The most consequential prediction in the current debate about AI is not that the technology is dangerous. It is that the people who agree it is dangerous expect to fail at containing it. Writing in the Financial Times, the economist Martin Wolf reaches three conclusions in a single sentence: AI is definitely dangerous, we should certainly try to regulate it, and the attempt will in all probability fail. For UK organisations, that third clause is the one to plan around. If the external guardrails most leaders are quietly waiting for are unlikely to arrive in any binding form, then the governance you build in-house stops being a compliance nicety and becomes the only control surface you actually have.

The business problem behind a geopolitical argument

Wolf frames AI’s dangers in three categories: a collapse in fundamental human values, a set of huge specific dangers, and widespread economic disruption. He draws the comparison most technologists resist — the atomic bomb — and argues it holds, because AI, like nuclear weapons, could bring great harms whilst being far more general in its reach. His proposed remedy is a technological disarmament pact between the United States and China, on the logic that a treaty between the two state-level competitors is the precondition for restraining the company-level race beneath it.

That is a foreign-policy argument. The business problem hides inside its pessimism. Wolf cites the Hoover Institution’s Niall Ferguson, who describes two simultaneous arms races: one among a handful of frontier companies, and one between the US and China. The US has chosen not to regulate the contest among its own firms, and neither superpower is trying to govern the contest between them. Ferguson’s claim, which Wolf endorses, is that the first race is largely explained by the second — that company competition stays “mafia-like” precisely because the geopolitical race gives every regulator a reason to stand aside.

For anyone running a UK organisation, the implication is uncomfortable but clarifying. The capability you are buying, embedding and depending on is being produced by a competition that no one is currently willing to slow down. The vendors at the frontier are not being restrained by their home government, because restraint is read as unilateral disarmament against China. You should not, therefore, price in external regulation as a near-term risk control.

Strategic Reality: The case for a US–China AI treaty is also a confession that one is unlikely soon. Wolf’s own forecast is failure. Build your AI governance for a world where binding international rules arrive late, partially, or not at all — not for the orderly regime that would be convenient to assume.

What is really happening beneath the treaty talk

Strip away the diplomacy and Wolf is describing an accountability vacuum that organisations are already operating inside. His sharpest point is not about geopolitics at all. It is about responsibility. Companies, he notes, are already legal “persons”, but an AI company might make decisions with no people in the loop. A criminal chief executive can go to prison; there is no equivalent for a model. The Pope’s formulation, which Wolf quotes approvingly, is that AI is a tool — it does not suffer, does not bleed, and cannot bear moral responsibility.

This matters commercially because responsibility does not vanish when it cannot attach to the machine. It flows to the nearest accountable human entity, and in most deployments that is the organisation operating the system, not the lab that trained it. When an automated decision discriminates, when a generative system fabricates a defamatory claim, when an agent takes an action no one authorised, the question “who is accountable?” has a default answer in UK law, and it is rarely the model’s developer. It is the data controller, the service provider, the employer — you.

Critical Context: Wolf’s accountability gap is not abstract. Under existing UK frameworks — data protection, equality, consumer, and sector regulation — liability already lands on the deploying organisation. The absence of bespoke AI law does not mean the absence of legal exposure. It means the exposure routes through laws written before the technology existed.

He extends the point to surveillance, autonomous weapons, deepfakes and scams, all of them enhanced by AI and all of them raising the same question of who answers for the outcome. The provocation he borrows from Javier Milei — Argentina’s “non-human corporation” — and Yuval Harari’s response about an “AI state” are deliberately extreme. But the underlying mechanic is mundane and immediate: integrate a capable system into a process, and you have not outsourced the accountability for that process. You have concentrated it.

The specific dangers that change your risk register

Wolf moves from the philosophical to the concrete, and this is where his argument touches operational risk directly. The danger he singles out is the disruption of cyber-enabled civilisation. He points to Anthropic’s warning about a threat to cyber security, and reasons that because almost everything we depend on runs on such systems, an AI capable of disrupting them would make life radically insecure and the scope for extortion close to unbounded. He names the design of lethal pathogens as a parallel terror.

For most UK organisations the pathogen scenario is someone else’s problem, but the cyber escalation is not. The same frontier capability that writes your marketing copy is lowering the cost and raising the sophistication of the attacks aimed at you. Phishing that no longer reads as phishing, social engineering at scale, and automated vulnerability discovery are not future risks; they are the current direction of travel. The strategic error is to treat AI adoption and AI-enabled threat as separate programmes owned by different people.

Competitive Reality: Your attackers are early adopters too. The asymmetry of AI in security is that offence compounds faster than defence, because attackers face none of the governance, procurement, and assurance friction that slows a responsible organisation. Assume the threat curve is steeper than your adoption curve.

Hidden Cost: Treating “AI strategy” and “AI threat” as separate budget lines owned by separate functions is the most common structural mistake. The same capability sits on both sides of your perimeter. Govern it as one surface, or you will under-resource the side that is moving fastest against you.

