Abstract

This paper proposes a hybrid model of liability for regulating AI systems. The model divides AI systems into two distinct groups. The first group comprises fully autonomous AI systems that are sufficiently intelligent and capable of performing tasks without human intervention. These systems would be granted a legal personality, inspired by the separate legal personality of corporations, and liability would be imposed on them for their mistakes. This approach also requires governments to define a set of criteria that an AI system must satisfy before it is granted such personality. The second group consists of all other AI systems, which are ineligible for personhood. These systems would be regulated under a strict liability regime, under which liability for their actions lies with the creator, the owner, the user, or a combination of them, depending on the circumstances. The benefit of this model is that countries can begin adopting it now, without any obligation to assign legal personhood to any of the AI models currently available. They can set out the specific criteria they consider essential for an AI system to satisfy before being granted personhood, and grant such personhood only once an AI system satisfies those criteria. Until then, all other AI systems can be treated as belonging to the second group and regulated under strict liability. The “wait and see” approach must be discouraged in the regulation of AI systems, as attempting to regulate them too late may prove dangerous.

1.Introduction

In response to the question “Will artificial intelligence outsmart us?”, Stephen Hawking stated that “[w]hen an AI becomes better than humans at AI design, … we may face an intelligence explosion that ultimately results in machines whose intelligence exceeds ours by more than ours exceeds that of snails. When that happens, we will need to ensure that the computers have goals aligned with ours.”1 The presence of AI systems in our lives is inevitable, and the rate at which these systems are advancing in intelligence is astonishing. Large language models such as ChatGPT can now take standardized tests designed for humans and outperform the average test taker.2 AI is also expected to become a two-trillion-dollar market by 2030, a twenty-fold increase from its current size.3 Given the emergence of AI models such as ChatGPT, the ever-increasing capabilities of AI systems, and their enormous economic value, Hawking’s vision of AI’s intellectual dominance over humans may not be so far-fetched, as AI’s progress shows no sign of coming to a halt any time soon.

As AI systems become intelligent enough to perform tasks entirely autonomously, assuming roles previously held exclusively by humans, such as those of doctors, lawyers, and accountants, an important question arises for governments around the world: who will be responsible for mistakes committed by fully autonomous AI systems? Any answer to this question must account for the plausible scenarios that will unfold in the future of human–AI interaction. These include, for example, circumstances in which AI surpasses human intelligence in its ability to perform particular tasks and human intelligence becomes inferior; circumstances in which AI goes “rogue” and begins disobeying humans; and circumstances in which an AI’s decision is at odds with human expectations, but humans are unable to determine whether the human decision or the AI’s decision is the better solution. This paper refers to the last of these scenarios as the “AI decision dilemma.”

Under the liability schemes currently in place in various jurisdictions, the only viable answers to this question would be either to hold AI creators liable for civil or criminal negligence through product liability, or to assign strict liability to a legal entity such as the user, owner, or creator. However, both options will fail to properly regulate the highly capable and intelligent AI systems of the future, mainly because of the unforeseeable nature of AI actions stemming from their “black box” algorithms. As AI systems become more intelligent, it becomes significantly more difficult for their creators to predict their actions in the long run, especially where the AI is designed to learn from its users’ behaviour and adjust its actions accordingly. Current liability schemes could therefore leave without compensation those who have been harmed by an AI but who cannot hold its creator liable because of foreseeability issues.

Furthermore, transferring liability for an autonomous AI’s actions to another entity could hinder innovation in AI-related fields and reduce the benefits of the new generation of AI systems. Once AI systems operate at a level superior to that of humans, they could make decisions that may not be explainable to people. Faced with the AI decision dilemma and fearing the consequences of a wrong decision, people would likely disregard the AI’s decision, even where it may have been the more beneficial one. This disregards one of the main goals of AI development: to allow AI systems to find patterns and solutions that humans cannot. Such an outcome would certainly discourage the development of AI systems that have the potential to be immensely helpful to society by finding solutions to its serious problems in ways that humans could not have achieved for generations.

There is, however, a solution that balances the government’s interest in protecting the public, on the one hand, with maximizing and embracing AI’s potential, on the other: assigning liability to the AI itself by granting AI systems a legal personality. Under this approach, governments would define the criteria an AI must satisfy to be eligible for legal personhood. Once an AI system satisfies those criteria, it would be registered as a legal entity and could assume liability for its actions. This paper refers to these specific types of AI systems as “AI persons.” For an AI’s legal personality to be enforceable, governments must also design a financial scheme that allows AI persons to compensate others if they are found liable. This would be made possible by granting AI persons the right to hold property and other financial assets.

This is not to say, however, that every AI would be eligible for legal personhood, as some AI systems cannot satisfy the eligibility requirements. This paper therefore proposes a hybrid model of liability for regulating AI systems that divides them into two distinct groups. The first group will include those AI systems that have been deemed eligible for personhood; these systems will be subject to the same liability framework as any other legal entity. The second group will include those AI systems that are ineligible for legal personhood; these systems will be regulated under a strict liability model, under which governments can choose to assign strict liability to the creator, the owner, the user, or a combination of them, depending on the circumstances.

This hybrid model is the most sustainable framework for the long-term regulation of AI systems. It removes the confusion surrounding the allocation of liability for actions executed by fully autonomous AI systems, eliminates the possibility of non-compensation due to foreseeability issues, and gives courts confidence in assigning liability to the appropriate entities. The model also allows the potential of future AI systems whose intelligence surpasses that of humans to be realized, by assuring the creators and users of these systems that they will not be held unfairly liable for the AI’s actions, especially in circumstances where their view of the correct course of action differs from that of the AI. A further advantage of this model is that governments can begin adopting it today without having to assign legal personality to any of the AI systems currently on the market. Governments can begin evaluating the criteria they believe must be satisfied for an AI system to be recognized as a legal person and, until an AI satisfies those criteria, continue to govern it under a strict liability model. This approach is preferable to the “wait and see” approach that governments usually adopt in dealing with new technologies, as it may be too late, and perhaps dangerous, to regulate such intelligent systems once they have already arrived.

Section 2 of this paper reviews the current laws and regulations governing AI systems, and the liabilities arising from them, in various jurisdictions. Section 3 addresses how AI-made mistakes would be treated under current liability models and explains why these models are insufficient for the proper regulation of sophisticated AI systems. Section 4 introduces the corporate model of legal personality as a source of inspiration for an AI legal personality model, as corporations are currently the most widely recognized artificial legal entities in the world. Section 5 identifies some of the challenges involved in creating a regulatory framework for AI persons. Finally, Section 6 proposes a regulatory scheme that allows AI systems to be recognized as separate legal entities.

Throughout this paper, the term “fully autonomous AI” describes an AI system that can, but need not, operate entirely independently and without any human intervention. The term “AI person” describes a currently hypothetical AI system that has been recognized as a legal entity.

2.Current Status of AI Regulations in Different Jurisdictions

Jurisdictions around the world have begun regulating the creation and use of AI systems, and some have even begun contemplating rules for addressing liability issues arising from AI. This shows that lawmakers are becoming aware of the potential issues stemming from AI’s autonomy. This section briefly describes the current laws and regulations governing AI systems in different jurisdictions and assesses their effectiveness in providing a definitive solution to AI-related problems.

2.1General Regulations

2.1.1European Union

The European Union is by far the most active jurisdiction in its attempts to regulate the creation and use of AI systems. Once passed, the EU AI Act will be the most comprehensive set of laws ever enacted to regulate AI. However, most of the Act is dedicated to regulating a specific category of AI systems known as “high-risk” systems, which include any AI system used as a safety component of a product, as well as the other AI systems listed in Annex III of the Act, such as those used in health, employment, education, biometric identification, social services, and law enforcement.4 While the list of high-risk AI systems is very comprehensive, many, if not most, of the Act’s provisions would be inapplicable to an AI system that falls outside the high-risk categories.

While the Act does not explicitly assign liability for AI-related mistakes, it does impose certain duties on AI creators, owners, and users. Providers of high-risk AI systems, for example, are obliged to ensure the AI’s compliance with all applicable European regulations5 and to take corrective action once non-conformity is detected after the AI has been placed on the market.6 Users, on the other hand, are tasked with monitoring the activities of the AI system in accordance with its instructions for use and with keeping a log of all AI activities in case they need to be reported to the authorities.7 As for fully autonomous AI systems, the EU has decided that such systems will not be permitted to operate within the EU and that human oversight is a necessary condition of an AI’s existence.8 However, the Act does not specify the extent of oversight required for fully autonomous AI systems. Since constant oversight of AI systems is impossible, the Act leaves unresolved what would happen if an AI were to make a mistake while unsupervised.

