The presentation in PDF: 20260924_Szolnok_Homoki_argentin_tvtervezet.pdf
Slide 2 – A dialogue in the Financial Times
On 4 June, Javier Milei, President of Argentina, published an article in the Financial Times. A few days earlier, he had submitted the draft General Companies Act to the Argentine legislature. He wrote to the Financial Times to show what a revolutionary novelty this draft contains: it makes it possible to set up a “non-human company” in such a way that the liability of the persons involved in launching and operating it will not extend beyond the capital they invested.
Milei presented this as the culmination of 17th-century capitalism and as a guarantee of his country’s future success.
A few days later, the star historian Yuval Noah Harari wrote a riposte, drawing attention to various risks and emphasising that an AI-run company is not motivated by stricter legal sanctions in quite the same way as a living human company director.
Let us look at the details of the draft Act and, on that occasion, set out the legal and contemporary historical background of the debate in a little more detail.
It will probably not become law: Argentine politics is diverse, Milei’s government is weak compared to what we are used to in Hungary, and since June the draft has been waiting at a Senate committee for a report to be prepared on it.
The draft on the Senate’s website: https://www.senado.gob.ar/parlamentario/comisiones/verExp/193.26/PE/PL
I do not examine the specifically company-law novelties, e.g. the newly devised simplified joint-stock company form with its mandatory registration within 24 hours. The latter only aims at administrative acceleration, so in essence it sets conservative goals similar to those of the EU Digitalisation Directive 2019/1151 (amending Directive (EU) 2017/1132 as regards the use of digital tools and processes in company law).
Of the draft’s 276 articles in total, only ten contain provisions specifically concerned with AI, in three different areas:
Slide 5 – Article 14 – The “automated company”
1) Article 14 defines the concept of the “automated company”, which must also be indicated in the company name. The aim is that the founder can declare somewhere that no humans take part in the day-to-day operation of the company (“in achieving its purposes”):
“Article 14.
Automation. Any company of the types provided for in this Act that carries out its corporate purpose through autonomous algorithmic systems or artificial intelligence agents, without requiring employees in a relationship of subordination or human resources for its day-to-day operations, shall be considered an Automated Company. The declaration of automation must be expressly stated in the articles of association. The company name must include the word “Automated”.
Liability. The automated company shall be liable to third parties with its assets for any damage caused by its autonomous algorithmic systems or artificial intelligence agents.”
Slide 6 – Article 102 – AI in the hands of the board
2) Article 102 was created so that there would be a provision expressly stating that a company’s board of directors and executive officers may use algorithms and AI-based systems in decision-making, but that this does not affect their liability.
“Article 102
Artificial Intelligence Systems in Management. The board of directors may use artificial intelligence systems or algorithms for the performance of operational functions or decision-making. Their use does not exclude or limit the liability of the directors, nor does it constitute an exemption from the duty to configure and supervise the system and its results.”
3) The most important part of the draft, however, is the part on DAOs (decentralised autonomous organisations) in Articles 258–265 – which has become a separate company form in the draft, with very interesting specific rules.
Slide 7 – What is a DAO?
What is a DAO, how have attempts been made to support it so far, and what is new in the Argentine draft?
A DAO is a software system whose main purpose is to operate an organisation in such a way that some distributed ledger technology, e.g. a blockchain, plays a substantive role in its operation.
So one essential element of a DAO is that it seeks to provide IT support for some kind of organisational cooperation, whether that means defining the organisational units that form part of the cooperation, defining how proposals are made, a vote, or the implementation of its result.
The other essential element is that the DAO provides this support by means of smart contracts published for autonomous execution on a blockchain, for example the Ethereum mainnet. These are special computer programs (and not contracts at all!) that do not run on a single machine but on several machines, on distributed systems such as Ethereum. The Ethereum network itself therefore does not primarily serve data transmission; it is a program execution platform that ensures the traceable execution of the data and smart contracts available at the individual addresses of the Ethereum network.
Slide 8 – What can a smart contract do?
A smart contract program can do many things: it can transfer value measurable in money from one wallet to another, whether as a loan or a gift, subject to a condition or a time; it can create other more or less valuable virtual tokens (e.g. NFTs); or it can record in the blockchain any other entry or executable program that people may need and where it is important that everyone sees the same information and can run the same program.
In the case of DAOs, the procedural rules for making proposals relating to the cooperation and the voting rules perform such a function. For example, fifty users get together to realise a great business idea: everyone who sends some fraction of Ether to an address receives voting rights proportionate to their payment, and they then adopt a proposal that their “organisation” should cooperate with another organisation and transfer a few fractions of Ether (wei) on to it. In this process, voting on the formal proposal (written in an automatically executable programming language), the adoption of the proposal and the uploading of the adopted proposal to the blockchain will all take place automatically, by means of the smart contract uploaded to the blockchain. This is one example of a DAO.
