Javier Milei’s Free-the-Machines Bill Gets Limited Liability Backwards
Limited liability worked because states built accountability around it; Milei's A.I. proposal starts by removing human responsibility
This is the headline on an editorial that Milei just published in the Financial Times:
The president of a sovereign nation, with the backing of his chief economic advisor, has proposed a law that would allow autonomous entities to command organizations with no human accountability at all.
The reform goes way beyond garden-variety deregulation. Consider: under this law, the entity itself would still be subject to the same civil and criminal penalties that worry human chief executives. But why would an artificial CEO care about criminal penalties? Fear of virtual prison? All it would consider are civil penalties that reduce shareholder value. As long as the benefits outweighed the costs (on a risk-adjusted basis), these autonomous corporations would be free to do … anything.
Why would Milei do this? Well, the real reason is probably Peter Thiel.1 But Milei offers two justifications for the decision:
Technological: autonomous corporations will never exist without enabling legislation;
Historical: the creation of limited liability in the 1600s was a good thing, and he is just following in Dutch footsteps.
Of course, we have to compare the editorial to the actual proposal. Milei likes to say outrageous things, and here he has incentives to play up any “reform.” Do the details of the proposal live up to the presidential hype?
Let’s take these in turn. This post will first flesh out how Milei’s idea, as presented in the FT, would allow robo-corporations to break laws and damage others with impunity. The second part will examine the actual genuine real bill on the floor of the Argentine Congress. The third will provide some historical context. Have we really been here before or is Milei misreading history?
SPOILER: Milei’s solution solves no real problems, unless you think that accountability is a problem. In wrecking accountability, Milei gets the history of limited liability wrong. It is a history of solving practical problems and building institutions to recreate the accountability that had been provided by personal liability. It is not a history of farsighted policymakers giving ever more autonomy to private corporations. Moreover, Milei’s reforms will increase the risk that nonaligned A.I. could go wild — this is a bad time in human history to lessen accountability.
The Terminator as manager
Milei claims that autonomous organizations will not emerge without changes to limited liability.2 That is almost certainly correct. But is that a bad thing?
Right now, the directors and shareholders of a corporation are generally free from criminal or civil liability should the corporation engage in wrongdoing. If Procter & Gamble markets a soap that accidentally turns some people green, then Procter & Gamble is liable for damages and its share price will go down. The executives who authorized the new soap can be punished under civil and (perhaps) criminal law. But the victims cannot directly go after the shareholders or their representatives (in U.S. parlance, the “directors”).
There are a few exceptions. If the directors authorized the wrongdoing, then they are liable. For example, under Delaware law directors, are liable for any decisions that they direct, order, ratify, approve, or consent to carry out. The directors are also liable for any harms that executive wrongdoing may inflict on the company’s creditors or shareholders. They cannot authorize illegal distributions or dividends, obviously. Nor can they breach their fiduciary responsibilities to the shareholders. And they cannot just not even bother oversee corporate executives.
Shareholders themselves are liable in a few even more restrictive situations. If a single entity holds a majority of the shares, then it can be held liable for corporate actions. Ditto for controlling shareholders (even without a majority) or the corporate owners of subsidiary companies. Finally, if you can prove that the corporation effectively acted as an agent of some shareholders (in legalese, that “the veil of corporate responsibility was pierced”), then those shareholders will be liable.
But the above conditions are fairly restrictive and no substitute for the responsibility held by human executives. Milei’s proposal would keep mostly intact the conditions under which directors and shareholders are liable, but allo the human executives to be removed from the loop. Should a corporate A.I. harm people, then the only way to hold a human being to account would be to prove that the directors signed off on the action. Since it wouldn’t be at all hard to train an A.I. to just, you know, not tell the directors about edge-case decisions, that doesn’t sound too effective.
The actual proposal
It wasn’t easy, but I managed to track down the text:
Is the reform as radical as Milei’s editorial implies? Not entirely. It builds in some safeguards to restrict the actions of computer-run companies, but not all that many. Article 14 allows regular companies to become “automated companies” (sociedades automatizadas) operated by artificial intelligences. The corporate entity is still liable for damages. And Article 92 requires that even if you place the company under the control of a “legal person,” you are going to have to designate one responsible “human person.”
But it is still pretty radical.
Should the bill become law, if somebody is damaged by an operational act, the that “responsible human person” can be punished. The problem is that it will be far from clear whether they actually approved the operation, or if anybody did. The directors can then just throw up their hands and let their sacrificial lamb be sacrificed (see page 51 of the bill).