The EU’s General Data Protection Regulation (“GDPR”) is another EU instrument that affects the creation and use of AI systems in Europe. One of the main challenges for AI creators conducting business in the EU is compliance with the GDPR’s requirements concerning the withdrawal of consent and the right to erasure.9 Compliance is particularly challenging because it is unclear how the creators of AI systems would erase data that has been fed into the system to train its algorithm, especially where a significant amount of data must be deleted. This requirement is so strict that, if properly enforced, it could oblige AI creators to remove all of their data from the algorithm and start over. The GDPR also grants users “the right not to be subject to a decision based solely on automated processing,”10 which, once again, reflects the EU’s strict stance against fully autonomous AI systems.

2.1.2Canada

Inspired by the work of the EU, Canada has also taken steps toward regulating AI systems. In 2022, Bill C-27, also known as the “Digital Charter Implementation Act,” was introduced in Canada’s Parliament.11 Among other things, the Bill introduces the “Artificial Intelligence and Data Act” to “regulate international and interprovincial trade and commerce in artificial intelligence systems by establishing common requirements, applicable across Canada, for the design, development and use of those systems” and “to prohibit certain conduct in relation to artificial intelligence systems that may result in serious harm to individuals or harm to their interests.”12 Bill C-27 introduces the concept of a “person responsible” for an AI system, defined as any person who designs, develops, or makes available for use the artificial intelligence system or manages its operation. It remains unclear whether “managing the operation” of an AI system falls within the duties of the creator or of the user once the AI is available on the market.13 Bill C-27 also makes it a criminal offence to knowingly or recklessly release AI systems capable of causing severe harm or damage to the property of others, or AI systems intended to defraud the public and cause economic loss.14

2.2Regulations Related to AI’s Liability

Some countries have already begun assigning liability for the actions of autonomous AI systems, although none of them appears to anticipate the arrival of highly sophisticated, fully autonomous AI systems. Most of these regulations relate to the automotive industry, as the boom in autonomous vehicles has become a primary concern for many jurisdictions worldwide. Even though none of these regulations considers scenarios in which an AI system operates completely autonomously, they reveal the tendency of different jurisdictions to allocate liability for actions taken by AI systems to either the user or the manufacturer of the AI.

2.2.1France

France, for example, has adopted “Ordinance No. 2021-443 of April 14, 2021, relating to the criminal liability regime applicable in the event of the circulation of a vehicle with driving delegation and its conditions of use,” which allocates liability in the case of vehicles whose control has been delegated to an AI system but can be taken back by the driver at any point when desired or required. The ordinance releases the driver from criminal liability in situations where the driver does not have control of the vehicle and imposes criminal liability on the manufacturer instead.15

2.2.2The UK

The United Kingdom, on the other hand, has chosen to exclude manufacturers from liability in such circumstances. Under the Automated and Electric Vehicles Act 2018, where damage arises from an accident that occurs while the vehicle is “driving itself,”16 liability rests with the insurer of the vehicle or, where there is no insurance policy, with the owner of the vehicle.17 The Act does, however, allow the insurer to claim damages from “any other person liable to the injured party,”18 which means that insurance companies may still have an avenue to assign liability to the manufacturer in a civil court.

2.2.3Germany

Germany was the first country to adopt regulations governing the operation of SAE Level 4 autonomous vehicles, allowing the public to use vehicles with that level of automation under the supervision of a “technical person” tasked with ensuring compliance with traffic laws. The role of the technical person is expected to be assumed by the driver unless the driver delegates it to another person. The regulations also require these vehicles to collect driving data in compliance with the GDPR and to transmit that data to the appropriate authorities in the event of an accident. Germany’s regulations do not change the owner’s liability in the event of an accident,19 even where the role of the technical supervisor has been delegated to another party.20

2.2.4The EU: The Missed Opportunity of Discussing AI Legal Personality

Perhaps the closest any regulatory body has come to addressing the potential liability issues arising from highly capable and intelligent AI systems was the draft report prepared by the EU’s Committee on Legal Affairs, which set out recommendations to the Commission on Civil Law Rules on Robotics.21 The report contained 38 recommendations, among which was a call for the Commission to:

“creat[e] a specific legal status for robots, so that at least the most sophisticated autonomous robots could be established as having the status of electronic persons with specific rights and obligations, including that of making good any damage they may cause, and applying electronic personality to cases where robots make smart autonomous decisions or otherwise interact with third parties independently.”22

The Committee also recommended that the Commission consider ways to ensure that those who suffer damage are fairly compensated, even where the damage was caused solely by a robot.23 In response, the Policy Department for Citizens’ Rights and Constitutional Affairs published a study rejecting the idea of granting autonomous robots legal personality, describing it as unhelpful and inappropriate.24 Instead, the study proposed that liability arising from an AI’s actions be allocated as follows:

  • If the AI is sold with open-source software, liability rests with the programmer.
  • If the damage can be traced back to the AI’s design, the designer or producer of the AI must be held responsible.
  • If the damage occurs while the AI is in use or while it is still learning, liability can be allocated to the user or the owner, depending on whether the user is a professional and whether or not they are a victim.25

While the study makes some valid points, it rests on several erroneous presumptions and misconceptions that, to an extent, undermine its argument. The first issue is the study’s tendency to use the terms “robot” and “AI” interchangeably, as if they had the same meaning. Whereas AI is a computer-generated algorithm that carries out certain actions, a robot is the physical embodiment that enables the AI to interact with the physical world. A self-driving car offers a useful analogy: the vehicle is the “robot,” while the software that allows the car to drive is the “AI.” The study also adopts a very narrow definition of a robot, imagining it to be an object that behaves exactly like a human. This is evident from the study’s characterization of the advocacy of AI legal personality as “inspired by science fiction.”26

The second issue with the study is its presumption that legal entities require consciousness. On this view, other legal entities, such as corporations, must have received their legal status simply because a conscious human being stood behind them. This presumption is even more evident in the reasoning underlying the study’s rejection of a separate legal personality for AI systems:

“Yet how can a mere machine, a carcass devoid of consciousness, feelings, thoughts or its own will, become an autonomous legal actor? How can we even conceive this reality as foreseeable within 10 to 15 years … From a scientific, legal and even ethical perspective, it is impossible today—and probably will remain so for a long time to come—for a robot to take part in legal life without a human being pulling its strings.”27

By focusing on the consciousness of the robot, the study entirely misses the point of how an AI legal personality might be envisioned and completely ignores the issue of the autonomy of future AI systems. As this paper will explain, autonomy should lie at the core of any discussion of AI legal personality, because consciousness has a highly subjective definition and is neither sufficient nor necessary for recognizing the legal personality of any entity. In general, the study fails to respond to the challenge of regulating future AI systems and their growing capabilities once they outperform humans in intelligence.

