A DeepMind Veteran Warns the ‘AI Arms Race’ May Be the Wrong Story

The phrase “AI arms race” has become a default way to describe the scramble among governments and tech companies to build the most capable models. But Verity Harding, a former senior policy executive at Google DeepMind, argues that the metaphor is not just misleading — it may be actively shaping the future in the wrong direction.

In a new essay collection she curated, Harding and contributors from politics and academia challenge the idea that artificial intelligence development should be understood as a zero-sum contest between the United States and China, or between rival labs such as OpenAI and Anthropic. In her view, the language of war narrows the policy debate, encourages nationalism, and makes cooperation harder precisely when it is most needed.

Harding’s warning carries unusual weight. Between 2016 and 2020, she spent years briefing political leaders around the world, including Barack Obama and Emmanuel Macron, while leading global public policy at DeepMind. That work placed her at the center of early attempts to explain AI’s promise and dangers to governments that were still trying to understand the technology’s implications. She now says the conversation has shifted from international coordination to geopolitical rivalry — and that shift, if left unchecked, could produce worse and less widely shared outcomes.

Her argument lands at a moment when AI policy is becoming more nationalistic, export controls are tightening, and the largest companies are racing to outdo one another with ever more powerful systems. Harding does not deny that competition exists. Her concern is that calling it an arms race encourages policymakers to treat collaboration as weakness, when in practice many of the most important challenges — safety, security, supply chains, energy use, access to talent and critical minerals — are too interconnected for any one country to solve alone.

Why the arms-race metaphor took hold

Harding says the popularity of military language around AI is easy to understand. It offers clarity in a confusing field, especially for leaders who want a simple way to grasp what is at stake. But that simplicity comes with a cost.

She argues that the metaphor compresses a complicated technological and political landscape into a binary contest, one that makes every development look like a strategic move against an adversary. In her view, that framing is seductive because it is emotionally powerful and easy to communicate. Yet it can also trap decision-makers inside a narrow worldview.

The shift, Harding suggests, did not happen overnight. In the earlier years of AI governance, the dominant mindset was that the technology was exciting but needed to be handled through international cooperation. That approach treated safety, standards, and oversight as shared problems. Over time, however, the discourse hardened. AI increasingly came to be seen through the lens of civilizational competition, especially between the West and China.

According to Harding, two forces helped drive that change. One was genuine fear: some policymakers and experts came to believe AI would be dangerous unless democratic countries held the lead. The other was more political and commercially useful. For critics of regulation, invoking China became a way to resist rules at home. The argument, in effect, was that any constraint on domestic AI developers would hand an advantage to a rival power.

How ChatGPT changed the conversation

Harding points to late 2022 as a turning point. The release of ChatGPT drew public attention to AI in a way that previous systems had not. Suddenly, a technology once discussed mainly by experts was on the agenda of journalists, lawmakers and business leaders around the world.

But she also argues that the timing mattered. ChatGPT arrived in a world already primed for strategic anxiety. The COVID-19 pandemic had disrupted assumptions about borders, trade and interdependence. Russia’s invasion of Ukraine made geopolitical conflict feel newly tangible again, while questions about AI and military use became harder to dismiss as abstract.

That combination, Harding says, made it easy for the “AI arms race” concept to spread quickly. Once the phrase took hold, it began to be mapped onto historical analogies that many policymakers already understood, especially the Cold War and the nuclear competition between superpowers. The result was a narrative that felt familiar, urgent and dangerous — and therefore politically useful.

Yet Harding believes that analogy can also distort reality. AI is not a bomb or a missile. It is a broad-purpose set of technologies that touches medicine, logistics, education, communications, public administration and military systems alike. In her view, describing it only as a weapon reduces public understanding of the ways it can be built, governed and shared.

What changed inside the industry

Harding’s critique is not limited to governments. She also believes that large AI companies have benefited from the competitive framing, even if they did not create it alone.

