The danger is the race, not the robots
“I don’t worry about a robot takeover. At least, it’s not in my top five.” That is Oxford professor Michael Wooldridge, in a long interview with Steve Rose for the Guardian, published on 20 May. The headline line is the reassuring one. The more useful part of the interview is his explanation of what does worry him, and it comes from game theory, the subject of his book published that month. A handful of rich companies are racing to build AI while saying they fear it could go badly wrong, and none of them stops, he says, because each believes that if it steps back, a rival will carry on.
Strategic Insight: Wooldridge’s account moves the AI risk debate away from what the machines might do and towards how the companies building them are incentivised to behave. That second question is one a UK reader can test against real decisions.
The interview ran five months ago. One of the UK assets he singles out, NHS patient data, is bound up in a government contract decision due by December.
Why a sceptical scientist describes the race as a trap
A prisoner’s dilemma with a handful of players
Wooldridge has worked in AI for three decades and won the Royal Society’s Faraday prize in 2025 for explaining science to a lay audience. In the interview he says that, if he could, he would slow AI development down “just so that we have more time to understand what’s going on”. He then explains why nobody will: the situation is a classic prisoner’s dilemma.
In that parable, two prisoners would both be better off keeping quiet, but neither can trust the other, so the logic of the game pushes each to confess. Applied to AI, the Guardian’s write-up spells out the result: competition between the firms drives ever more spending, resources and power-hungry datacentres, with no net benefit to humanity. Wooldridge’s own summary is blunt. The companies are “busy pursuing AI, while at the same time saying that they are afraid that something’s going to go horribly wrong with it”, and they keep going “because they think if we back down and we don’t pursue it, somebody else will”.
Critical Context: This is a structural argument, not an accusation of bad faith. In a prisoner’s dilemma each player acts rationally on its own terms, and the collective outcome is still poor. Appeals to any single company’s good intentions do not change the incentives.
How the race got started
Wooldridge’s account of how the industry got here is unglamorous. He says “the core techniques that drove the current AI revolution were invented by the mid-80s”, and that what was missing was computing power and data. GPT-3’s breakthrough in 2020, in his telling, rested largely on a bet by OpenAI that doing the same thing ten times bigger would work. He was among the many who doubted it, and he concedes that the bet paid off.
That history explains why the race is about resources. If progress comes mainly from scale, the players with the most compute and the most data set the pace.
| Point | What Wooldridge says | Why it matters |
|---|---|---|
| Robot takeover | Not in his top five worries | Existential-risk talk is not where he puts his concern |
| A “Hindenburg moment" | "Entirely plausible” given how far AI is embedded in computing infrastructure | Shifts attention to failure in systems people rely on |
| Training compute | GPT-3 needed “20,000-odd AI supercomputers”; Oxford has “probably a couple of hundred” | A leading university’s compute is a fraction of one model’s training run |
| Wikipedia’s share of GPT-3 training data | 3%, by his account | Data, not ideas, is becoming the limit |
Data is the bottleneck, and the NHS holds a lot of it
Where the next tenfold comes from
Wooldridge names compute and data as the limits on progress, adding that “data is now a real constraint”. If all of Wikipedia was only 3% of GPT-3’s training data, he asks, where does ten times more come from next time?
His answer points directly at a UK institution. “The NHS is sitting on a huge amount of data about human beings. That’s the most valuable kind of data imaginable.” Private companies would pay heavily for it, he says, “but I suspect that whoever signed off on such a deal would live to regret it”. He also imagines a dystopian future in which using the NHS depends on agreeing to wear monitoring technology.
⚠️ Warning: The wearables scenario is a thought experiment, not a forecast. The point underneath it is simpler: if data is the binding constraint on AI progress, then whoever holds large datasets about people holds something AI developers want.
What the NHS says about its data today
The live UK case is not a sale of patient data. It is the NHS Federated Data Platform, provided by Palantir under a £330 million contract awarded in 2023. According to AFP, the NHS says sensitive patient data stays in the UK, belongs to the health service rather than Palantir, and is used neither commercially nor for training AI. Doctors’ groups and campaigners want ministers to trigger a break clause so the deal is not renewed past March 2027, and the British Medical Association has told doctors not to use the platform. AFP reported in October that the contract is under review, with a ministerial decision expected by December. We covered the pressure on the Palantir deal at the time.