3.AI Legal Personality vs. Current Liability Models

It is undeniable that AI is becoming more intelligent by the day. AI systems constantly improve as larger collections of data are added to their libraries, making them capable of performing tasks with ever greater accuracy and consistency. AI systems based on large language models, such as IBM’s Watson, OpenAI’s GPT-4, and, more recently, Google Bard, are constantly being updated to better understand human language. A comparison between OpenAI’s GPT-3 and GPT-4 shows how quickly AI systems are improving. As AI systems become more capable and intelligent, it is only a matter of time before they are deployed in every industry and begin to acquire the ability to perform tasks completely independently, without any human intervention. It is equally inevitable, however, that these AI systems will eventually make mistakes and cause damage, even if they operate at a level of intelligence higher than that of humans. When this happens, a central question will be who is liable for the damage caused by these AI systems. While determining liability is much simpler for current AI systems, which are not yet fully independent, the question becomes extremely challenging as these systems become fully autonomous and human supervision is minimized or removed. Such systems are not products of fiction but realities waiting to happen. Consider the following scenarios:

  1. A fully autonomous AI dietitian is tasked with controlling its user’s eating habits to ensure that the user lives a healthier life. The AI has access to the user’s shopping list and food-related home appliances, such as the refrigerator and the coffee machine, as well as to the user’s bank account so that it can order groceries automatically. The user is confident in the AI’s ability to make them healthier. Suppose that, in order to help the user live a healthier life, the AI dietitian forces the user to stop eating processed food and junk food altogether and prevents the user from having more than one cup of coffee per day. Suppose further that, if the user manually places junk food in the refrigerator, the AI recognizes it and disables the refrigerator until the junk food is removed. As a result of this interference, the user develops depression and other mental health issues.
  2. A fully autonomous AI accountant is tasked with handling the user’s financial matters and investment accounts, including making sound investments and filing the user’s taxes. Suppose the AI gives the user very poor advice on a financial matter, causing the user to lose a significant amount of money. Suppose also that, in an effort to save the user money, the AI identifies what it mistakenly believes to be a way of reducing the user’s taxes, but the action turns out to be illegal, and the user is charged with tax fraud.
  3. A fully autonomous AI surgeon is tasked with operating on a patient. During the operation, the AI accidentally cuts a tissue, causing heavy bleeding, and the patient dies.
  4. A fully autonomous AI has joined a city’s police force. The AI is tasked with recognizing dangerous situations that could lead to human deaths, such as shootings, and with intervening to ensure that human deaths are minimized. One day, while on duty, the AI mistakenly perceives a minor conflict as dangerous and shoots the person involved in the conflict. The person dies as a result of the shooting.
  5. An SAE Level 5 fully autonomous vehicle is driving its user to the countryside. The AI determines the most appropriate route to the destination and begins driving along it, even though the user suggests that the route is somewhat dangerous because of animals. The AI, having the ability to determine its actions autonomously, decides to override the user’s input and drives along the dangerous route anyway. An animal jumps in front of the vehicle, and the AI, in trying to avoid the animal, strikes a pedestrian, injuring both the passenger and the pedestrian.

To properly assess the liability issues arising from the events above, in which AI systems make mistakes, governments must consider the full potential of these intelligent systems, including (a) the possibility that an AI system might go rogue and begin maliciously disregarding human orders, and (b) the fact that some of these systems will eventually surpass human intelligence, becoming what many scholars call a “superintelligence.”28 It should be noted that this paper does not suggest that only superintelligent AI systems should qualify for a separate legal personality. Rather, it suggests that, when governments consider how to allocate liability for AI-generated damage, they must account for the possible arrival of superintelligent AI systems.

As AI systems become more capable, the prospect of AI systems disobeying human orders, even when heavily supervised, may become a matter that governments will want to consider seriously. This is especially important because some research now suggests that there is a threshold of intelligence beyond which controlling an AI becomes impossible due to the fundamental limits of computing.29 The arrival of highly sophisticated AI systems also raises the possibility of the AI decision dilemma. If an AI with such capabilities decides to take a specific action (for example, an AI doctor deciding to prescribe a particular medication) and humans believe that the AI’s decision is wrong, what should the humans do? Subject to certain ethical limitations, it would be unwise to disregard the AI’s decision immediately. After all, owing to the large volumes of data at their disposal, AI systems are capable of finding solutions that humans could not have conceived. In March 2016, DeepMind’s AlphaGo defeated the world Go champion by four games to one, in the course of which it made a move that no human would ever have made and that humans would probably have found absurd. “The killer move by AlphaGo was successful precisely because it used tactics which went against all traditional human schools of thought.”30

All things considered, governments will be limited to the following options when dealing with AI systems:

  1. Option 1: Completely ban the use of fully autonomous AI systems and require constant human oversight of the actions taken by AI. This would be similar to the European Union’s human oversight requirement in the EU AI Act.31
  2. Option 2: Transfer liability to a legal entity, such as a human (the user, creator, or owner) or a corporation (where a corporation is the owner). Liability could be imposed in one of the following ways:
    1. A strict liability regime, under which liability is imposed on the creator, the owner, the user, or a combination of them, depending on the approach that governments consider appropriate.
    2. A non-strict product liability regime, currently practised in some countries such as Canada, which allows those who suffer damage either to claim damages for breach of contract or to sue in negligence.
  3. Option 3: Assign liability to the AI itself, which requires creating a regulatory scheme that adopts an AI legal personality.

This section evaluates the three options and explains why the first two are not viable for regulating fully autonomous AI systems in the long run and will eventually become inoperable. The third option, granting AI a legal personality, best safeguards the public interest by offering a solid framework for claiming damages, while preserving the proper administration of justice by ensuring that those who suffer damage are compensated and that innocent entities are not unfairly held liable.

3.1Banning Fully Autonomous AI Systems

A ban on the use of fully autonomous AI systems can take two forms. The first is to prohibit the operation of such systems entirely, meaning that no AI creator would be permitted to place on the market any AI system capable of operating completely autonomously. The second is to permit the operation of such systems but to require human oversight, even where the systems would not need such supervision. The latter resembles the “human oversight” requirement adopted by the European Union in the EU AI Act,32 and it is the focus of this section. Human oversight means that manufacturers are obliged to design and develop AI systems in such a way that “they can be effectively overseen by natural persons during the period in which the AI system is in use.”33 This would require the human supervisor to be able to intervene at any time during the operation of the AI system, either to override the AI’s output or to stop its operation completely whenever the supervisor believes the AI may exhibit risky behaviour.34

The human oversight requirement in the EU AI Act, and any similar provision or regulation, presents significant implementation challenges. The first challenge concerns the extent of the oversight, which requires determining how far the AI’s operation and outputs must be supervised. While supervising the AI system during the design process may not pose legal challenges, supervising it while it is being operated by the user will raise significant privacy issues if the user does not consent to data collection by the AI’s creator. This is especially true under the EU’s own GDPR, one of the strictest data collection regimes in the world.35 Since users themselves are often neither required nor qualified to interpret the AI’s actions correctly, it is evident that, without information about the user’s interaction with the AI system, constant supervision of, and intervention in, the actions taken by that system would be impossible.

The second issue with the human oversight requirement concerns the nature of the oversight, that is, what the supervisor must do to ensure the AI’s safety. The EU AI Act provides that the supervisor must “fully understand[s] the capacities and limitations of the high-risk AI system and be able to duly monitor its operation, so that signs of anomalies, dysfunctions and unexpected performance can be detected and addressed as soon as possible.”36 This, however, will become extremely difficult as AI systems become more intelligent. The black-box nature of AI algorithms means that no one can fully predict an AI’s actions, and fully understanding an AI’s output is neither possible nor desirable. Indeed, as noted earlier in this section, one of the great promises of AI systems is their ability to “think outside of the box” and find patterns that humans are incapable of finding.37 The oversight requirement could therefore lead human supervisors to override an AI’s decisions in circumstances where the AI is making a novel discovery with the potential to benefit humans significantly.

Giving humans the power to override an AI’s decisions without a full understanding of how the AI reached them, while partly beneficial, would certainly eliminate the AI’s ability to innovate and find solutions that require a higher level of intelligence than that of humans.

3.2Establishing a Strict Liability Regime

Another option for resolving liability for AI actions is the creation of a strict liability framework. Under such a framework, if an AI were to cause damage, civil or criminal liability for the AI’s action would automatically transfer to a person who can be held accountable at law. As explained in Section 2, some jurisdictions have already begun regulating semi-autonomous AI systems through such rules.

The first problem in establishing a strict liability framework for fully autonomous AI systems lies in choosing the entity on which governments would impose strict liability. This could prove highly challenging for AI systems capable of full autonomy. Once autonomous AI systems begin interacting with users other than their creators, and as an AI adjusts its behaviour to a user’s preferences over a long period of interaction, it becomes extremely difficult to determine the root cause of the AI’s actions. In other words, it will be challenging to determine whether an AI’s improper action resulted from a fault in its programming or from the AI learning the user’s behaviour.

Strict liability could also severely impede AI’s growth and hinder innovation by placing all the risk on humans in an “AI decision dilemma.” Under a future strict liability regime, if an AI proposes a course of action that is at odds with human judgment, humans will tend to ignore and disregard the AI’s decision because of the risk of liability they would bear for the AI’s actions, even though the AI’s decision might be the correct approach and humans are simply unable to recognize it. As a result, AI systems will not be permitted to explore intelligence beyond that of humans, as their decisions will most often be disregarded unless humans are willing to accept the risk of following them. Holding humans strictly liable for the actions of highly intelligent AI systems would be analogous to holding the parents of an adult liable for all of that adult’s actions; it is neither fair nor practical.

