Researchers Split AI Ethics Into Four Problems, Say 'Slow Down' Debate Is Dead

Two researchers at Anglia Ruskin University in Cambridge say the fight over whether to speed up or slow down AI has stopped being useful. Their fix: stop treating AI ethics as one giant problem and break it into four. Kirti Prakash and Ian van der Linde of the university's Advanced Computing Research Centre call it the 'Four As' — Action, Access, Authority and Allocation — and lay it out in the journal AI and Ethics.

Each A gets its own owner. Action covers whether humans can actually intervene in what a system does, which the authors hand to developers. Access covers who may hold a capability that could cause serious harm, and goes to regulators. Authority asks which judgements may be handed to a model without meaningful human review, landing on deploying institutions, while Allocation covers how benefits, burdens and power get distributed, which they assign to governments.

The four are genuinely separate, the authors argue, and they run an independence test to prove it. Fully solving one doesn't make the others vanish: controlling a system doesn't settle who gets to use it, and legitimate access doesn't settle what decisions it's allowed to make. They also flag a second failure mode they call cross-locus displacement, where cutting risk in one category pumps it up in another — restricting access to a powerful model makes misuse harder while concentrating power in fewer hands.

Prakash isn't shy about the target. 'The debate around frontier AI is often framed as a choice between accelerating development and slowing it down,' he said. 'We argue that this is too blunt.' The paper reads the White House Accord on Super Intelligence — with its internal monitoring, internal oversight, external evaluation and board supervision — as weighted almost entirely toward Action, leaving access, delegated authority and distribution untouched.

Politics and money pull against any global coordination. The authors note Sam Altman, Elon Musk and Demis Hassabis have backed stronger pacing or oversight, while President Donald Trump rejected a substantial slowdown to keep the United States ahead of China, and China has pushed back on proposals it reads as limits on its development. They add that safety rules raising the cost of frontier work may favour the firms already sitting on the most compute and capital — and that calling systems exceptionally powerful can prop up their commercial value. Commercial interest doesn't invalidate genuine risk, they write, but precaution itself redistributes power.

The historical comparisons are deliberately imperfect. Nuclear governance, recombinant DNA and gene editing aren't AI, but each forced a society to decide when capability creates obligations toward people who never chose the risk. Asilomar in 1975 is their anchor: scientists working on recombinant DNA didn't call for abandoning the research, matched containment to risk, shelved experiments that outran existing safeguards, and rewrote the rules as evidence came in.