G7 leaders warned of the need to govern AI from the people building it. But is consensus on how really necessary? • City St George's, University of London

Artur d’Avila Garcez, Professor of Computer Science, shares how his recent proposal for a framework on AI accountability may provide insights.

By City St George's Press Office (City St George's Press Office), Published Thursday, 25^th June 2026

At their recent summit in France, G7 leaders sat down to a working lunch with the heads of OpenAI, Google DeepMind and Anthropic. The message from Sam Altman, Demis Hassabis and Dario Amodei was unanimous on one point: G7 governments need to act on the governance of AI, and fast.

However, beyond that observers such as Artur d’Avila Garcez, Professor of Computer Science at City St George’s, University of London, believe that consensus collapsed. He shares:

Garcez argues that the divergence is exactly why his recent framework paper in pre-print matters. He says of the paper, co-authored with Chris Percy, University of Warwick, and entitled, ‘The AI Accountability Ecosystem in the Era of Large Language Models’:

Our proposal doesn't ask governments to pick a geopolitical lane or referee a fight between labs. It asks for something much harder to object to: accountability distributed across the whole AI supply chain - frontier developers, application builders and users alike - backed by continuous outcomes monitoring rather than one-off audits, and assurance mechanisms like benchmarking, certification and incident reporting that don't require waiting for international consensus.

He argues that what makes the framework especially relevant is that it sidesteps the exact fault line the three CEOs got stuck on:

Amodei, Altman and Hassabis were really arguing about control: who gets to set the rules, who gets boxed out, whose infrastructure the rest of the world should run on.

Garcez believes that is part of the problem:

Centering accountability on ‘a bounded actor’, such as a product or machine learning team made sense five years ago, when AI risk sat mostly inside individual applications, but it no longer does. Now, the frontier labs, the application developers building on them, and the millions of end users prompting them are all co-producing the outcomes that go wrong.

Crucially, Garcez argues the mechanisms of the framework he proposed with Percy (benchmarking, certification, incident reporting, contractual assurance up and down the supply chain) don't need a G7-wide treaty or a single dominant coalition to start working.

They can be built through procurement rules, platform policies and market pressure. In other words, where the CEOs' proposals all require the G7 to bet on one institutional structure over another, the accountability framework gives leaders something they can move on immediately, without resolving the harder argument about who should ultimately be in charge.
Whatever position Altman, Amodei or Hassabis lands on next, none of them have a credible argument against transparent supply chains, monitored outcomes and shared responsibility for misuse. Where CEOs disagree on power, liberty or control, promoting accountability in AI shouldn't be controversial at all.

Hashtags