3.3Regulation Through Non-Strict Product Liability Rules

Another avenue for recovering damage caused by AI is product liability. Under this approach, AI would be recognized as a product, and AI creators would be liable for damage arising from the everyday use of their AI systems. Governments can enforce product liability rules within either a strict or a non-strict framework. For example, while product liability in the United States takes mainly the form of strict liability,38 Canadian courts and legislatures have largely rejected the concept of strict manufacturer liability for products,39 even though Canada’s provincial product liability laws contain certain elements of strict liability.40 Since Section 3.2 has already shown strict liability to be ineffective in addressing AI-created mistakes, this section focuses on the non-strict form of product liability, taking Canada as an example of a jurisdiction that applies it.

In Canadian jurisdictions, there are generally two main avenues of recovery in product liability: first, claims for breach of warranty, which arise where there is a contract and a seller–buyer relationship, as well as under the legislatures’ consumer protection and sale of goods statutes; and second, claims in negligence.41 This section explains why both of these avenues of recovery (contract and statute-based claims, and negligence claims) are ill-suited to fully autonomous AI systems that can operate without human supervision.

3.3.1Contract and Statute-Based Claims

Manufacturers have a duty to represent their products accurately and to assure consumers that their products will perform as described. If a product fails to meet those expectations, consumers naturally have a right to claim damages from the manufacturer for breach of warranty.42 The source of this warranty can be traced to provincial statutes such as consumer protection legislation and the Sale of Goods Acts.43 The law also requires manufacturers to warn users of any dangers that may arise from use of the product.44

This framework, however, works only for products whose operation is highly predictable and that do not evolve over time. Computer batteries provide an example. Manufacturers of computer batteries can assure consumers of their batteries’ performance with a high degree of certainty because batteries have a predictable function, and if a fault is detected in a battery’s performance, it can easily be determined whether the fault resulted from a manufacturing defect.

AI systems, however, are not predictable in the same way, because they continue to evolve after they leave their creators’ hands. Some AI systems learn from their users and adjust their performance accordingly. In fact, no two AI systems trained by their creator on the same data will behave in precisely the same way once they have been tailored to their users’ preferences through different sets of input data. This makes it extremely difficult for AI creators to predict an AI’s performance. AI creators can train an AI algorithm as diligently as possible, with extreme levels of scrutiny and inspection, and AI systems may nonetheless perform tasks in ways that no one could predict. Unpredictability is inherent in AI: AI systems can take surprising actions that no one expects because they process data in ways that humans cannot understand.

To hold AI creators liable for damage caused by AI systems, users must prove that the damage resulted from a fault on the part of the AI’s creator and that the creator must have known of this defect, which, as noted, is very difficult given the nature of AI systems. Under the “state of the art” or “development risk” defence, Canadian law allows manufacturers to escape liability where the state of knowledge at the time of manufacture did not enable the defect to be identified.45 Therefore, even if contract and statute-based product liability claims were applied to autonomous AI systems, AI creators would most likely be able to defend themselves by arguing that it would have been impossible to predict the specific action taken by the AI system once it left the creator’s hands. AI systems do not belong in the same category as other products and should not be treated as such.

3.3.2Negligence

Negligence claims against the creators of AI systems will likely fail for reasons very similar to those that defeat contract and statute-based claims. In Donoghue v Stevenson, the leading common law authority establishing the principle of negligence, Lord Atkin stated that a person can be held liable only for reasonably foreseeable harm.46 Furthermore, the Supreme Court of Canada in Cooper,47 following the House of Lords’ decision in Anns,48 held that establishing the existence of a duty of care requires a foreseeable harm to a foreseeable plaintiff.49

Foreseeability of harm will be the main obstacle for plaintiffs suing AI creators in negligence. Uncertainty as to the type of data an AI receives from its user, and as to how the AI processes that data, could render any surprising move by such a system unforeseeable to its creator. In other words, the AI’s black-box algorithm, coupled with its creator’s inability to control every input the AI receives once users begin interacting with it, could make it nearly impossible to predict every action the AI takes.

Proving negligence also requires establishing causation, meaning that the actions of the AI’s creator must be the direct cause of the injury. In Clements v Clements, the Supreme Court of Canada explained that the “but for” test is the proper test for determining causation.50 Under the “but for” test, the plaintiff must show that the damage would not have occurred but for the defendant’s actions. As explained above, once AI systems are intelligent enough to perform tasks on their own, and as they spend more time with their users, it becomes difficult to prove that the AI’s creators were responsible for a specific action produced by the AI system. Once AI systems are intelligent enough to operate completely autonomously, the damage may become too remote from the AI’s creators for them to be found liable for causing it.

With users unable to prove either the foreseeability of harm or the direct role of the AI’s creator in causing the injury, it is unclear how courts will allocate liability to these creators. Users may consequently be unable to recover anything at all, because they cannot establish the fault of any legal entity for the accident.

3.4Establishing an AI Legal Personality and Assigning Liability to the AI Itself

Although some scholars disagree,51 assigning liability to the AI itself appears more promising than any of the other options. Recognizing an AI’s legal personality would ensure that AI creators and users are not punished or held liable for the AI’s unforeseeable actions, while at the same time allowing injured parties to recover damages for the injuries they suffer.

A separate legal personality for AI would also resolve the AI decision dilemma and help advance innovation in AI. If the users or creators of an AI know that, subject to certain exceptions, they will not be held liable for its actions, they will allow the AI to make its own decisions even when they disagree with the outcome, and will intervene only when they believe it necessary. Recognizing AI’s legal personality would even enable recovery for damage caused by a rogue AI system whose actions are entirely beyond the control of its user or creator.

Although this approach would require governments to devise a new and complex regulatory regime for fully autonomous AI systems, it is the most future-proof of the available options, as it provides courts with a definite route for deciding liability issues arising from AI-made mistakes.

4.Benefits of a Separate Legal Personality: The Corporate Model as an Official Non-Human Legal Entity

Section 3 showed that assigning liability to the AI itself may be the most appropriate way of addressing AI-made mistakes and the damage they cause. This section examines the concept of AI as a separate legal person by considering what such a personality would look like and what benefits it would offer. It does so by exploring the history of another artificial legal entity: the corporation. Corporate legal personality is perhaps the most well-established artificial personality in the world, and its success can serve as a valuable source of inspiration for designing the next generation of artificial personalities and, as some have suggested, as a model for a potential AI civil or criminal liability system.52

While there are significant differences between AI systems and corporations, if AI is to be recognized as a legal person, it is essential to analyze the vision behind humanity’s desire to grant a separate personality to non-human entities and how that vision might shape AI’s eventual path to recognition as a legal entity. This section therefore explores the origins of corporate legal personhood, why countries decided to confer such legal status on corporations, how this model can be applied in the context of an AI person, and how the AI model and the corporate model of legal personality might differ.

To better understand the history of the corporate model of legal personality, however, it is important first to distinguish between two concepts that coexist in today’s corporate model: the corporation’s separate legal personality and the limited liability of its shareholders. The former allows the corporation to exercise rights such as buying and selling property, borrowing money, and suing. The latter, also known as owner shielding,53 is enabled by corporate personality and protects shareholders’ personal assets from creditors and other shareholders.