As she sees it, the language of an arms race can increase the perceived importance of frontier labs by implying that only a few companies have the expertise to steer such a powerful technology. That, in turn, can reinforce the idea that regulation should be light-touch or deferred, because the stakes are supposedly too high for ordinary political processes to manage.

For Harding, this creates a self-reinforcing loop: the more AI is described as uniquely dangerous and strategically decisive, the more power accumulates around the labs building the largest systems. The more power those labs gain, the easier it becomes to argue that they, rather than governments or broader coalitions, are best placed to define the solution.

She does not suggest that all competition is harmful. Rather, she says the problem is when competition is treated as the only legitimate mode of progress. That logic, she argues, crowds out the possibility of joint safety standards, shared research, and coordinated governance.

Harding argues that once AI is cast as a geopolitical weapon, the debate becomes far more rigid: collaboration starts to look naïve, and every country is pushed to choose sides.

National security, sovereignty and the limits of self-sufficiency

One of Harding’s central points is that the push for technological sovereignty is understandable, but often overstated. She says it is reasonable for Europe and the United Kingdom to want stronger domestic capacity in AI. No major region wants to depend entirely on foreign platforms and infrastructure for a technology with such broad economic and security implications.

At the same time, Harding argues that complete self-sufficiency is unrealistic. Even the United States and China, despite their scale, cannot independently control every input needed for AI development. Chips, rare minerals, energy, cloud infrastructure, frontier talent and manufacturing capacity all sit in a global web of interdependence. That makes the idea of a fully sovereign national AI stack more aspiration than achievable reality.

In practice, she says, countries are already locked in strategic bargaining over chokepoints. One side may control advanced processors; another may control access to minerals, research talent or markets. Those dependencies can be weaponized through export restrictions, procurement decisions or market exclusion. But Harding warns that treating the entire ecosystem as a contest of total separation is a mistake.

She believes the right answer is not blind globalism. Rather, it is a balance between national capacity and international cooperation. Countries should develop their own capabilities, she says, while also preserving the channels that make collective action possible.

The Trump administration and the return of nationalist AI policy

Harding sees recent U.S. policy under the Trump administration as evidence of how quickly the arms-race framing can translate into nationalist action. She points to an executive order on AI that used overtly patriotic language and, in her view, reflected a more insular approach to the technology’s development and deployment.

She also cited the administration’s pressure that effectively pushed Anthropic to remove its latest frontier model from the market, describing it as another sign that AI policy is becoming tightly entangled with national security and industrial competition.

For Harding, these moves are not just isolated policy decisions. They illustrate how the rhetoric of competition can spill into restrictions, market pressures and political symbolism. Once leaders begin assuming that AI development is an arena for national triumph, the incentives shift toward control and exclusion rather than sharing and coordination.

That said, Harding does not frame sovereignty and collaboration as opposites. Her position is more nuanced: countries can compete, build their own strengths and still cooperate where interests overlap. In other words, the challenge is not to eliminate rivalry, but to stop rivalry from becoming the only story policymakers tell themselves.

A coalition of middle powers

One of Harding’s most concrete proposals is the creation of a coalition of middle powers — countries strong enough to matter, but not so dominant that they can dictate the terms alone.

She has suggested that such a group might include Canada, France, Japan, South Korea, India and the United Kingdom. The logic is partly about scale and partly about leverage. Each of these countries brings a different advantage to the table: India offers enormous diffusion potential and population scale; the UK has research talent and a lively startup ecosystem; Canada has important resources such as critical minerals; Japan and South Korea have deep industrial expertise; France brings scientific and diplomatic heft.

The purpose of such a coalition, Harding argues, would be to resist a world in which AI policy is reduced to a two-player game. If mid-sized powers accept that the future belongs only to the U.S. and China, they risk becoming spectators at best and dependent clients at worst.

By coordinating, Harding says, these countries could gain more influence over standards, safety norms, compute access, model governance and trade rules. They could also help ensure that AI development is not defined solely by the priorities of the biggest companies or the largest militaries.

