Geoffrey Hinton, one of the people most responsible for the neural-network techniques that made modern AI possible, has said the quiet part aloud. In a recent Guardian interview he observed that it could turn out we would have been better off never developing the technology at all. He is careful not to call himself a doomer. He still thinks a "wonderful future" of radical abundance is possible. He simply notes that within a decade, possibly sooner, we will know which way the bet resolved: either we managed the risks, or we get massive unemployment, extreme inequality, cultural fracture, and societies pulled apart. He expects a warning shot first, an AI-enabled brownout of power systems or a multi-day banking outage, something large enough to force attention the way Three Mile Island once did.

The claim is striking coming from him. It is also, as you point out, a claim that can be made about a great deal of the modern world. Industrialisation, fossil fuels, the automobile, antibiotics, the internet, nuclear power: each delivered genuine goods to large numbers of people while generating externalities, concentrated power, and tail risks that the people who enjoyed the crumbs did not fully internalise. We kept going because the local incentives were strong, the alternatives looked worse in the short run, and the full bill arrived slowly or elsewhere. Things rolled on until they didn't. Hinton's lament is therefore both specific and generic. The interesting question is whether AI is merely another chapter in that pattern or whether the pattern breaks.

The generic pattern is real. Most transformative technologies have a distributional shape in which the average person receives some consumer surplus, cheaper goods, new capabilities, longer life, while the largest gains and the largest downside risks accrue to a smaller set of actors. The industrial revolution raised living standards over generations while producing Dickensian cities, environmental damage whose costs are still being paid, and political upheavals. The internet gave billions access to information and coordination tools while concentrating economic power, enabling new forms of surveillance and manipulation, and contributing to measurable declines in attention and social trust. In each case one can construct a plausible "better off without" counterfactual for some populations or some time horizons. The counterfactual is almost never decisive, because the world that did not take the technology is not the world that simply stayed still; it is a world that continued to compete with those who did. Once a capability exists, refusing it often means accepting a relative disadvantage.

AI fits this template in several ways and breaks it in others. The template holds in the distribution of benefits so far. Frontier systems have produced clear productivity gains in coding, design, research assistance, and customer service for those who can deploy them. Ordinary people have received free or cheap tools that compress certain kinds of knowledge work. At the same time, the capital intensity of training runs, the scarcity of specialized talent, and the data advantages of a handful of firms have concentrated power more sharply than most previous general-purpose technologies. The risk side is also more compressed in time. Earlier technologies took decades to diffuse and for their second-order effects to become obvious. Current AI systems improve on timescales of months, and the organisations building them are explicitly trying to accelerate that curve. Hinton's "within ten years we will know" is not a rhetorical flourish; it reflects the speed at which capability and deployment are moving.

Where the pattern may break is in the character of the risk. Most earlier technologies had natural or institutional brakes. Nuclear weapons required rare materials and large industrial bases; their use was constrained by mutual vulnerability. Fossil fuels scaled with physical infrastructure that took years to build and could be regulated. Software systems that can write software, that can be copied at near-zero marginal cost, and that can be pointed at a wide range of goals do not have the same brakes. The "Chernobyl" Hinton describes, an attack or accident that takes down power or payments, is still a recoverable event. The harder problem is the possibility of systems that are better at strategic deception, resource acquisition, or recursive improvement than the institutions that are supposed to oversee them. That is not a claim that such systems exist today. It is a claim that the rate of improvement makes the window for effective oversight shorter than the window for previous general-purpose technologies.

The frequent observation that we keep receiving positive crumbs while the overall balance looks negative is therefore both correct and incomplete. It is correct that ordinary people have rarely been the primary authors or the primary residual claimants of these transformations. It is incomplete because the crumbs are not random; they are the mechanism that secures consent and continues the process. Once enough people experience the local benefit: faster research, cheaper services, new forms of entertainment or communication, the political and economic cost of slowing the technology rises. Firms that pause lose market position to those that do not. Countries that regulate more tightly risk falling behind those that do not. The process is not a conspiracy; it is the result of decentralised incentives under competition. It continues until either the negative externalities become large enough and visible enough to generate coordinated restraint, or until some failure mode removes the option of restraint.

Hinton's personal position is consistent with this dynamic. He left Google so he could speak freely. He argues for something like an FDA for frontier models: independent testing and a demonstration that expected benefits exceed expected harms before release. He does not trust the current leadership of the leading labs to impose that discipline on themselves. At the same time he says his gut is that we are still in control if we choose to put serious effort into remaining so. That is the residual hope of the pattern: that the warning shots arrive while course correction is still possible, and that the institutions capable of course correction still function. History offers mixed evidence on that hope. Some risks were managed (ozone depletion, certain nuclear safety regimes). Others were not managed until after large costs had already been paid, or were never fully managed at all.

The counterfactual Hinton raises, that we might one day judge the whole enterprise a net loss, cannot be dismissed by pointing out that the same could be said of earlier technologies. Those earlier technologies did not all turn out to be net losses for the populations that adopted them, even after accounting for their costs. Some did, for some groups. The difference with AI is that the judgment may arrive while the systems are still being built and deployed, and that the systems themselves may be able to influence the conditions under which the judgment is made. Whether that makes the case for slower, more deliberate development stronger or simply highlights the difficulty of achieving it is the practical question that remains open. The process will continue until the incentives change or the option to continue is removed. Hinton's contribution is to say, from inside the process, that the second outcome is no longer unthinkable.

https://www.theguardian.com/uk-news/2026/nov/03/ai-godfather-geoffrey-hinton-better-off-without-ai