Verity Harding, who led global public policy at Google DeepMind from 2016 to 2020, argues the dominant framing of AI development as a US-China arms race is both inaccurate and self-fulfilling. In an interview tied to her new essay anthology, Reframing the AI Arms Race, Harding says the metaphor concentrates power in a handful of frontier labs, forecloses international cooperation, and forces smaller economies into vassal positions behind one superpower or the other. She spent four years briefing world leaders including Barack Obama and Emmanuel Macron on the trajectory of the technology.
Harding's central claim is that language shapes policy, and the language chosen for AI has hardened in one direction since late 2022. She describes the shift as a move away from the collaborative, internationalist frame that governed early frontier research toward a civilizational binary of the West against China. The anthology gathers contributors including historian Lawrence Freedman and Japanese politician Taro Kono to make the case that the framing itself is the policy lever most in play.
“I just think it's a sexy framing. It's one of those things that feels very clarifying, but if you dig deeper, it restricts your thinking.”— Verity Harding, Former head of global public policy, Google DeepMind
She dates the tipping point to the November 2022 launch of ChatGPT, which arrived alongside a bordered pandemic world and the war in Ukraine. Those conditions, Harding argues, made it easy to map AI onto the last arms race in living memory, the Cold War, and to talk about frontier models in the same register as nuclear weapons. Two forces then pushed the framing further: a sincere belief among some policymakers that democracies should hold the keys to a dangerous technology, and an anti-regulatory current that found China a useful bogeyman.
Key facts
- 01Verity Harding ran global public policy at Google DeepMind from 2016 to 2020, briefing leaders including Barack Obama and Emmanuel Macron.
- 02Her new essay anthology, Reframing the AI Arms Race, includes contributions from historian Lawrence Freedman and Japanese politician Taro Kono.
- 03Harding dates the shift toward arms-race framing to the November 2022 ChatGPT launch, which coincided with the pandemic and the war in Ukraine.
- 04She proposes a middle-powers coalition of Canada, France, Japan, South Korea, India, and the UK as an alternative to US-China binary alignment.
- 05In the two weeks before the interview, Donald Trump signed a nationalist AI executive order and Anthropic pulled its latest frontier model from the market.
The commercial stakes accelerated the shift. Harding says the amount, speed, and freneticism of capital rushing into the sector reshaped the rhetoric, though she does not think money alone accounts for what has happened. She points instead to how the arms-race narrative benefits the labs themselves, positioning them as uniquely capable of managing a technology described as uniquely powerful.
“If you regulate us, you let China win.”— Verity Harding, Former head of global public policy, Google DeepMind
Harding sees the current US posture as the clearest evidence that the worst-case scenario is taking shape. She cites the Trump administration's nationalist executive order on AI, issued in the two weeks before the interview, and the administration's effective pressure on Anthropic to withdraw its latest frontier model from the market. The political culture in the United States, she argues, is now the single largest input into how global AI policy turns out.
Her prescription is not a rejection of national capacity. Sovereign AI capability in Europe and the UK matters, she says, but isolationism has become the only driving force in policymaking, obscuring other options. Even the United States and China cannot build the full stack alone, which is why chokepoints around chips, critical minerals, and scientists keep multiplying. A fully sovereign AI stack for every country, she argues, is not a realistic ambition.
The counter-proposal is a middle-powers coalition. Harding sketches a group of Canada, France, Japan, South Korea, India, and the UK, each contributing a distinct asset: India for scale and diffusion, the UK for talent and startups, Canada for critical minerals. The point, she says, is leverage, and the refusal to accept a binary board on which every other country is a chess piece.
“Talking about AI as an arms race accrues power to them, by saying it's so powerful, new, and unique, that only we have the answer, and only we should be in charge of the solution.”— Verity Harding, Former head of global public policy, Google DeepMind
She reserves particular criticism for the major labs, arguing they are complicit in the shift from cooperation to competition because the framing accrues power to them. If AI is described as so powerful that only its builders can manage its risks, the builders become the de facto policymakers. Harding, who worked inside one of those labs, is explicit that she includes former colleagues in the critique.
The end state she describes is centralization: excessive government control of frontier systems, weaker collaboration on shared problems like security, food security, and disease, and a growing cohort of countries reduced to picking a superpower to align behind. Cooperation, she warns, is a muscle that atrophies when it is not used, and the current rhetoric is not building it.
Harding's argument lands at an awkward moment for the frontier labs. OpenAI, Anthropic, and Google are all publicly positioning themselves as national champions in the US-China contest, a posture that has proven effective at winning federal contracts and easing export-control debates in their favor. The arms-race narrative is, in short, working for them commercially. Any serious middle-powers coalition would have to purchase compute and models from those same US labs, which is why Harding's framing gets applause at policy conferences and limited traction in procurement offices. The interesting question is whether governments in her proposed coalition will build the coordination machinery to negotiate as a bloc before the next model generation locks in dependencies that are much harder to unwind.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