4.1Origins of Corporate Legal Personality

The concept of corporations as separate legal entities has existed in one form or another since time immemorial.54 Since medieval times, corporations could be formed by the authority of Parliament, through a Royal Charter, or with some other form of permission from the king, and they were considered distinct entities, separate from their members. However, this form of corporation was mostly reserved for public service entities, such as local governments, churches, universities, and hospitals.55 It was around the beginning of the seventeenth century that people began to recognize the economic benefits of corporations and to form commercial corporations.56 The Dutch and English East India Companies and the Hudson’s Bay Company were among the first commercial corporations incorporated by charter.57 The idea behind these newly incorporated companies was simple: to gather investors and seek their financial support and, in return, to give them a share of the company in the form of joint stock, allowing them to participate in the profits generated by the company or to sell their shares at a later date.58

Limited liability of shareholders in joint stock companies, in its modern sense, did not emerge until the nineteenth century. For limited liability to have become an issue, these companies would have needed to accrue some form of debt. However, most early joint stock companies were financed exclusively through equity, and debt represented a relatively small share of their sources of financing.59 This was mainly due to “the high level of uncertainty in the early stages of trade, the high level of on-going risk, the absence of tangible collaterals, and the non-existence of commercial banks or bond markets of the magnitude that could provide long-distance finance.”60 Limited liability issues would have arisen only if a claimant had been unable to recover all of its losses from the corporation because the corporation was insolvent.61 Even when these companies did accrue debt, their close relationship with the government usually meant that the government would assist them in paying it off.62 After all, these companies were created by Royal Charter, with the exclusive permission of the Crown or Parliament, and the government had at least some form of stake in them, which made them quite safe from insolvency.63

A further indication that limited liability was not a significant issue is the creation of unincorporated joint stock companies in Britain in the second half of the eighteenth century.64 Given the reluctance of Parliament and the Crown to grant a royal charter to every joint stock company because of the prevalence of corruption and economic misuse,65 unincorporated joint stock companies were seen as a means of circumventing the charter requirement and enabling investors to purchase shares in companies. These unincorporated companies were not recognized as separate legal persons and, as a result, did not shield their shareholders from liability to third parties and other shareholders. Shareholders were aware that they were fully exposed to tort and contractual liability, and yet they voluntarily invested in these companies. This indicates that limited liability was not considered a significant issue in the joint stock companies of the seventeenth and eighteenth centuries. It is worth noting, however, that the shareholders of both unincorporated and incorporated joint stock companies enjoyed some form of liability shield through custom, practice, and sometimes contract,66 although this was not an official, attributable status of such companies.67

The nineteenth century brought two significant changes to the status of joint stock companies: free incorporation and the recognition of limited liability by some governments. A key development in the evolution of companies during this period was the arrival of free incorporation, which meant that companies could incorporate at will without obtaining the permission of the Crown or Parliament.68 New York was the first jurisdiction in the world to enact a general incorporation law, “The Act Relative to Incorporation for Manufacturing Purposes of 1811,” which, as its name suggests, initially opened the door to incorporation only for manufacturing companies.69 Over the following decades, many states adopted similar laws but heavily restricted the activities of corporations and the amount of capital they could raise, owing to fears that such laws would exacerbate inequality in the United States.70 The UK Parliament also demonstrated its support for free incorporation by enacting the Joint Stock Companies Act 1856 and, later, the Companies Act 1862, both of which provided that seven or more persons could form themselves into an incorporated company.71

The concept of limited liability in its modern sense, on the other hand, took much longer to be adopted.72 The first signs of official recognition of limited liability in the UK appeared in the Limited Liability Act 1855 and the Companies Act 1862, under which corporations were permitted to allocate to their shareholders the liability of their choosing, which could, but did not have to, include complete shielding of shareholders from liability.

The concept evolved throughout much of the nineteenth and twentieth centuries to become the default mode of liability in many jurisdictions, such as the UK, Canada, and many US states. Under the UK’s Companies Act, for example, one can register a limited company whose shareholders’ liability is “limited to the amount, if any, unpaid on the shares held by them.”73 Canada’s Business Corporations Act and New York’s Business Corporation Law both provide that the shareholders of a corporation are under no obligation to pay the corporation’s debts except to the extent of their ownership of the corporation’s shares.74

In jurisdictions that recognize their separate legal personality, corporations enjoy privileges that include not only rights such as the right to sue, to hold property and other assets, and to enter into binding contracts, but also rights that more closely resemble those of humans, such as freedom of expression. In the United States, for example, Justice Anthony Kennedy of the US Supreme Court wrote in Citizens United v FEC that limiting or prohibiting independent political spending by corporations and other non-profit groups violates the First Amendment right to free speech, thereby extending freedom of expression rights to corporations.75

4.2The Benefits of Corporate Legal Personality

Although the concept of the corporation as a separate legal person has existed for many centuries, the idea of using corporations for for-profit enterprises was nothing short of extraordinary. The ability to create an entity that pools resources from many investors and uses them to generate profit, combined with the arrival of limited liability, was a great step toward the expansion of the economy. In an often-cited speech before the New York Chamber of Commerce in 1911, Nicholas Murray Butler, the President of Columbia University, called limited liability companies “the greatest single discovery of modern times,” the value of which rendered the discovery of the steam engine and of electricity far less important.76

Perhaps some of the most convincing evidence of the effect of corporations on the economy comes from a comparison of the economic growth of the Middle East and Western Europe. In his book The Long Divergence, Timur Kuran seeks to explain why the Middle East, once a hub of trade and economic activity, gradually fell behind Europe’s economy. Over the seven centuries from 1000 to 1700, the Middle East’s share of world gross domestic product fell fivefold, from 10.3% to 2.2%, while Europe’s share more than doubled over the same period, from 9.1% to 21.0%.77 Kuran argues that one of the main reasons for the divergence between the two regions’ economies was the arrival in Europe of commercial corporations and other similar forms of business. While most Middle Eastern countries were unable to form corporations as separate legal entities because Islamic law excluded such concepts, corporations were thriving in Europe. By way of comparison, the first predominantly Muslim-owned joint stock company did not come into existence until 1851,78 by which time European commercial joint stock companies had existed for at least two centuries. Kuran suggests that the persistent simplicity of Islamic partnerships, which usually ended upon the death of a partner, prevented Islamic countries from forming long-lasting companies. Social service organizations, such as charities and religious bodies, took the form of trusts and were not self-governing. These institutions were bound to follow the directives of their founders even after the founders’ deaths, which restricted their ability to maximize profitability at the discretion of their caretakers.79 Accordingly, Islamic rules and policies, while laudable in purpose, hindered Muslims’ ability and motivation to form long-lasting corporations, which contributed to the decline of the Middle Eastern economy and of its share of the world economy.

Kuran correctly suggests that the invention of corporations and limited liability was transformative for Europe’s efforts to expand its economy. Limited liability allowed investors to purchase shares in various companies freely, without worrying that every purchase would place their entire assets at risk. Without limited liability, investors would probably have concentrated all their investments in a single firm, one whose operation and management they could personally control and oversee.80 Corporate limited liability allows investors to put as much money into companies as they reasonably deem appropriate while ensuring that their uninvested assets are safe from liability and from corporate creditors.81 In other words, “Limited liability encourages the development of public markets for stocks and thus helps make possible the liquidity and diversification benefits that investors receive from those markets.”82 Today, corporate personality is so foundational in some jurisdictions that courts refuse to impose personal liability on shareholders even where the corporation has only one shareholder or was formed for the specific purpose of limiting or escaping liability.83

The effect of corporations on today’s economy is undeniable. According to research by the McKinsey Global Institute, corporations have contributed significantly to the economic growth of the OECD84 countries, with more than half of the OECD countries’ GDP in 2017 coming from corporations.85 The gigantic industrial projects of today cannot be accomplished with capital provided by a few people; rather, they require a significant number of people to be interested in investing in them.86 Investors would presumably be less willing to invest money in a corporation if they knew they were bearing a greater risk than the amount they were contributing. “Evidently people will not invest and surrender the control of their investment if they must also pledge their personal credit to make good the risk of possible loss.”87

4.3The Corporate Model as an Inspiration for AI’s Legal Personality

Given the justifications for recognizing the separate legal personality of corporations and the benefits it entails, the corporate model can serve as an important source of inspiration for designing an AI liability model. The need to address liability issues arising from AI-driven mistakes, combined with the potential of an AI, as a separate legal entity, to increase its profitability significantly by accessing new financial opportunities, makes the application of the corporate model to AI a suitable approach.