The human factor Wolf puts at the centre

What lifts Wolf’s piece above a standard risk inventory is his insistence that the deepest danger is to what humans are, not just what they can do. Humans think, create and act, he writes, and asks what happens when machines do the thinking, creating and acting for us — whether we will still struggle to understand, or become spoon-fed. He frames it as a question of identity: will AI change not just what humans do, but who we are.

Inside an organisation this lands as a question about capability and judgement. When a model drafts the analysis, proposes the decision and writes the justification, the institutional skill of forming and defending a judgement starts to atrophy unless it is deliberately preserved. The risk is not dramatic automation of jobs; it is the quiet erosion of the human competence that is supposed to supervise the automation. An organisation that lets its people become editors of machine output, rather than authors who use machines, hollows out exactly the expertise it will need on the day a model is confidently wrong.

Success Factor: Preserve the human capacity to originate and contest judgements, not just approve them. Teams that use AI to accelerate work they could still do themselves keep their supervisory competence. Teams that use it to do work they no longer understand lose the ability to catch its errors — which is the entire point of keeping a human in the loop.

What UK organisations should actually do

Wolf ends on a note of qualified hope: fear of disruption is rising, the public no longer believes the future should be left to a few tech masters, and that resentment may eventually open room for regulation. He is right that the window may open. He is also right that it has not opened yet, and may open unevenly. The practical posture for a UK organisation is to govern as though you are the regulator of last resort for your own use of AI — because, functionally, you are.

That does not mean building a bureaucracy. It means putting a small number of durable controls in place before they are forced on you, on terms you choose rather than terms a future statute imposes.

Implementation Note: Treat internal AI governance as the control that substitutes for absent regulation, not as preparation for it. The two look similar on paper. The difference is urgency: you are not getting ready for rules that are coming, you are filling a gap that may stay empty.

Priority actions, sequenced by maturity:

• If you are early: Map where AI already touches consequential decisions — hiring, credit, pricing, content, customer outcomes. You cannot govern what you have not located, and shadow adoption means the map is always larger than the official one.
• If you have a foundation: Assign named accountability for each consequential use. Wolf’s whole argument is that responsibility cannot rest with the model; make sure it rests with a specific person in your organisation, documented before an incident, not after.
• If you are maturing: Unify your AI adoption and AI-threat governance under one owner with one risk view. The capability is one thing; stop budgeting for it as two.
• At every level: Preserve human judgement deliberately. Decide which decisions a human must be able to originate and defend without the machine, and protect the skills that make that possible.

Four challenges that will not announce themselves

The non-obvious risks in Wolf’s analysis are the ones that arrive without a headline.

The first is regulatory whiplash. Wolf expects regulation to fail in the near term, but he also expects the public mood to shift. An organisation that builds nothing now, betting on permanent inaction, can be caught flat by a sudden, politically-driven rulebook. Mitigation: build the controls that are good practice regardless of the regime, so any future rule finds you compliant by default.

The second is vendor opacity priced as assurance. Because the frontier race is unrestrained, vendor claims about safety and capability are marketing in a competition, not guarantees backed by a regulator. Mitigation: contract for transparency, audit rights and exit, and assume the assurance you are given is shaped by the race Wolf describes.

The third is the accountability default. The gap Wolf identifies does not stay empty; existing law fills it, and it fills it with you. Mitigation: treat every consequential automated decision as one a named human must be able to explain and defend under current UK law.

The fourth is competence drift. The erosion of human judgement is gradual and feels like productivity right up to the moment it fails. Mitigation: measure whether your people can still do the work the machine accelerates, and treat a “no” as a risk, not a saving.

Reality Check: None of these four will appear as a line item until they have already cost you something. They are the risks of a governance vacuum, and a vacuum produces no warnings — which is precisely why the controls have to be in place before the event, not designed in response to it.

The strategic takeaway

Wolf’s article is a plea for international cooperation that he himself doubts will succeed. Read as policy, it is a warning. Read as operational guidance for a UK organisation, it is unexpectedly clarifying: stop waiting. The disarmament pact that would slow the frontier race is, on Wolf’s own forecast, improbable in any binding near-term form. The arms races he describes will keep producing more capable systems faster than any institution is currently willing to govern them. That is the planning assumption.

The organisations that come through this well will not be the ones that lobbied hardest for regulation or the ones that ignored the risk entirely. They will be the ones that treated the governance vacuum as their own responsibility and built three things whilst waiting for rules that may never come in usable form: clear human accountability for every consequential use, a unified view of AI as both capability and threat, and a deliberate effort to keep human judgement sharp enough to supervise the machines they are deploying. Wolf is asking the world to act. Whilst the world decides whether it can, the only governance you can rely on is the governance you build yourself.

Analysis based on Martin Wolf, “Why the world must agree to regulate AI”, Financial Times, published 10 June 2026. Source: https://www.ft.com/content/8724874e-f387-4848-bfc4-2411d6f0797f. This article represents independent strategic analysis by Resultsense, interpreting the source for UK business leaders. Quotations and attributed positions, including those of Niall Ferguson and Yuval Harari, are drawn from Wolf’s article.