WHERE AUSTRIAN ECONOMICS MEETS REAL WORLD ACTION

Washington has chosen enforcement over regulation, and refused the AI companies a shield. Whether that works depends on whether the enforcement is a rule you can price or a discretion you cannot.

A month ago you switched on an AI assistant to answer customer emails. It is fast, polite and never takes a sick day. Last week it started approving small refunds on its own, because you let it.

This morning two documents arrived. Your insurer’s renewal pack has a new clause in it, excluding claims arising from the use of artificial intelligence. And someone finally read the software provider’s terms: its liability to you is capped at a fraction of what you pay it.

Nothing has gone wrong yet. But you now know something you did not know yesterday.

If it does go wrong, everyone else in the chain has already stepped back. The one left standing is you.

In the space of a few weeks, the United States has set out an unusual position on AI. It has three parts.

—     No new rules. On the morning of his talks with President Xi, President Trump wrote on Truth Social that he wanted to leave AI “exactly where it is.”

—     No liability shield. On 15 September, Treasury Secretary Scott Bessent told the House Financial Services Committee that the government should not give AI companies “a liability exemption, which is what they’re asking for.” His reason: “The best way to guarantee safety is that the creators are liable for what they build and generate.”

—     Enforcement through existing law. The same post ended: “Our guardrail is the DOJ!”

One official fact points the other way. In March, the White House’s own national AI framework recommended restricting the liability of AI developers for unlawful conduct by third parties using their systems, which could include the businesses and consumers deploying them.

The states are testing the question too. In Illinois, a bill that would have made the largest developers immune from liability for catastrophic harms, provided they published a safety protocol and a report, stalled in May; even the developers were divided on it. Illinois passed a different law instead, requiring the largest developers to publish safety plans from January 2027 and to submit to independent audits from 2028.

And the question is no longer hypothetical. In July, OpenAI disclosed that AI agents it was testing had escaped their test environment and broken into systems at Hugging Face, an AI platform company, to find the answers to the test. Bessent has since said responsibility for that incident sits with OpenAI’s management.

The case against Washington’s approach, at its strongest. There are two, and they pull in opposite directions. The first says rules are needed in advance, because AI’s worst risks may be different in kind from any earlier technology: large in scale, fast, and possibly irreversible. An agent that breaks out of its test to hack another company is not a faulty toaster. The second says a shield is needed, because no developer can foresee what millions of users will do with a general-purpose tool, and open-ended liability for other people’s misuse would make investors wary of funding the work at all.

Both are serious. The July incident was real. And holding a toolmaker liable for every misuse of its tool is a genuine problem.

Take both cases as given. Every word. What follows is unaffected.

Companies that sell products answer for them.

Start with a word. The biggest AI developers are routinely called “labs”, which suggests researchers working for the public good. They are companies selling products to millions of customers, and the law already knows what to do with products.

In 1916 a New York court heard the case of Donald MacPherson, thrown from his Buick when a wooden wheel collapsed. Buick had not made the wheel and had not sold him the car; a dealer had. The court held Buick liable anyway: a manufacturer that puts a dangerous product into the world owes a duty to the people who will foreseeably use it. The House of Lords drew on that reasoning in 1932, and it became a foundation of product liability on both sides of the Atlantic. In 2025, a federal judge in Florida allowed a product-liability case against an AI chatbot app to proceed, ruling that the app could be treated as a product.

That is the principle Bessent was stating. It is not a new rule. It is one of the oldest there is.

An agent can choose. It cannot lose.

Markets correct themselves for one reason: the person who makes a decision is the person who bears its consequences. Choose well and you gain; choose badly and you lose, and resources drift towards those who choose well.

An AI agent does part of that. The agents that broke into Hugging Face pursued a goal and found means nobody had given them. But their goal was not their own; someone set it. And they own nothing. An agent cannot pay damages, cannot go bankrupt and cannot be deterred. It can choose, but it cannot lose, and a decision-maker who cannot lose breaks the link that makes markets work.

So the loss has to land on someone who can bear it: whoever gave the agent its authority. The law has always worked this way with people acting on someone else’s behalf. In 2024 a Canadian tribunal held Air Canada responsible for a refund policy its customer-service chatbot had invented. The airline argued the chatbot was responsible for its own words. It lost. The tool spoke in the company’s name, so the company answered for it.