Wooldridge does not mention Palantir or the platform. The connection is ours: his argument is a reason to watch the terms on which any company gets near NHS data.
Reality Check: The NHS position is that its platform data is not used to train AI. The commercial pressure Wooldridge describes would show up as pressure to change that position, so the commitment is the thing to watch, not the platform itself.
Who loses when Silicon Valley owns the story
Wooldridge is also frustrated with the public conversation. “We have seen the narrative stolen by Silicon Valley,” he says, describing a version of AI that “certainly me and an awful lot of my colleagues have no interest in promoting or building”. He calls it “kind of depressing” for someone who has “spent their career trying to build AI to make a better world”.
His complaint is that large language models “suck all the oxygen out of the room”, hiding a wide range of benefits that go unnoticed. His example is from Oxford: a team building an AI-assisted tool to read a heart scan taken with basic ultrasound equipment and passed to a GP by mobile phone. He describes this as “the kind of expensive stuff that the NHS struggles to provide, all of a sudden available at negligible cost”.
Two UK groups sit on either side of this argument:
- University researchers, who by Wooldridge’s numbers have a fraction of the compute of the largest companies, and who he says have watched those companies take over the story of what AI is for.
- NHS leaders, who hold what he calls “the most valuable kind of data imaginable” and are also potential users of the cheaper, narrower tools he says get overlooked.
Competitive Reality: By Wooldridge’s figures, the whole of Oxford has around 1% of the machines one model took to train. His heart-scan example points to a different kind of value, where AI is applied to a specific clinical problem and the NHS is the user rather than the supplier.
What remains unresolved
Who changes the game
Earlier in the interview, discussing the game of chicken and the US-Iran conflict, Wooldridge describes one way a damaging game can shift: an outside party offers one side a reason to act differently. He does not apply this to AI. If the AI race is a prisoner’s dilemma, though, the same logic suggests that it will not end through the players’ own restraint. Whether any government is willing or able to be that outside party is a question the interview does not answer.
Whether the scale bet keeps paying
Wooldridge thinks the expectations of human-level AI within a few years, from figures such as Sam Altman and Demis Hassabis, are over-optimistic. His test is practical: a chatbot will discuss quantum physics with you in Latin, yet no AI can enter a house it has never seen, “locate the kitchen and clear the dinner table”, which a minimum-wage worker can do.
What a “Hindenburg moment” would look like
He calls an AI-related disaster comparable to the Hindenburg crash “entirely plausible”, because “computer programs go wrong in all sorts of ways” and society relies on computing infrastructure with AI increasingly built in. We reported his February warning about a Hindenburg-style disaster. The interview does not say what such a failure would look like or who would be accountable.
What to watch
- The Palantir decision, expected by December. A decision to use the break clause would show ministers putting control of NHS data ahead of continuity. A renewal on unchanged terms would leave the current commitments on commercial use and AI training as the main safeguard.
- The March 2027 break point. Whatever ministers decide, the wording of any new or renewed contract on AI training will show whether the commitment AFP reported survives.
- Any NHS data partnership with an AI developer. Wooldridge’s warning is about deals, not platforms. An agreement giving a model developer access to NHS data for training would be a direct test of whether whoever signs it lives to regret it.
Wooldridge’s interview is useful because it refuses both the doom and the hype. His worry is not that AI will turn on us. It is that a handful of firms are locked in a race none of them will step back from first, at a time when data is running short and the NHS holds what he calls the most valuable kind. The decisions that follow from that are British ones, and one of them is due by December.
Source: ‘I don’t worry about a robot takeover’: AI expert Michael Wooldridge on big tech’s real dangers (and occasional blessings) (The Guardian, 2026), interview by Steve Rose. NHS Federated Data Platform details from AFP via France 24 (October 2026).
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching how ministers settle the Palantir contract, and whether any deal opens NHS data to AI training. Read more analysis at Insights, or get in touch.