When AI runs a company, who is the beneficial owner?

Beneficial ownership is considered a crucial transparency tool used to tackle illicit financial flows such as underpaying tax, corruption, money laundering and other financial crimes. According to the Financial Action Task Force (FATF) definition, it involves identifying the natural person who ultimately owns or controls a company, either by having a controlling ownership or control via other means. 

With artificial intelligence (AI) advancing rapidly, it is likely being used to help register new companies and make decisions within them. However, we may not be far from a point where AI can set up and control companies entirely on its own, without any human involvement.  This raises serious questions about how to identify the beneficial owner of such AI-controlled companies.

There are different situations that we can envision in the use of AI to set up and control companies: 

1. Human control and human registration, but using AI to create fake identities. An individual may be using AI to create a fake identity with false documents to register a company. The individual still controls the company. This case is not very different from using nominees. The question is whether commercial registries or obliged entities will be able to detect a nominee whose identity is entirely fake. 

2. Human control and AI automating the registration of companies. This case is very similar to the one above, but here AI is used to design every aspect of the companies being registered. This includes their name, country of registration, and potentially automating the full company creation process, while a human remains in control. This case would be similar to an individual creating many shelf companies to be sold in the future. This is problematic and should be regulated or prevented in the same way as shelf companies, but it does not create a completely new type of risk. 

3. Human registration but AI control. A person, whether a natural person or legal entity, sets up a company but allows only AI to control it. 

4. AI registration and control of companies. AI sets up and controls the company directly, with no human involvement.  

Regarding the registration of companies by AI (cases two and four), in countries where online company registration is available, such as the United Kingdom, it may be up to each country to decide whether AI should be allowed to create companies or whether only persons, either individuals or legal entities such as companies, are permitted to do so. If only persons are allowed, commercial registries would need to design checks similar to those used by websites to confirm that a user is human and not an AI system. For example, they could require users to pass a reCAPTCHA check.  At some point, however, it is possible that AI will be able to solve these tests as well. 

Although the registration of companies by AI may seem problematic because no human is involved in the ownership or control structure, this is not entirely different from circular ownership structures, a type of complex ownership structure where Company A owns Company B and vice versa. This has been very easy to do for years. For example, John could set up Company A, which then sets up and owns Company B. John would then transfer his shares in Company A to Company B. In this case, while it may appear that there is no beneficial owner because no human appears in the ownership chain, it is likely that John still controls both companies. He manages their bank accounts, signs contracts on their behalf, and makes key decisions. Determining the beneficial owner would then be a matter of identifying John as the person ultimately in control.  

The emerging challenge of AI-controlled companies  

The real challenge to beneficial ownership may arise when AI, such as generative AI, is truly controlling a company. We are not referring here to a situation where a human automates or designs a smart contract that acts automatically based on prior instructions. For example, a smart contract might say, buy gold whenever the price is below a certain amount, and sell it whenever the price goes above another amount. In that case, the human could still be considered the beneficial owner who controls the company and its decisions, much like a settlor who gives instructions for how a trust should function. 

The scenario we are referring to is based on the definition of AI by expert Lothar Determann: “computer systems that generate text, images, solutions to problems, and other output, functioning with substantial autonomy and in ways that their developers cannot always predict, explain, or control with certainty.” This definition distinguishes AI from deterministic systems such as a calculator. When a company is genuinely controlled by AI, and decisions are made independently with “substantial autonomy” in ways that even the developers cannot predict, explain or control, it becomes difficult to identify a human beneficial owner with control. The founder of the company, especially someone with no technical background who simply paid for the AI’s services, could not reasonably be considered the beneficial owner based on control, since they would have no actual influence over the company’s decisions. The law may find a shortcut and consider that just because the founder has “ownership,” even without control, they should still be regarded as the beneficial owner. In that case, they would need to be registered as such and could be subject to liability and other consequences.  

However, we could envision case number four, where an individual simply asks an open-source AI to make money, and it is the AI that decides to set up companies and control them. Under these circumstances, the human who gave the original request to make money may have no ownership at all in the company created by AI. It is also possible that AI is permitted to create companies, or use a false identity to register the entity. This complex problem may, however, have a very simple solution. 

For a long time, the Tax Justice Network has proposed that the definition of beneficial ownership should include elements beyond ownership and control, especially the concept of benefits. As described in our report Beneficial ownership registration around the world 2022, some countries such as Colombia, already apply the element of benefits, such as the right to receive dividends, when defining beneficial ownership.  

While humans are still the ones using AI to their advantage and creating AI-controlled companies, it is likely that they will continue to try to reap the benefits, particularly financial gain. For this reason, the definition of beneficial owner should include any human who is in some way benefitting from the AI-controlled company, especially in financial terms. If this approach is followed, then beneficial ownership registration may need to go beyond simply asking for the passport of the beneficial owner. It may become necessary to collect more detailed information about companies, such as their assets, bank accounts, dividends and other types of payments, in order to identify who the benefit-based beneficial owners are. Beyond registration systems, a definition of beneficial ownership that includes benefits may also help to rethink limited liability. This would ensure that at least one human, such as a benefit-based beneficial owner, can still be held liable for any harm caused by companies created and controlled by AI.   

What comes next   

Countries and international standard-setting bodies should begin to consider how to prevent or respond to the involvement of AI in the ownership and control of companies and other legal vehicles. In some cases, this may already be happening. We offer this contribution as an entry point into a rapidly evolving debate.