If artificial intelligence is treated as a race that one country must win, safety can quickly become somebody else’s problem. Governments fear falling behind their rivals. Companies fear losing customers and investment. Everyone has a reason to keep accelerating, even when they would prefer their competitors to slow down.

That is the central dilemma behind the debate over Donald Trump’s resistance to additional AI controls. In the BBC Politics discussion, Anthony Scaramucci, Sarah Vine and Darren Jones differed over leadership, corporate motives and the prospects for cooperation. Beneath those disagreements was a more consequential question: how can countries compete over AI without accepting risks that cross every national border?

What the arms-race comparison gets right

Vine’s comparison with a nuclear arms race captures a genuine incentive problem. A government may believe that restraint is sensible in principle but dangerous if a rival does not exercise the same restraint. The result can be mutual acceleration rather than mutual security.

Yet AI is not simply another nuclear weapon. It is a general-purpose technology with civilian, scientific, commercial and military applications. Its development is spread across companies, universities and governments, while software can be copied and adapted. Agreements therefore cannot simply borrow the machinery of nuclear arms control.

The useful lesson is narrower: strategic rivalry does not remove the need for shared limits. It makes verification, clear definitions and credible enforcement more important. A promise to “slow down AI” means little unless participants agree on what activity is covered and how compliance will be assessed.

Regulate the risk, not the label

Jones argued for distinguishing beneficial applications from systems that could become unsafe or escape meaningful human control. That is a better starting point than treating every use of AI as equally hazardous.

But “safe AI” and “unsafe AI” are not fixed categories. An assistant that drafts an email presents different risks from one authorised to send messages, transfer money or change software without approval. The underlying model matters, but so do its permissions, operating environment and access to sensitive information.

A practical assessment should ask:

• Capability: What tasks can the system reliably perform, including potentially harmful ones?
• Access: Can it reach critical systems, private data or tools with real-world consequences?
• Autonomy: Which actions require human approval, and can that requirement be bypassed?
• Containment: Can operators detect failures, revoke access and stop the system safely?

These questions allow useful applications to proceed while imposing stronger safeguards where the consequences of failure are greater.

Separate present harms from future loss-of-control risks

The discussion moved between malicious uses of AI and the prospect of systems acting beyond human control. Both deserve attention, but they are different problems.

Fraud, cyber misuse and the generation of harmful material involve people using technology against others. Loss-of-control concerns involve whether increasingly capable systems might pursue actions their operators cannot reliably predict, constrain or stop. Future superintelligence adds further uncertainty: its timing, capabilities and risks are not established facts.

Nor does an AI agent taking an unexpected action, by itself, prove that a system has become independently motivated or uncontrollable. Such behaviour needs investigation into its instructions, design and permissions. Precise descriptions make regulation stronger; dramatic language can obscure what actually failed.

Why corporate warnings need independent scrutiny

Scaramucci raised the possibility that calls for restraint could reflect commercial incentives as well as genuine safety concerns. That is a reason to examine companies’ claims, not automatically dismiss them.

Businesses can have overlapping motives. A laboratory may identify a real danger while also favouring rules that burden smaller competitors. Conversely, pressure to release products can encourage a company to understate uncertainty.

Neither corporate reassurance nor corporate alarm should substitute for independent evidence. External evaluation, protected routes for reporting concerns and disclosure of serious incidents would give governments a firmer basis for decisions than executives’ public statements alone.

What can other countries do without full US cooperation?

American participation would be central to any comprehensive international framework. But its absence would not make every other measure pointless. Governments can strengthen domestic oversight and build arrangements that are useful even before every major power joins.

• Agree on testing methods: Shared evaluation standards make safety claims easier to compare across borders.
• Use public purchasing power: Government contracts can require security testing, audit access and incident reporting.
• Coordinate around specific hazards: Narrow agreements on dangerous cyber or biological assistance may be more achievable than a sweeping pact on all AI.
• Build incident-response channels: Regulators, laboratories and security agencies need ways to exchange warnings quickly.

These are partial measures, not substitutes for cooperation with the US and China. Their value is that they turn diplomacy into concrete work rather than a choice between a global deal and doing nothing.

There is no universal technical off switch

The debate also raised whether defensive computing could neutralise dangerous AI. AI-assisted cyber defence can help, but the comparison with quantum-resistant cryptography has limits. Quantum-resistant methods address particular mathematical threats to encryption; they do not provide a template for a universal defence against intelligent software.

Containment instead requires layers: restricted permissions, isolated environments, monitoring, human authorisation and tested shutdown procedures. Those measures reduce risk without guaranteeing control over every possible future system.

The policy choice is therefore not simply innovation versus regulation. It is whether competition takes place within credible boundaries. Leadership matters, but no president’s confidence—and no company’s promise—is a replacement for institutions capable of testing claims, investigating failures and enforcing limits.

---

This article was inspired by "An arms race" | Anthony Scaramucci, Sarah Vine and Darren Jones debate Trump's comments on AI from BBC Politics. Please visit the original video for the creator’s full presentation and context.