Stanford: US Should Spend Up to $310 Billion a Year on AI Risk Reduction

Stanford Graduate School of Business has published an analysis concluding that the United States should spend at least $310 billion a year on reducing risks associated with artificial intelligence, and in some scenarios more than 20 percent of the country's GDP. The calculations, by economist Charles Jones, rest on the standard cost-benefit analysis economists have used for decades to value health care and human life.

The starting point of Jones's work is a question economists have long asked in other contexts: how much is it worth paying to lower the probability of a catastrophe. In valuing road safety or environmental regulation, governments have for years attached a specific dollar figure to a human life. Jones applied the same logic to the risk that the development of artificial intelligence could lead to the extinction, or the permanent, severe weakening, of humanity.

Where the 10 Percent Figure Comes From

The estimates are based on the 2024 AI Impacts survey, in which nearly 2,800 researchers working on artificial intelligence took part. The results showed that the median risk of catastrophic AI outcomes for humanity, as estimated by respondents, is 5 percent, but a significant share, between 37.8 and 51.4 percent depending on how the question was phrased, rated that risk at 10 percent or higher. This data has circulated in the AI safety debate for years, but Jones is the first to translate it directly into a specific sum of money society should be willing to pay to reduce it.

In his model, Jones outlines two main worst-case scenarios. The first is a situation in which the capacity to cause mass harm falls into the wrong hands, for example a terrorist group using an advanced model to design a biological weapon. The second is a scenario in which artificial intelligence itself, developing beyond human cognitive abilities, begins acting against the interests of its creators, like the superintelligent alien life form he uses as a metaphor, one that eliminates its host.

Economics Instead of Apocalyptic Rhetoric

Jones stresses that his goal isn't to scare people, but to apply a standard economic tool to a problem usually discussed in terms of science fiction or an ideological dispute between AI enthusiasts and skeptics. The model accounts for both the potential benefits of AI development, such as faster economic growth, and the cost of forgoing that growth in exchange for greater safety.

I can't tell you exactly what the right number is, but the right number is a lot bigger than anything we're spending now - Charles Jones, professor of economics, Stanford Graduate School of Business

You don't have to believe the probability is 90 percent to want to do something about it. Even if the probability is 1 percent, we're willing to take actions that are economically large and significant - Charles Jones, professor of economics, Stanford Graduate School of Business

Gap Between Theory and Budgets

Comparing these figures with reality reveals a huge gap. US government spending on AI safety research is currently measured in the hundreds of millions, not hundreds of billions, of dollars, even though the AI sector itself is attracting investment running into the trillions. Critics of such analyses point out that it's difficult in practice to price the probability of an event that has never happened before, and that Jones's model necessarily rests on numerous simplifying assumptions.

Still, the work fits into a broader trend in which economists and academic institutions are trying to give the debate over AI risk a harder, more quantifiable framework, rather than leaving it solely to computer scientists and commentators. A similar approach has been taken in other papers cited among researchers of existential risk, combining growth economics with models of catastrophe probability.

What It Means for the Public Debate

The key takeaway isn't the specific figure but the approach itself: the conversation about AI safety is ceasing to be purely an ideological dispute between acceleration enthusiasts and supporters of a moratorium, and is becoming subject to the same cost-benefit analysis used to value health or climate policies. The parallel public debate, including recent calls to pause development of the most powerful models, suggests the question of how much should be spent on AI safety will keep coming up more often, as governments and regulators, including in the EU, continue refining their own frameworks for overseeing the technology's development.