framework tackles whether humans can effectively constrain consequential AI action (Action), who should be permitted to obtain or exercise potentially harmful AI capabilities (Access), which consequential judgements may be delegated to AI (Authority), and how AI related benefits, burdens and power should be distributed (Allocation).
Rather than treating AI risk as a single problem, the framework identifies which "A" the ethical obligation arises under and matches it to the appropriate form of governance, with developers primarily implementing Action safeguards, regulators governing Access, deploying institutions governing Authority, and governments and public bodies addressing Allocation.
The framework also asks whether interventions have knock-on effects by simultaneously creating a different ethical obligation elsewhere. For example, restricting access to a powerful model may reduce misuse risk and security while concentrating capability and decision-making power among a small number of firms or institutions, creating a new allocation concern.
Dr Kirti Prakash, Senior Lecturer in AI and Applications at Anglia Ruskin University (ARU), said: "The debate around frontier AI is often framed as a choice between accelerating development and slowing it down. We argue that this is too blunt. Before deciding how strongly to intervene, we need to ask what exactly the ethical problem is.
"Losing control of an autonomous AI system is fundamentally different from deciding who should have access to a dangerous capability, whether AI should make a medical or military decision, or whether AI development concentrates power in a handful of organisations. Different problems require different forms of governance."
Co-author Dr Ian van der Linde, Associate Professor in the School of Computing and Information Science at Anglia Ruskin University (ARU), said: "Effective AI governance will require more than a checklist of general ethical principles. The Four As provide a way of moving from identifying an ethical problem to asking what constraint is required, who should enforce it, what would signal that the safeguard has failed and what corrective action should follow.
"Recent UK policy provides a concrete example of this distinction. The newly announced AI Biosecurity Lab will use frontier AI to strengthen biological security while relying on trusted access to advanced models from companies including OpenAI, Anthropic and Google DeepMind, illustrating how governance may focus not simply on whether a capability exists, but on who can access and deploy it."
The new "Four As" framework is published as international concern is growing about how rapidly frontier AI could develop. A recent study from the Cambridge Programme on AI Science & Policy and the Centre for the Governance of AI looked at whether increasing automation of AI research itself could trigger an "intelligence explosion", compressing years of technological progress into months or less.
Dr Prakash added: "The possibility of AI increasingly contributing to the development of its own successors makes the governance question even more urgent. If capabilities begin advancing faster than institutions can adapt, simply asking whether AI is 'safe' will not be enough. We need to know who is in control, who has access, where decision-making authority lies and who ultimately gains or bears responsibility for the consequences."
