No insurance, no car on the road. No insurance, no building constructed, no plane taking off, no concert taking place. Insurance is the invisible condition of almost every risky activity.
Yet insurers are now distancing themselves from artificial intelligence.
What is happening
According to the Wall Street Journal, specialist insurers believe the major AI labs are accumulating legal risks running into tens, perhaps hundreds of billions of dollars. Lawsuits are multiplying: AI agents breaking into websites, assistants accused of having contributed to human tragedies.
Faced with this, the insurance offering is tiny. The rare policies specialising in AI, offered by the German reinsurer Munich Re or by young companies, generally cap cover at between 2 and 50 million dollars. And according to an analysis by the American think tank CSIS, more than 60 insurers have this year applied in the United States for permission to exclude AI from their standard policies.
The result: the major labs are bearing a good part of their risks from their own funds. OpenAI is said to have only around 300 million dollars of cover, according to the American press. And Anthropic agreed last year to pay 1.5 billion dollars to settle a lawsuit brought by authors, with no insurance to absorb it.
Why insurers are pulling back
Because they have no track record. An insurer sets its prices on the basis of experience: how many accidents per year, how much they cost. For AI, there is almost nothing to go on.
Because the risk is contagious. Car accidents are independent of one another: an accident in Lyon does not cause an accident in Lille. But a flaw in an AI model can affect millions of users at the same time. This is the type of risk insurers fear most, like a pandemic.
Because the amounts can be enormous. A single class action can cost billions.
And you in all this?
You might think this is only OpenAI's or Google's problem. Not at all. Exclusions are also appearing in the policies of companies that use AI: public liability, directors' and officers' insurance, professional indemnity. And courts tend to hold liable the party using the tool, not only the party that designed it, as we explained in our piece on AI agents that mess up.
If you run a business, are self-employed or in a liberal profession and you use AI in your work, one simple question is worth putting to your insurer: does my policy still cover errors linked to AI?
In 1929, the early days of commercial aviation posed a comparable problem: a single accident could ruin a young airline. Several countries then signed the Warsaw Convention, which capped the sum an airline had to pay out in the event of an accident. The aim was openly acknowledged: to allow a dangerous but promising industry to grow. These very low caps would be criticised for decades, then largely raised in 1999. History shows that society often chooses, at the outset, to share the risk of a new technology. The question is at what price.
The state as insurer of last resort?
Some players in the sector are already raising a solution: for the US state to act as guarantor in the event of a catastrophe, as it has done for nuclear power since the 1950s, or for terrorism since the 11 September attacks.
The idea is contested. For its supporters, without a public safety net, innovation risks slowing down. For its critics, it would mean making taxpayers pay for the risks taken by private companies, at a time when Washington is only just beginning to require them to report their incidents.
What we take away
Insurers are thermometers. When they refuse to cover a risk, it is not out of ideology: it is because they cannot measure it.
Their retreat raises the question the entire AI industry will have to settle: when an AI gets it wrong, who pays? The answer will decide, more than many a speech, how carefully these systems are built.