Now, the directors’ immunity isn’t complete. Article 102 is short and reads as follows:
The administrative body can use artificial intelligence systems or algorithms in order to carry out operational functions or make decisions. Their use does not eliminate nor limit the administrators’ responsibility nor exempt them from the duty to configure and supervise the system and its output.
The problem is that if negligence can’t be proven, then there will be nobody at whom to point the finger, neither a human CEO nor the “responsible human person” in the absence of a CEO.
Worse yet, Article 101 of the bill reduces the directors’ liability even if they are negligent. Currently, Article 51 of the Argentine corporate law mandates administrators and representatives to act with loyalty and the diligence of a good businessman. Those who fail are liable unlimitedly and jointly for damages resulting from their action or omission. Under the new law, however, they would only be liable for direct damages and objective liability would be eliminated. That is to say, you would now have to prove intent in order to sue — any administrator who could credibly claim to have acted or failed to act in “good faith” will have nothing to fear.
So imagine that a company trains an artificial “operational CEO” to maximize subscription revenue. The A.I. learns that it can increase profit by hiding cancellation buttons, delaying refunds, and sending misleading renewal notices. The company is still liable for damages. But there is no responsible human being. Why wouldn’t the directors let the A.I. do whatever it wants, as long as the benefit to the company’s bottom line exceeds the direct costs?
But it gets worse. The bill also allows for the creation of “decentralized autonomous organizations” (DAO) where nobody would be in charge. At least in regular corporations, all decisions would go through a human “responsible person” at some point. Article 274 of the current law says that said human is legally accountable to shareholders and third parties for bad performance of office, violation of law or regulation, and damage caused by fraud, abuse of powers, or gross negligence, and Article 274 will remain in force for most companies. The human decision-makers will probably not want to risk their livelihood and freedom in order to allow a computer to misbehave.
But if the company organizes itself as a DAO, then even though Article 260 says that there still has to be a human representative, Article 257 and Article 262 will remove any liability for that human representative’s regardless of bad performance or negligence.
Moreover, under a DAO means that the damn A.I. could be doing God only knows what, with no one around with any incentive to monitor it as long as the quarterly reports look good. I honestly don’t want to think about what that could mean.
The history of limited liability
Milei says he is just following the footsteps of the people who invented limited liability, without which large capitalist enterprises would not exist. But that gets the story exactly backwards. Large existing private enterprises existed before limited liability. Those enterprises then ran into problems. The government and the financiers cobbled together piecemeal solutions which in turn generated new problems. New laws and institutions then solved those problems. It wasn’t at all libertarian in origin and it wasn’t invented in advance of a problem. Rather, it emerged as an accretion of makeshift solutions and it required the state to invent institutions to replace personal liability. It took centuries for limited liability to look like the institution we all know.
Limited liability is a weird creation, when you think about it. The owners of a business aren’t responsible for the actions of that business? What? That’s crazy. How can you align incentives without personal liability?
In the 17th century, the Dutch government and a collection of Dutch merchants found themselves needing to answer those questions. They wanted to fund giant enterprises to go out and do stuff out there in Asia that could generate income. But the Republic didn’t want to finance those enterprises with state money — the Dutch Republic probably couldn’t have done so even had its leaders wanted to. Alternatively, the Dutch could have imitated the Spanish, who financed one-shot expeditions under private partnerships, rewarding success with control over indigenous labor, government posts, or giant blocks of land. That model didn’t seem too attractive either, not least because it required the private backers of the expeditions to physically migrate to the colonies or forego most of the benefits.
The problem was that the risks of the proposed operations proved too large for equity investors. The enterprise that would operate half a planet away, all the while getting into wars with the Portuguese and the Spanish and the English and the local kingdoms and God knows what else. And then, if they succeeded, the enterprises might then do all sorts of crazy and risky things to the local population. Moreover, it might fail to repay debts. If you were going to be held personally liable for all those things, then the risk was just too large to risk large amounts of equity capital. But if there wasn’t plenty of equity capital at stake, then nobody would want to lend to the enterprises. Lenders want owners to have skin in the game!
So in 1602, the Dutch Republic decided that maybe it would help if they increased the expected profits. If you can’t lower the risks, increase the rewards! The government did that by giving one company, the Dutch East India Company (henceforth VOC, for its Dutch initials) a monopoly over the Asian trade and the right to whatever it needed to do out in the East Indies in order to increase that trade.
