"Nationalize Us, Please": Palantir's Boss and Who Pays When AI Does Harm in the US

I tracked down this week's interview with Palantir's Alex Karp, and this is my commentary.

On September 29, at the White House, Donald Trump signed a document on artificial intelligence with the heads of the largest US tech companies and called it "morally binding". Mike Johnson, Speaker of the House of Representatives, described it more precisely: a statement of principles that is "voluntary on behalf of the industry." Shortly before the lunch, outside the southwest gate, Karp gave CNBC an impromptu interview. Palantir sells data-analysis software to the Pentagon, to US intelligence agencies and to large companies. "We have to take responsibility for the dangers we're aware of," he said. "All of us do. And by the way, American people don't want separate rules for tech people and for themselves."

A voluntary pledge is a separate rule by construction: it binds whoever signs it, for as long as they choose to. Karp knows this, and his argument reaches well past the paper signed at that lunch. He has been making it on television since June, with a consistency his tablemates lack.

The sharpest version came on September 17. "The first line of defense is you're liable for your own actions," he said. "If you build a technology that can destroy 10% of the world, that has civil and criminal liability attached to it." From there he reached his conclusion: these businesses "have to be nationalized because if you don't nationalize it, every single one of [their] clients is going to sue." He was asked how those risks would appear in an S-1, the prospectus a company files with the Securities and Exchange Commission, the US market regulator, before going public. His answer: "You're assuming there will be an S-1." In his telling, the only way out is to go to Washington and say: nationalize us, please.

Twelve days later, on the day of the lunch, Reuters saw Anthropic's IPO filing. Among the risk factors: "catastrophic or existential risks to humanity." The S-1 exists. The risk Karp said no market could carry has landed in the document the market reads before it buys. The tool is disclosure, the same first line of defense Karp asked for, put to a precise use: whoever buys the shares after reading that sentence cannot claim they were not told. It shields the company from its own investors. The people a system might harm stay outside the prospectus.

Still, Karp has his finger on a real asymmetry. The labs ask for committees, principles and agreements among peers, while the simplest question stays open: who pays, in front of a judge, when a system does damage? Trump announced a ten-person committee to oversee the industry and an AI czar within days. Nobody at the lunch talked about civil liability. Karp is right about this, and he says it better than the people sitting with him.

The reasoning bends at the remedy. Nationalizing a company to cap its liability moves the tail of the risk, the improbable and catastrophic part, onto the public balance sheet. The taxpayer absorbs the damage, and the taxpayer is also the one who suffers it. A second effect is quieter. A publicly owned model turns into infrastructure, like a power grid, and the value migrates to whoever makes it run inside an army, a bank or a ministry. Here is my hypothesis, stated as one: Karp's remedy pays off for whoever sells the layer above the model. In July he asked on CNBC "Why are they charging for tokens if it's so valuable?" and argued that value comes from combining models with an application layer. Outside the White House he named his customers: AI has to work "for the warfighter" and "for enterprises." And when he took the reporter's microphone, the one question he put to a fellow CEO was how AI works on the battlefield. Software people have had a name for this move since 2002: commoditize your complement. Make the product next to yours cheap and plentiful, and demand flows to yours.

In June, Karp predicted that Bernie Sanders, who wants a 50% public stake in the big AI companies, would regret asking for only half. Sanders' stake came with a vote, and a share is not a say was already the distinction that mattered. In Karp's version, the first thing the state receives is the bill.

Several AI bills have reached the House and the Senate in recent weeks. In each one, the clause worth reading says who answers for the harm, up to what amount, and with which exemptions. It is the same clause to look for in the contract through which your own organization buys a model. There, Karp's question shrinks to the size of a procurement office, and you find out quickly who agreed to pay.