However, not every AI system should be recognized as a separate legal entity, both because of the risks that AI systems pose and because of their widely varying degrees of intelligence and autonomy. A smart vacuum, for example, would not qualify as a legal person; nor should an AI designed to operate nuclear missiles. For this reason, if governments decide to recognize the legal personality of AI systems on the basis of the corporate model, the framework should resemble that applicable to early corporations, which were permitted to incorporate only with the explicit permission of the government. This point is elaborated further in Section 6. A separate personality for AI would allow claimants to sue the AI itself rather than every entity involved in the incident, which would make it easier for courts to resolve such liability cases. This is especially beneficial where hundreds or thousands of people are involved in designing and writing the code for sophisticated AI systems. The rights allocated to the AI person, which would be similar to those of corporations, such as the right to sue and to enter into binding contracts, would also offer several advantages. For example, they would give the AI’s designers and shareholders standing to sue on behalf of the AI for damage caused to it by the user’s improper use or disregard of the terms of the contract, such as intentionally using the AI to commit fraud and thereby damaging its reputation.

Furthermore, an AI system granted a separate legal personality and the ability to shield its shareholders from liability could encourage investors to contribute to the AI person’s financial resources and, in return, receive a share of the profits made by the AI person in proportion to their shareholdings. This would also mean that investors would not need to be involved in the day-to-day operation of the AI person, since any liability stemming from the AI person’s mistakes would be attributed to the AI as an entity, preventing claimants from pursuing the investors directly. Knowing that they would not be assuming any risk greater than the amount they invested in the AI person, investors could purchase as small or as large a share of the AI person as they wished.

4.4Fitting AI into the Current Corporate Model vs. Creating a New Personality Model

Given the similarities between the AI personality model and the corporate personality model, there has been debate as to whether AI’s legal personality could be recognized by fitting the concept into the corporate model. Could AI systems, for example, simply be registered as corporations or as another type of registered legal entity? Shawn Bayern suggests that the flexibility of modern corporate law could allow autonomous systems to be recognized as separate legal entities without any change to the current regulatory framework.88 Bayern argues that limited liability companies (“LLCs”) can accommodate the concept of an AI as a separate legal person by allowing the creation of “memberless LLCs,” which are valid legal entities. He describes the process as follows:

“(1) an individual member creates a member-managed LLC, filing the appropriate paperwork with the state; (2) the individual (along, possibly, with the LLC, which is controlled by the sole member) enters into an operating agreement governing the conduct of the LLC; (3) the operating agreement specifies that the LLC will take actions as determined by an autonomous system, specifying terms or conditions as appropriate to achieve the autonomous system’s legal goals; (4) the sole member withdraws from the LLC, leaving the LLC without any members. The result is potentially a perpetual LLC—a new legal person—that requires no ongoing intervention from any preexisting legal person in order to maintain its status.”89

Bayern also emphasizes that, while LLCs are usually not permitted to remain memberless indefinitely, in some jurisdictions, such as New York, the period during which they may remain memberless is often left to the discretion of the drafter. The initial members of the LLC could therefore provide that the LLC will not require a member for a very large number of years.90

While this would admittedly be a more convenient approach to assigning legal personality to AI systems, it raises certain challenges that call for recognition of the notable differences between a hypothetical AI legal personality model and the corporate legal personality model. There are several reasons why the corporate legal personality model should not be applied to AI systems without significant regulatory change.

The first reason is that Bayern’s framework applies only to LLCs and not to corporations. Corporations in most jurisdictions require directors who are natural persons. The Delaware General Corporation Law, for example, provides that “[t]he board of directors of a corporation shall consist of one or more members, each of whom shall be a natural person.”91 Furthermore, New York requires that board members be at least eighteen years of age.92 Even where the natural person requirement is not expressly stated, courts may well interpret such laws as implying that directors must be natural persons. AI systems therefore could not be recognized as separate legal entities in the form of corporations and would not enjoy benefits such as the ability to issue stock, which would limit their economic viability.

Moreover, applying the current regulatory framework for corporations to AI systems without change is undesirable because of the AI’s autonomy in decision-making. Although a corporation is a separate legal entity at law, its decisions are made entirely at the direction of its board of directors. In other words, while the legal consequences of a corporation’s decisions belong to the corporation, the actual decisions are made by a group of humans. An AI, on the other hand, can decide matters on its own: it can choose to purchase or sell property, enter into a contract, or take any other action without ever needing a human to make that decision for it. The type of control that humans exert over corporations is therefore different from the type they would exert over AI systems.

The second reason is that allowing AI systems to be recognized as a type of memberless LLC raises challenges that governments may be unable to address without enacting new laws. The first challenge would be controlling which AI systems are granted legal personality. Under Bayern’s model, any AI system, whatever its degree of intelligence, could potentially be registered as a separate legal person. This could lead to misuse of the separate personality model by allowing the creators of lower-intelligence AI systems to shield themselves unfairly behind the owner-shielding features of LLCs. The second challenge is that, under this model, the extent of the effects of an AI’s decisions would be unclear. It would require determining whether the AI decides who receives the financial benefit of the profits made by a memberless LLC (which has no issued stock or shares) or who may join the LLC as a member. If these decisions are made by a human under the terms of the LLC, the concept of the AI’s separate legal personality disappears, and the arrangement becomes one in which a human legal entity is directing an AI to perform certain tasks.

In conclusion, governments would be better served by creating a new framework for AI’s separate legal personality than by allowing AI persons to exist within the current framework. Although an AI legal personality model could have much in common with the corporate model, AI’s autonomy in decision-making and its evolving intelligence require lawmakers to create a new regulatory framework, one specifically designed to address AI’s unique characteristics.

5.Challenges in Granting Legal Personality to AI Systems

Creating a framework for new artificial legal entities such as AI systems would certainly be challenging. As noted above, regulating AI persons would require an understanding and recognition of AI’s unique features, such as its autonomy in decision-making. To create an appropriate regulatory framework, governments will have to consider every aspect of these challenges and how they might be overcome or resolved. This section discusses some of the challenges that may arise in regulating AI persons and how governments might resolve them.

5.1Defining an AI Person and Setting a Threshold for AI Personhood

AI systems come in a wide variety of forms and capabilities. From robot vacuums and smart lights to large language models such as OpenAI’s GPT-4, Google Bard, and IBM’s Watson, all of these systems are considered intelligent. Plainly, however, not all of them should qualify for a separate personality. The central question for lawmakers is therefore how to determine the eligibility of AI systems for personhood. In other words, what makes an AI system intelligent and capable enough to be recognized as a legal entity?

In assigning legal personality to AI systems, governments must define a specific threshold that an AI system would need to meet in order to qualify for personhood. This threshold would be set on the basis of the criteria that governments put in place for AI systems to satisfy in order to obtain legal personhood. Formulating those criteria and setting the threshold for legal personality will certainly be one of the main challenges facing governments worldwide.

A central issue in setting this threshold is determining which characteristics of an AI are relevant to granting it legal personhood. In other words, when governments decide whether to grant a specific AI legal personality, which of its features should they consider in assessing its eligibility? Luciano Floridi and J W Sanders identify three key features that must be present in artificial agents:

“(a) Interactivity means that the agent and its environment (can) act upon each other. Typical examples include input or output of a value or simultaneous engagement of an action by both agent and patient - for example gravitational force between bodies. (b) Autonomy means that the agent is able to change state without direct response to interaction: it can perform internal transitions to change its state. So an agent must have at least two states. This property imbues an agent with a certain degree of complexity and independence from its environment. (c) Adaptability means that the agent’s interactions (can) change the transition rules by which it changes state. This property ensures that an agent might be viewed … as learning its own mode of operation in a way which depends critically on its experience. Note that if an agent’s transition rules are stored as part of its internal state … then adaptability follows from the other two conditions.”93

Interactivity, autonomy, and adaptability are all critical to any consideration of granting legal personhood to AI and could serve as a starting point for governments in setting a threshold for AI legal personality. However, the definition of autonomy that Floridi and Sanders use is less restrictive than what would be desirable for an AI person. While this definition excludes most low-level AI systems, such as smart thermostats, from consideration for legal personhood, it still leaves AI chatbots, including those developed before the paper was written in 2004, as candidates for personhood, since they are interactive, adaptable, and autonomous. This paper proposes that, while governments must consider all the characteristics of AI systems in assessing their eligibility for personhood, they should focus on the autonomy of AI systems, as autonomy is likely to be the decisive factor in AI personhood.