That might have worked had the monopoly profits been high enough. The problem was that they weren’t. The original plan had been to rely on old-fashioned circulating capital, meaning that investors would get their money back when the enterprise was liquidated. That meant setting an end date on the operation — they picked 1612, ten years after it began.
The problem was the VOC kept losing money, and there was no way it was gonna get into the black by 1612. Instead of generating quick returns, the VOC was a hot mess for decades. In retrospect, this isn’t that surprising, considering as the VOC found itself fighting a bunch of medium-sized wars out in the East Indies. Wars are pretty capital intensive, as the VOC’s management soon discovered.
The VOC’s losses had to be covered if the initial investors weren’t going to lose their shirts. In 1609 the board started casting around for permanent equity capital, but it wasn’t immediately successful. The VOC charter exempted directors from liability for unpaid wages but nothing else. Directors were regularly threatened with debtors’ prison over the company’s debts. Moreover, the directors couldn’t individually shoulder that burden even aside from the risk of debtors’ prison. Their individual reputations weren’t strong enough to sustain debts of the size they needed.
Limited liability emerged in fits and starts as a solution to the problem. In 1613, the directors insured a giant debt issue, ostensibly freeing them from personal responsibility for the loan. That didn’t work so well, not least because many of the VOC’s insurers were also its shareholders.
Therefore, in 1617 the directors declared that they would all be jointly responsible for the VOC’s existing debt. That was still not limited liability, of course. It helped, but the company’s debts kept rising. It wasn’t clear who was responsible for all these new bonds. The directors? If so, which directors, the ones in Middelburg or the ones in Amsterdam? If not, the officials who signed the debt? With the company’s demand for credit rising insatiably, something needed to be done.
In 1623 the VOC took the bold step of unilaterally rewriting its debt contracts to remove all personal liability from the directors. The company felt entitled to do that because in 1621 the Supreme Court of Holland and Zeeland had ruled that the directors were not personally responsible for frauds committed by company clerks, even if the company was still liable. Precedent in hand, the directors made their declaration. The markets accepted it, more or less, although even then more reform proved necessary over the next few decades.3
In short, nobody ever sat down and thought, “What do we need to do to get make it easier to finance large-scale enterprises?” And they certainly never thought, “Hey, we’re expanding human freedom from the state!” Rather, they chartered the companies first and then hammered out the kinks later as problems presented themselves.
And in 1623 you still didn’t have limited liability the way we think of it today. It took until 1862 for the U.K.’s Companies Act to really consolidate the institution. Even then, the Companies Act excluded banks and insurers because Parliament understood the obvious moral-hazard problem: if shareholders could enjoy the upside but walk away from the downside, then they would have few incentives to refrain from taking wild risks and gambling with other people’s money.
The idea spread around the world, but messily, and not always because its implementors saw the economic benefits. Rather, they saw limited liability as a tool that could solve political problems — it provided a cheap way for governments to benefit key supporters. A quarter-century ago I wrote a book about the Mexican example, so let’s start there.4
Porfirio Díaz came to power in 1876. The country was in chaos and most of it was effectively out of Mexico City’s control. Díaz had few resources that he could use to impose order and no one was about to lend significant amounts to the penurious federal government. Moreover, it is hard for a weak state to generate and distribute economic rents — what good are privileges provided by law if the government cannot enforce the law? And Mexico was a weak state; it could barely protect private property at all, let alone extract taxes or other resources. But it needed to generate some sort of benefit that would depend on keeping it in power and that could be used to buy off powerful and violent local interests.
Limited liability provided a useful tool to overcome this problem. Díaz could grant corporate charters to select enterprises, whose operations would be strategically dispersed around the country. Those enterprises could in turn raise capital by selling shares on the Mexico City markets. Board seats and investment stakes would be carefully distributed to local warlords, who would thereby enjoy a share of the profits. This worked because even though the federal government was too weak to protect everyone’s property rights, it could still protect some people’s property rights a little bit better than others. As a result, the enterprises chartered as limited liability companies enjoyed a bit more flexibility and security than others, which translated into a competitive advantage and higher profits — and most importantly, gave their beneficiaries a stake in the Porfirian system.
You might be surprised to learn that the early United States didn’t look much better than Porfirian Mexico. American corporations weren’t havens for John Galt to heroically pursue his capitalist visions. Rather, they were patronage vehicles that state legislatures distributed to select named firms. You didn’t see much impetus to create impersonal general incorporation laws until a wave of state bankruptcies in the 1840s opened up space for reform.