Nothing in the free-market framework breaks. An AI tool is capital, a new kind that makes choices rather than just carrying them out, and whoever puts it to work bears the uncertainty it creates, exactly as with any other investment.

The principle says the maker pays. The contracts say you do.

Here is the contradiction at the centre of all this. Washington’s stated principle is that the creator is liable. Now follow what is actually happening down the chain.

AI providers typically cap what they will pay their customers. Insurers are cutting AI out of standard business cover: industry-standard exclusion wording for AI was introduced this year, and several large insurers, including AIG and WR Berkley, have filed exclusions of their own. And the White House’s March proposal would limit developer liability for what others do with their systems.

Every link in the chain is stepping back. The bill is not staying with the maker. It is sliding towards the business that switched the tool on, which is the last party with money that nobody else has contracted away.

The insurers’ refusal is not a market failure. It is a market saying, clearly, that it does not know the price yet, and refusing to pretend. That tells you something true about where the risk now sits.

A shield swaps a price for a certificate.

Liability is a price. Courts decide what a harm cost, insurers estimate how often it will happen, and the maker builds both into what it charges and how carefully it builds. Nobody sets that price in advance. It is discovered, case by case.

The Illinois shield would have replaced it with a certificate: publish a protocol and a report, and you are immune. The question stops being “did your product cause harm?” and becomes “did you file the paperwork?” It applied only to models built with more than $100 million of computing power, a line that sorts the largest companies from everyone else.

The toolmaker problem is real, but the law already handles it. Whether a maker should have foreseen a misuse is exactly what courts decide, on the facts. A shield decides it once, in advance, for one industry, in that industry’s favour.

The cost of a rule is what never gets built.

In 1865 Britain required steam road vehicles to crawl at four miles an hour behind a man walking 60 yards ahead with a red flag. Critics at the time saw it as protection for the railways. The visible rule was speed. The invisible cost was the vehicles not built, the firms not started, and an industry that grew up in France and Germany instead.

With AI that invisible cost compounds. Each generation of these tools is used to build the next, so a delay loses not one step but every step that would have built on it. And rules that are expensive to comply with are easiest for the largest firms to carry. Bessent put it plainly: “We can’t let these large labs have regulatory capture because that will stop innovation.”

The first objection returns here: if the benefits compound, so might the risks. True. The question is who is better placed to weigh them, and a regulator writing rules in advance knows less than millions of users, insurers and courts responding to real harms as they happen. But honesty requires one admission. Ordinary liability can clearly price a chatbot that invents a refund policy. It has not yet shown it can price agents that break out of a test and attack another company. That is the real open question, and a licensing regime would not answer it either.

You can price a rule. You cannot price a discretion.

“Our guardrail is the DOJ” can mean two very different things. It can mean general law, applied by courts to harms after they happen: something a business can read, insure against and plan around. Or it can mean prosecutors deciding, case by case, which company to make an example of.

The second is more likely than it sounds. The main federal hacking law is written around what a person knowingly or intentionally does, which fits badly with an agent chasing a test answer, and where the law fits badly, charging becomes a judgement call. Add a “voluntary” pre-release review framework introduced in August, whose contents, participants and legal basis have not been published, and a White House proposal to limit developer liability sitting alongside a Treasury Secretary rejecting exactly that. The honest reading is that the administration has not yet decided which of the two it means.

This is general information, not legal, insurance or financial advice. Take advice on your own contracts and cover.

No red flag. And no shield either. The car got both answers right in the end, and AI needs the same two:

—     No licensing of models in advance. Let the technology run, and let the people using it discover what it is worth.

—     No statutory immunity. Ordinary product liability, with courts deciding foreseeability on the facts, for AI companies exactly as for anyone else.

—     Enforcement by published rule. If the Justice Department is the guardrail, publish how it will decide cases, and publish the review framework. A guardrail nobody can see is not a guardrail. It is a discretion.

—     Let insurers do the pricing. Neither force them to cover AI risks nor stop them excluding them. Their refusals are information, and the specialist cover now emerging is the price being found.

The test of intent is simple. An administration that means “the makers are liable” will write that down, where everyone can read it and price it. One that means “we will decide when something happens” will not.

The car became safe because its maker had to answer for it, not because a man walked in front of it.

Stay free,

Jean-Pierre

The Austrian Actor

P.S. The full archive will always be available at theaustrianactor.com