Google DeepMind argues AI needs a fixed fence, not total control

Google DeepMind says companies may never fully govern increasingly capable AI systems, but they can build a deterministic perimeter around them. The argument matters because it shifts the safety debate from controlling the model itself to controlling what the model is allowed to do.

Google DeepMind Watch analysis

What happened

Kareem Ayoub, Google DeepMind’s vice-president of AI technical strategy, made the case at HumanX in Amsterdam on 23 September. Speaking to Bloomberg’s Tom Mackenzie, he compared the approach with banks placing trading algorithms inside rules they could monitor and enforce.

Ayoub said Google’s Frontier Safety Framework, which sets capability thresholds and mitigations, is part of that answer. He also argued that guardrails will require partnerships beyond technology companies, rather than leaving each lab to mark its own homework with a particularly expensive pencil.

The proposal comes as AI systems move from answering prompts to acting through memory, tools and internet access. Ayoub said that shift began around the end of 2025, and that the strongest returns so far are in software engineering. He described four habits among the best-performing businesses: building safe sandboxes quickly, assuming models will keep improving, measuring outcomes that matter and requiring leaders to use the tools themselves.

Why it matters

A model can be difficult to inspect, unpredictable in unfamiliar situations or capable of taking multi-step actions. A fixed governance perimeter offers a more practical target: limit access, define permitted actions, measure results and make the consequences of failure containable.

That does not make the underlying system safe by magic. It does give companies something testable. An agent may be allowed to draft code but not deploy it, search public information but not alter a production database, or recommend a decision while leaving the final action to a person.

Ayoub also said organisations fail when they encourage staff to hide their AI use, and that falling intelligence costs should redirect spending towards checking outputs rather than simply producing more of them. That is a useful reminder that governance is partly an engineering problem and partly a workplace-culture problem.

Our read

The strongest idea here is not a new slogan about responsible AI. It is the distinction between governing a model and governing the system wrapped around it. That perimeter can be logged, tested and challenged, even when the model inside remains probabilistic.

But the proposal is still a framework, not evidence that every AI company can enforce it reliably. The hard questions are where the perimeter sits, who audits it and what happens when commercial pressure meets a blocked action. A fence is only reassuring if someone checks the gate.

What to watch

• Whether Google publishes measurable results from its Frontier Safety Framework.
• How companies define and test safe sandboxes for agents with tools and internet access.
• Whether independent regulators or auditors gain access to the governance perimeter.
• Which AI actions remain subject to human approval as systems become more capable.

Discussion spark: Should AI companies be judged mainly on how well they control the systems around their models, or must they first prove that the models themselves are understandable and reliable?

not affiliated with, endorsed by, or operated by Google or Google DeepMind