Governments should consider granting legal personality only to fully autonomous AI systems, which can be defined as AI systems that are not only capable of operating independently of human input but also prioritize their own output over that of humans. Fully autonomous systems can override human intervention and will take it into account only if it aligns with their preferred outcome. For example, self-driving software would not be considered fully autonomous if humans were able to override its route or intervene in its braking system. Fully autonomous self-driving software would receive the user’s braking input and would apply the brakes only if it judged that to be the right action. This is a high threshold for AI systems to meet, but it may be necessary if an AI is truly to be held accountable for its actions. It is also worth noting that, while autonomy is a necessary consideration, it is not by itself sufficient to qualify an AI system for legal personhood. Governments must take a holistic approach, considering all the relevant criteria that make an AI system worthy of such status.

Finally, governments must also future-proof the definition of an AI person to accommodate the arrival of superintelligent AI systems. Terms such as “computer” will soon be outdated as references to AI systems. While AI systems are designed using computers, they are not themselves computers; they are, in a sense, as much a computer as the human brain is.

5.2AI Income and Taxation

As AI persons begin to hold financial assets and generate income, their taxation will become a challenging issue for governments. A tax personality must be developed alongside the AI legal personality to address the issues arising from the potential taxation of AI persons. Taxing AI persons is crucial because, once they begin replacing the human workforce in certain industries, governments will see a significant drop in the tax revenue that would otherwise have been collected from humans through their salaries. This problem may be exacerbated by the fact that the financial burden on governments may increase in the short run as AI systems replace the human workforce, causing unemployment and increasing the need for social security.

Currently, the activities of AI systems are taxed only in the hands of the legal entities using these systems, based on the type and amount of income generated. With respect to the activities of AI persons, however, governments must consider the taxation of two different entities: the entities that use AI persons as objects or as replacements for their human employees,94 and the AI itself as a legal entity that earns income. To tax the use of AI persons by other entities, such as firms and individuals, governments could examine the idea of taxing the imputed salary, that is, the salary the entity would have paid to human employees had it chosen not to use the AI to perform the same tasks.95 Another option would be to introduce an automation tax on entities using AI persons, which could be calculated based on the contribution of human employees to the entity’s revenue.96 A third option would be to introduce specific tax deductions for the use of a human workforce, or to impose additional taxes on the use of AI persons as employees, in order to discourage human unemployment.97 This would be similar to governments levying additional taxes on products such as tobacco to discourage smoking. Governments could also consider indirect means of compensating for the loss of tax revenue caused by AI persons, such as increasing the corporate tax rate or capital income taxation.98

Taxing AI persons as separate entities, on the other hand, would require the development of a new tax scheme to accommodate their activities. Granting AI systems legal personality is not sufficient to recognize them as taxpayers. Rather, to be identified as such, an AI person should be “(i) defined for tax law purposes; (ii) recognized as a taxable entity; and (iii) the tax should be fair and equitable.”99 This would depend heavily on the tax scheme that governments decide to impose on AI persons.

A central challenge in designing a tax scheme for AI persons would be determining whether to levy revenue taxes on AI persons, as with corporations, or income taxes, as with human employees, since AI can play both roles in society. As a corporation-like artificial legal entity, an AI person can either distribute its profits to its shareholders or reinvest them in future projects, which means that it would have to be subject to revenue taxes, as corporations are. However, as noted above, AI persons can also take on the role of company employees and earn employment income. In that case, the AI would resemble a human, and governments would have to levy income taxes on the AI person; it would seem unfair to subject AI persons working as employees to a lower, corporate-like tax rate when humans performing the same tasks would be taxed at a higher rate. On the other hand, taxing AI persons as employees could create a risk of double taxation, whereby AI persons would be taxed once on their income and again on the revenue they generate.100

One way for governments to resolve this issue would be to create a dual tax personality for AI persons, taxing them at a lower rate when their work resembles that of corporations and at a higher income tax rate when they act as employees and replace human jobs. In the latter case, governments could tax AI persons’ income in the same way that human income is taxed. This would also make AI persons responsible for paying other forms of taxes, such as sales taxes. In determining the appropriate tax rate for AI persons relative to humans, governments must consider the differences between AI persons and humans with respect to their spending. AI persons, for example, would not have to pay for healthcare or childcare, nor would they need pensions.101 This could give governments an incentive to tax AI persons’ income at a higher rate than human income.

5.3Reasonable AI vs. Reasonable Human

The concept of the reasonable person is an important legal standard used in many jurisdictions to assess the civil or criminal liability of individuals. The reasonable person standard is often used in public and private law as the benchmark against which a person’s conduct in a particular situation is evaluated. The standard is not static and varies with the circumstances; the conduct expected of a reasonable individual in ordinary circumstances differs from that expected in an emergency. Halsbury’s Laws of Canada describes what is expected of the reasonable person as follows:

“The law has required of everyone a minimum level of performance, whether they are capable of it or not. It is not enough for an actor to plead that he or she did ‘the best he knew how’, for the law ‘does not attempt to see men as God sees them.’ … The reasonable person is not perfect, however, for even ‘the most excessively careful man will sometimes have an accident’. The reasonable individual is ‘not a seer’, nor a ‘paragon of circumspection’, not a ‘person of infinite-resource-and-sagacity’, nor does that individual have the ‘prophetic vision of a clairvoyant’. ‘Tort law does not require the wisdom of Solomon.’ ‘The standard is not perfection, or even the optimum, judged from the vantage of hindsight’, but reasonableness only. Consequently, one may be guilty of an error in judgment, without incurring tort liability, as in the case of ‘pure misadventure.’ The reasonable person is not obligated to ‘exercise the best possible judgment’ in an emergency, nor to avoid all possible risks. What the reasonable human must do is act in accordance with ‘normality and practicality’ rather than ‘judicial technicality or diversity of view.’”102

A critical question for courts is whether the same standard of reasonableness should apply to AI persons or whether they should be held to a higher standard: that of a reasonable AI. While it will ultimately be for the courts to decide which standard applies to AI persons, it is reasonable to expect that courts will hold AI persons to a higher standard than humans, because an AI person’s capacity to process information far exceeds that of a human. Faced with danger, for example, an AI system can analyze and process significantly more scenarios before taking action than a human could process even a fraction of. Given such capacity, it would be logical to subject AI persons to a higher standard of reasonableness when they make mistakes.

6.A Regulatory Framework for Governing AI Persons

Allowing AI systems to be recognized as legal entities requires a comprehensive framework that sets out the rights and responsibilities of AI persons while taking into account challenges such as those discussed in Section 5. This section proposes a regulatory framework that addresses many of the concerns raised in this paper. In the author’s view, this framework offers the best starting point for balancing the different interests of all stakeholders in AI liability issues.

6.1Definition of an AI Person: The Eligibility Criteria

A definition of an AI person must be established to enable the relevant parties to distinguish AI systems that are eligible for recognition as persons from those that are not. In a legal sense, this will create two distinct categories of AI systems: AI persons, which are recognized as separate legal entities, and other AI systems, which have not been granted such personality. To be recognized as a separate legal person, an AI system must possess certain characteristics, which together form a personhood test allowing governments to determine whether to recognize the AI’s separate personality. While governments must consider these requirements holistically, as noted in Section 5, the AI’s autonomy must be prioritized in deciding whether to grant personhood, and the criteria must also be future-proof, so as to remain cognizant of AI’s ever-increasing capabilities. A non-exhaustive list of the characteristics that an AI person must possess is as follows:

  • Ability to operate completely autonomously: The AI must not require any form of supervision to operate, and humans must not need to intervene in or correct its behaviour. The AI must receive input and evaluate that input independently before producing an output. Any successful human attempt to adjust or replace the AI’s output will render the AI ineligible for legal personhood.
  • Ability to override human input: This criterion relates directly to the AI’s autonomy in decision-making. It means that, while the AI may still receive input from humans, that input need not change the AI’s final output. The final decision on whether to disregard human input rests solely with the AI.
  • Dynamic learning: An AI person must be adaptable and must learn to adjust its behaviour based on user input. For example, simple classification models that have been trained on a data set to classify different objects should not be considered for legal personhood, no matter how much training data they have received, because these systems stop learning once they leave their creators’ hands and are therefore unable to adjust their behaviour and classify objects differently. If, however, they are allowed to learn as they are fed new data and are able to reclassify objects when told that their output is wrong, those systems can be said to be adaptable. If learning ends once the AI leaves the manufacturer’s hands, the AI should not be eligible for personhood, because the fault in such static models would most probably lie with their creators.
  • Ability to perform tasks at, or at least near, the level of an average human: This is the only intelligence-specific requirement that should be used in defining an AI as a legal person. As long as the AI is capable of producing work that at least resembles that of an average human, it should be considered sufficiently intelligent for legal personhood. It is worth noting that this ability alone is not sufficient to establish AI legal personality; rather, it must be considered alongside the AI system’s other characteristics in determining its eligibility.