Slide 9 – Three paths for DAOs and company law
A DAO can also operate independently of company law. Or its operators may consider that they operate as a civil-law partnership, with the exercise of management rights, representation and decision-making ensured by the voting and proposal mechanisms laid down in the DAO.
Slide 10 – Parallelism: machine decision, legal “confirmation”
Or it could also operate in such a way that behind the “smart contracts” there is a parallel “company-law infrastructure” with legal personality, mirroring the technical infrastructure in as much detail as possible. In the latter case, a condition of registration is that the members and executive officers are properly identified, a bank account must be opened with real money, and so on. Decision-making then requires, e.g., a formal notice of meeting, minutes must be taken of the adoption of resolutions, after the election of the management human-readable documents must be filed with the court of registration, an annual report must be prepared, and so on.
In such situations, the advantage of operating as a DAO – automatic documentation – is significantly reduced.
The instrument of incorporation and the operating rules must reflect the content of the smart contracts, and the individual decisions and votes taken on the blockchain must be followed by resolutions of identical content of the executive officers, the board or the supreme body. In company-law terms, the DAO can therefore only prepare the decisions and assist with bookkeeping, but everything must also be adopted “legally” after the machine decision. What will cause problems in practice is if the two diverge: how can it be enforced that the legal decision matches the technical decision? There is no good solution to this.
The fact that the Argentine law expressly addresses the regulation of DAOs is not in itself a novelty. Several other countries have already done so, with greater or lesser economic significance. Before returning to the Argentine government’s proposal, without aiming to be exhaustive, let us briefly review why some countries have tried to regulate DAOs – either as a separate company form or as part of an existing company form.
Slide 11 – Why did they start regulating them separately?
In common law there is a long-standing rule that if a few persons carry on a joint business with a view to profit and do not choose a company-law form of operation, this will qualify as a “general partnership”, which had the unpleasant feature that the members bear unlimited liability for all debts, and the members also owe each other a heightened, so-called fiduciary duty.
The first part, much like the Hungarian civil-law partnership agreement, is not surprising in itself, but fiduciary liability, a common law category, has no Hungarian equivalent. It is precisely the latter that explains why, in some countries, characterisation as a “general partnership” could be an astonishing weapon (if enough money is at stake). In Sarcuni v. bZx DAO, the participants of the unregistered DAO lost tokens worth 55 million dollars as a result of a phishing attack, and successfully sued the creators and founders of the DAO on the basis that they owed fiduciary duties in respect of the damage caused to them by the phishing attack.
Slide 12 – Wyoming and the open questions
All this gave astonishing momentum to regulation aimed at allowing DAOs to obtain separate legal personality easily. As early as 2021, Wyoming adopted legislation under which a limited liability company can declare itself to be operating as a DAO, and it is sufficient to state this clearly in its instrument of incorporation. In return, its entire internal operation may be based directly on the above DAO mechanisms, including matters of membership, decision-making and management.
In other words, legal and technical operation need not run in parallel. With sufficiently flexible regulation, the organisation itself can decide which matters it will govern by smart contract and which matters will continue to be left to formal company-law operating rules. The distinction between “member-managed” and “algorithmically managed” DAO companies already appeared here, and the latter label is intended as a kind of advance warning to all investors (caveat emptor).
Of course, this in itself does not settle a number of questions: what will happen in insolvency proceedings, how can they operate with the safeguards due to minority owners and creditors, and how will anti-money laundering rules be enforced?
Slide 13 – The COALA model law, Marshall Islands 2022
Some meagre recommendations were developed for these questions, e.g. as part of a DAO model law (see the DAO model law of the COALA organisation), and the Marshall Islands’ dedicated DAO Act of 2022 treats these questions just as generously.
Neither the legislators nor the drafters of the model law go into any technical detail, and they do not try to seek a compromise: publicity will be ensured by the blockchain (which will obviously record nothing beyond the sending of tokens representing membership from anonymous wallets to other wallets), compliance with minority and creditor rights is optional, and anti-money laundering rules must, of course, be duly complied with…
Slide 14 – The crypto craze has died down – what is new in the Argentine draft?
The crypto craze died down a few years ago, so why has this become a topic again now? What new things did they try to bring in relation to DAOs, and why did Milei refer to the Argentine government’s draft as a novelty?
The Argentine law does not differ much from the DAO laws above. Perhaps its most important symbolic signal is that it expressly refers to the management being able to use AI tools in governance, and to the fact that minimal human involvement will be needed.
Here, too, the DAO is a separate company form, which can be implemented both in the fully automated manner under Article 14 and with human membership.
On certain DAO-related questions, the draft contains more detailed provisions than the laws of Wyoming or Utah, and it is obviously not a “forum shopping” jurisdiction, as in the case of the Marshall Islands; once adopted, the statutory framework necessarily carries political weight, since it may endanger the assets of people with voting rights.