With more incorporation, however, the problem of liability became more acute. As a result, the new general incorporation statutes imposed uniform rules on capital, debt, dividends, reporting, voting, and directors’ liability. When laws and institutions proved insufficient, state legislatures changed the laws. (Read John Wallis and Naomi Lamoreaux on these changes: the story is amazing.)5 In fact, limited liability didn’t fully triumph as a business form used by almost everyone until 1988, when a small IRS rule change meant that LLCs would be taxed the same as partnerships.
Limited liability was a tradeoff. It made shares less risky and more liquid. That in turn enabled the mobilization of more capital for large enterprises. But it also made it harder to monitor those enterprises. Creditors therefore needed other assurances that owners would not strip assets, undercapitalize the firm, or surreptitiously instruct managers to gamble with their money. So over time, governments created rules around boards of directors, fiduciary duties, accounting rules, disclosure, bankruptcy protection, shareholder voting, and, later on, securities regulation.
This is neither a triumphant nor a libertarian story. It certainly didn’t emerge out of the forehead of some brilliant head of government in 1602. And what did emerge in the late 19th century emerged was about a lot more than protecting owners. Institutions designed to increase accountability were a necessary accoutrement if limited liability was to work.
Back to Argentina and artificial intelligence
Milei’s proposed law is not something that the A.I. industry needs. For better or for worse, A.I. is doing just fine without it.6 It is, however, a law that future investors in A.I.-controlled firms may want. The firm as a corporate entity will still pay damages if damages can be proved and collected. But at the end of the day the operational wrongdoer will just be software. The human representative can say the protocol acted automatically while and the directors can claim that they configured and supervised the system in good faith. No one will be accountable.
The shield of limited liability worked because the state built substitutes for the personal liability it removed. Milei’s proposal offers the shield first and barely even gestures toward the substitutes later. The law might get a bunch of A.I.-managed firms to locate in Argentina. And maybe that will produce benefits for the country. But it could also allow for spectacular harms. It certainly increases the risk that non-aligned A.I.’s could drag us to disaster by allowing humans to give the automatons direct control of their enterprises with little fear of any repercussions.
In the end Milei’s autonomous company law is a solution in search of a problem. He invents a fake problem: nobody has yet put a business under the control of a computer because they will be held accountable for any and all of the computer’s mistakes or wrongdoing. Only in so doing Milei creates a real problem: how do you deal with companies designed to act and learn without a human manager and capable of doing harm without a human signature?
May God save us all.
My friend Nic Saldías actually made this point to me. Thiel has recently moved his family to Argentina and appears to be (unsurprisingly) a big fan of Milei.
Unless the Singularity arrives, needless to say. Well, that’s too strong. Old-fashioned limited liability might save our bacon even if it does.
The success of the Dutch East India Company (VOC) involved multiple institutional breakthroughs including legal personhood, permanent capital, transferable shares, separation of ownership from management, and only then limited liability. Oscar Gelderblom, Abe de Jong, and Joost Jonker tell the story in “The Formative Years of the Modern Corporation: The Dutch East India Company VOC, 1602-1623,” Journal of Economic History 73, no. 4 (2013). Most of the above discussion comes from their work.
That argument in turn built on earlier work by Steve Haber, which is why we wound up writing a bunch of books together.
Lamoreaux’s summary of the corporate-law sequence says special charters were patronage instruments rather than mainly public-good bargains. General incorporation stalled where legislatures could still grant special privileges. The later general incorporation statutes imposed regulatory provisions on capital, debt, duration, voting, reporting, dividends, and directors’ liability. Lamoreaux and Wallis’s brilliant 2021 article states that before the mid-nineteenth century most U.S. laws were special bills granting favors to specific individuals, groups, or localities.
I think worse, much worse, but hopefully I am wrong about that.
This time we are on the same page, strongly. I have been saying for a while, that the only thing I am afraid of about the AI is that humans would somehow make a mistake of granting it citizenship. Corporations are not quite citizens - but close enough. And, as you say, it has not been trivial to work out how to make their owners/managers responsible enough for the whole thing not to be a disaster. Argentina is playing with fire. Saner countries might be best advised from now on to disregard Argentinian jurisdiction as a matter of their own legislation.
The emergence of AI is a reason to strengthen liability laws. There should never be an action by AI for which a human being, capable of controling that AI, is not properly responsible. If that were the case, I would not lose much sleep over this particular round of technological change. Milei is the first one who made me really scared of it.