What the essay collection is trying to do

Harding’s new anthology, Reframing the AI Arms Race, is intended to widen the debate rather than end it. By bringing together voices from different disciplines and countries, the project aims to show that the language used to describe AI is itself a form of policy.

Among the contributors are Lawrence Freedman, a prominent historian of war and strategy, and Taro Kono, a Japanese politician known for his engagement with technology policy. Their inclusion underscores Harding’s argument that this is not merely a Silicon Valley problem. AI governance is now a geopolitical issue, a public-policy issue and a democratic accountability issue all at once.

The essay collection’s broader message is that framing matters because it shapes what people think is possible. If AI is described as a race, then winning becomes the primary goal. If it is described as a shared civilizational challenge, the policy toolkit broadens: safety standards, shared testing, public-interest research, cross-border coordination and diplomatic compromise all become more plausible.

Why metaphors matter in tech policy

Harding’s intervention sits in a larger debate about how language influences technology governance. Metaphors are not just rhetorical flourishes; they establish assumptions. If AI is framed as a weapon, then defense ministries and intelligence agencies become central actors. If it is framed as a consumer platform, market competition may dominate. If it is framed as public infrastructure, then regulation, access and resilience matter more.

Harding believes the current metaphor too often narrows the field to military logic. That can create a sense of urgency, but it also risks crowding out the slower work of governance. Security questions, for example, are not solved by outpacing rivals alone. They require standards, testing, red-teaming, incident reporting and international coordination.

The same is true for social benefits. AI systems can help with medical research, education, food systems and disaster response, but only if their deployment is governed in ways that make them reliable, affordable and accessible. A world obsessed with domination may miss those opportunities.

Why Harding thinks cooperation still matters

Harding’s criticism is ultimately rooted in pragmatism. She does not argue that countries will abandon competition or that major powers will suddenly trust one another. Instead, she says the very problems created by AI make cooperation unavoidable.

Security is one example. No country wants advanced AI systems to be vulnerable to misuse, theft or sabotage. Food security is another. Better forecasting, logistics and agricultural planning could all depend on stable and trusted AI tools. Public health is a third. If the technology is to help with disease detection, drug discovery or healthcare administration, then researchers and governments will need ways to share insights without surrendering control.

Harding fears that the current mood — defined by strategic suspicion and nationalist messaging — could erode the habits of cooperation needed to tackle those issues. In her view, collaboration is not a luxury reserved for peaceful times. It is the mechanism by which risky technologies are made safer.

Key facts and context

Timeline: how the AI race narrative intensified

What happens if the race framing wins?

Harding’s bleakest scenario is not a Hollywood-style AI catastrophe. It is something more structural: a world in which centralization deepens, international trust erodes and smaller countries lose meaningful influence over technologies that will shape their economies and societies.

In that future, governments may respond to fear by concentrating power over AI systems. Labs may become more entrenched. Cooperative research could shrink. Nations may harden into blocs that trade access, infrastructure and expertise as geopolitical weapons. Rather than producing safer systems, the race mindset could make them less transparent and less accountable.

Harding also warns that if the world stops practicing cooperation in AI, it may lose the muscle memory needed for cooperation elsewhere. That matters because many of the technology’s most important applications — from health to food systems — require exactly that kind of sustained international problem-solving.

The danger of treating one metaphor as destiny

The central lesson of Harding’s critique is that language can become self-fulfilling. If leaders keep repeating that AI is an arms race, they may eventually build institutions, markets and alliances that behave like one. Once that happens, the metaphor stops being description and becomes prescription.

That is why she keeps returning to the same concern: the story people tell about AI is already shaping the technology itself. And if the dominant story is one of military rivalry and national triumph, then the world may get exactly the kind of AI governance it has been warning itself about — controlled by a few, divided among many, and missing the cooperative solutions that might have made it safer.

Harding’s answer is not to pretend competition does not exist. It is to insist that competition should not be allowed to erase the possibility of shared responsibility. In an era when AI is increasingly entangled with geopolitics, that may be the most important distinction of all.