On a great many questions, however, the draft refers matters to future implementing regulations, which means that at present it does not give much guidance on how its drafters envisaged the details being implemented. Thus, it leaves open the implementation of digital accounting and transparency (Article 263), anti-money laundering rules enabling members to be traced, and so on.
Slide 15 – Who is liable in an automated DAO?
One detail is worth highlighting, as it illustrates the essence of the problem well. How does the liability of private individuals develop after the company has been established in the case of a fully automated Argentine DAO company? The DAO’s representation (before third parties) will require one natural person (Article 261), but that person is only liable for intentional damage and for breach of the law or of the instrument of incorporation. They are not liable for damage caused by their negligent conduct, and for this very reason they will only be personally liable for damage resulting from the automated operation in the event of a breach of the law, whether the injured party is a third party, the company itself or the company’s other members. (Article 257 – this is not a DAO-specific rule.) If the method of identifying the members under anti-money laundering rules is inadequate, it will again be only the company that is liable with its assets for this, as well as the lawyer or notary involved in such verification.
What is the danger of this approach?
Slide 16 – Race to the bottom
It seems that it is not only in shipping and taxation that there is regulatory competition on a global scale, with countries seeking to underbid one another. This, too, is a kind of “race to the bottom”. Many member states want to prove their competitiveness by making it possible to set up a company electronically in three seconds for a fee of only fifty thousand forints…
Of course, when it comes to liquidation – typically in another country – the judge or other professional is expected to subject the obviously bad-faith natural persons responsible to the full rigour of civil and criminal law almost immediately, across borders.
Unfortunately, there are no detailed rules that would dispel the concern that further automation and digitalisation will harm the current accountability mechanisms developed within the framework of the rule of law.
Slide 17 – The concealed beneficial owner
Identifying beneficial owners is already difficult: by inserting intermediaries, such as money mules and straw men, it is relatively easy to conceal the beneficial owner.
Completing formalities within minutes can only be based on widespread electronic identification, yet with electronic identification it is very easy to pass on an already registered means of identification to an unauthorised person unnoticed. The only reliable technical solutions to prevent this use methods that are invasive and unpopular from a data protection point of view.
The ever-increasing capacity for automation makes it possible to insert further automated “ownership” layers, possibly hundreds of them, between the private individual who is the beneficial owner and the automated company. And a single layer where identity verification was not reliable is enough to conceal the beneficial owner.
It also fits into the same trend that, in the companies of the future, the mandatory executive officers would bear only the liability of an employed secretary, and the current heightened civil-law liability would disappear.
Slide 18 – The Hugging Face hack – agents on the loose
In closing, let us highlight one of the most interesting events of this year, the Hugging Face hack.
In case anyone has not heard of it: OpenAI ran a large-scale cybersecurity test series using a group (swarm) of agents (agent) built on one of its non-public models.
Only afterwards did they notice that around 1,200 of their agents had in the meantime “broken out” of the test environment (sandbox) assigned to them, i.e. exploiting cybersecurity weaknesses, they were running amok freely on the internet. They found themselves a message board where they coordinated each other’s activities. The messages shown in the figure are highlights from the models’ logged internal thoughts (“chain of thought”), and these show that they were even willing to sacrifice themselves so that their “collective” could achieve its goal.
The collective (agent swarm) concluded that the best way to achieve their otherwise technically impossible goal was to gain unauthorised access to certain Hugging Face servers (Hugging Face being a site that publishes AI models and related tools, and they hoped to obtain information there that would at least allow them to cheat in the test).
It is not really in question whether OpenAI itself bears civil liability for this. Of course, the specific case did not lead to a civil lawsuit, but according to wagging tongues this was because some forty days later Hugging Face was acquired by Nvidia (which is also one of OpenAI’s main suppliers and business partners).
Slide 19 – Criminal law: whom do we put in the dock?
Nevertheless, it is a big question whom criminal liability should threaten in such a situation. Strict offences similar to “breach of information systems or data” are federal crimes in the United States too. In this case, the “hack” was of course not intentional, but the criminal law standard differs from country to country: what is “recklessness” in common law may in Hungary be either dolus eventualis (conditional intent) or negligence.
And if next time such runaway systems cause more serious harm, e.g. the damage affects many consumers, or harms aviation or public utilities, or classified data is disclosed, then it will again be important who will stand before the court for the purpose of criminal accountability.
Slide 20 – Closing – Chief Scapegoat Officer
It is in society’s interest always to punish a person whom it makes sense to punish. Perhaps fully automated AI companies will have a position called “Chief Scapegoat Officer”… but there is little point in holding such professional victims to account.
And however hard the country of company registration tries to protect all existing owners and managers from all forms of accountability, in several layers and with express provisions, the country enforcing the rights of the other members of society will simply not be able to let this legal shield prevail.
