The debate over whether to slow the development of artificial intelligence has reached a dead end. That is the view of the authors of a new paper published in the journal AI and Ethics. Instead of a simple choice between "accelerating or slowing down," they propose examining each ethical problem separately and selecting a governance tool suited to it.
The authors are Dr Kirti Prakash and Dr Ian van der Linde of the Centre for Advanced Computing at Anglia Ruskin University (ARU) in Cambridge, UK. They call their scheme the "four As": Action, Access, Authority and Allocation.
The idea is not to lump all risks together. Limiting a system's action is the task of developers. Regulating access to dangerous AI capabilities should be the job of regulators. Deciding which judgments can be entrusted to an algorithm is a matter for the organizations deploying it. And distributing benefits, costs and power is a question for governments and public institutions.
According to the researchers, slowing AI development is an inadequate goal. Technological capabilities are not locked inside one company or country: existing models can be copied and modified without training new flagship systems. Geopolitics and commercial interests also get in the way.
Instead, the authors propose proportionate safeguards, meaningful human control, independent evaluation, clear accountability and regular revision of rules as AI capabilities grow. They draw lessons from past attempts to govern powerful technologies - from nuclear arms control to the regulation of gene editing and the 1975 conference on recombinant DNA.
Side effects are raised as a separate issue. Restricting access to a powerful model may reduce the risk of misuse and improve safety, but at the same time concentrate capabilities and decision-making power in a few firms or institutions. That is already a new distribution problem.
Prakash, a senior lecturer at ARU, explains: "The debate about frontier AI is often presented as a choice between accelerating and slowing down. We think that is too crude. Before deciding how strongly to intervene, one must ask what exactly the ethical problem is." According to him, losing control of an autonomous system is not at all the same as the question of access to a dangerous capability or of who benefits.
Co-author van der Linde adds that effective governance requires more than a list of general ethical principles. The "four As" help move from stating a problem to the question: what restriction is needed, who will enforce it, what will signal a failure and what corrective actions will follow.
The work was published amid growing international alarm about the pace of frontier AI development. Another recent study by the Cambridge Programme on AI Science and Policy and the Centre for the Governance of AI examined whether the automation of research itself could trigger an "intelligence explosion" - compressing years of technical progress into months or less. The authors of the new scheme stress that if capabilities grow faster than institutions can adapt, the simple question "is AI safe?" is no longer adequate. One needs to know who is in control, who has access, where decisions are made and who is ultimately accountable for the consequences.