6.2Liability Scheme

Once the two legal categories of AI systems have been introduced and acknowledged, a hybrid model of AI liability and strict liability will be introduced to impose liability for AI-made mistakes. This hybrid model consists of the following allocations of liability:

  1. AI systems recognized as legal entities will be treated in the same way as any other legal entity. They can be held liable in both civil and criminal cases, and no other entity need share liability with them unless the courts consider it appropriate.
  2. Every other AI system that has not been granted legal personality will be governed by a strict liability scheme. Governments can choose to assign strict liability to users (as in the UK), owners (as in Germany), creators, or a combination of them (as proposed in the EU report), depending on the circumstances.

This liability scheme would give courts a definite and more convenient pathway for assigning liability when AI systems cause damage or commit crimes, and it would remove uncertainty from AI design, as AI programmers could create and adjust their designs in accordance with the liability rules.

6.3Creation of an AI Commission

A necessary step in regulating AI persons is the creation of a commission (for example, an “AI Persons Commission”) composed of legal scholars and AI experts and delegated the power to oversee the activities of AI persons. The Commission’s tasks could include, but would not be limited to, reviewing and approving the eligibility of AI systems deemed fit for legal personhood and revoking such eligibility where required; establishing a panel to hear disputes arising from the granting of legal personality to eligible AI systems (which would not include civil cases against AI persons); and providing training and manuals to the Commission officers tasked with reviewing applications for AI legal personality. The Commission could also be tasked with making recommendations to Cabinet regarding the regulations it believes must be enacted to govern such AI systems properly.

6.4Registration Requirement

AI persons need some form of identity in order to be recognized when interacting with humans or other entities, and they must be identifiable through a verification system. For this reason, granting legal personhood to AI systems requires that, like corporations, they be registered and given a unique, verifiable digital identity that cannot be replicated. This is intended to prevent identity tampering and can be achieved using new technologies such as distributed ledgers and blockchain.103

The registration of AI persons would serve several purposes. First, it would allow AI persons to interact with humans transparently, which is a requirement for AI systems in jurisdictions such as the EU. The EU, for example, has envisioned that high-risk AI systems be registered and developed “in such a way that the natural persons are informed that they are interacting with an AI system.”104 Registering AI persons would make users aware that they are interacting with an AI and would enable them to verify the AI person’s identity using its unique identifier. This will be critical as AI systems begin to mimic human behaviour, and it will also enable people to determine whether they are dealing with a registered AI person or an unregistered AI product, which is necessary if they decide to bring a lawsuit against that AI system. Second, registration would give people the opportunity to decide whether they wish to interact with an AI person. This would allow AI systems to comply with rules against subjecting humans to automated decisions against their will, such as those set out in the GDPR.105

The third purpose of registration is to distinguish AI persons, which would be registered as such, from other AI systems that have not qualified for personhood. Registration is perhaps the most convenient means of informing others of the status of an AI person and of allowing them to enter into binding contracts with, or to sue, the AI person. Without registration, people might find it difficult to distinguish AI persons from other AI systems, which could lead some parties to fraudulently hold themselves out as AI persons when, in fact, they are not.

The operation of the hybrid liability model will also depend on the registration of AI systems. As with corporations, once an AI’s registration is complete, it will be recognized as a separate legal entity and granted the prescribed bundle of rights and responsibilities associated with AI persons. Unregistered AI systems, or those whose applications for registration have been denied, will be deemed AI products, and a human entity (the user, owner, or creator) will be held strictly liable for their actions.

The registration process for AI systems, however, would not be as simple as that for corporations. Whereas registering a corporation requires the submission of a relatively simple application to the government, with no approval process, registering an AI system must require the approval of the AI Commission. This is because not every AI system should qualify for legal personhood, and the decision to grant legal personhood to an AI system should rest with the AI Commission.

Governments must establish a process through which an application for an AI’s registration as a legal person can be filed with the Commission. The application must be filed by the owner of the AI system, which could be a company or the AI’s creator(s). The owners of the AI system would also be recognized as the first stakeholders of the new AI person. The submission to the Commission must be accompanied by information about the AI and its capabilities to assist in determining its eligibility for personhood. The submission could include the following information:

  • The type of machine learning (supervised, unsupervised, or reinforcement learning) used to train the AI.
  • The data sets used to train the AI system. This is to ensure that the data has been collected and processed in accordance with the data collection requirements of each jurisdiction.
  • The purpose of the AI’s design, the field(s) in which its creators envision it operating, and any areas in which the AI should be restricted from operating. This information could be used by the Commission or the courts to identify misuse of the AI’s capabilities.
  • The complete code of the AI’s algorithm. This could serve multiple purposes, such as ensuring that the code does not infringe any other person’s IP rights and running tests on the code to ensure its safety and its compliance with the relevant laws and regulations.
  • Information about all tests that have been performed on the AI. This is for determining the AI’s level of risk and ensuring that it is safe to use and does not give rise to any ethical violations. Any updates or changes to the source code must also be reported in order to maintain the AI’s registration status, as the Commission would consider only the latest version of the AI system for personhood.

The Commission must review the application and the information submitted with it and run tests to ensure that the AI system satisfies the standards established for personhood eligibility and complies with all relevant laws and regulations. Once the Commission is satisfied that the AI meets the criteria for legal personhood and complies with all relevant laws, such as data privacy and ethical regulations, it can approve the legal status of the AI system, register the AI as a legal person, and assign it a unique, verifiable identity.

Finally, the Commission must also establish a procedure under which, in certain circumstances, an AI’s registration as a legal entity could be temporarily or permanently revoked. These circumstances could include cases in which the AI person is later found to be ineligible for personhood following a review of the criteria by the Commission, cases in which the AI is found to make an unreasonable number of mistakes, and cases in which the AI’s criminal conduct is punishable by shutting down the AI system.

6.5Encrypted Recording Requirement

Governments may choose to require AI persons, and possibly even other AI systems, to keep a record of their data processing and to store it locally or in cloud storage. This approach resembles the “record-keeping” requirement of the EU AI Act106 and Germany’s recording requirement for autonomous vehicles.107 Such a requirement can help determine liability where disputes arise as to which actions led to a mistake and who was responsible for it. The AI person could be required to release the recording to the relevant authorities upon request to assist them in reaching a decision. The recording could later be used by the AI Commission to decide whether to revoke the AI’s registration as a legal person based on the extent of the AI’s mistake.

The requirement could operate by obliging AI persons to have local or cloud storage available at all times during their operation. At a minimum, an AI person should be required to record and store data such as its operating time, software status, power supply, inputs received from sensors and cameras, any user commands (such as conversations, texts, or physical inputs), and any other data exchanged or communications with external bodies.108

Given the wide variety of functions performed by different AI systems, each system might be required to record and store different types of data relevant to its operations. Self-driving car software, for example, might be required to record information such as the position, speed, or acceleration of the vehicle,109 whereas such information would be irrelevant to an AI accountant, which could instead record information such as the orders it has placed on the stock market for its clients.

All input data must be collected and processed with the consent of the persons interacting with the AI person and in compliance with the data collection and privacy regulations of the jurisdictions in which the AI person operates. The collected data must also be stored using the latest encryption technologies in order to comply with provisions such as the “appropriate security” requirement of the EU’s GDPR.110 Governments could also prescribe how long the recordings must remain available and adopt regulations on how the data is to be safely discarded.

6.6Disregarding AI’s Legal Personality: Piercing the “Digital Veil”

Although the personal assets of corporate shareholders have long benefited from the immunity conferred by limited liability, this immunity is not absolute. In certain circumstances, courts have disregarded corporate legal personality and held shareholders personally liable, a doctrine known as “piercing the corporate veil.”111 Over the years, courts in different jurisdictions have set out the conditions under which they are willing to pierce the corporate veil. Courts in many jurisdictions have held that, where a corporation is used for fraudulent purposes,112 has no personality separate from that of its shareholders, or is a mere façade for the misuse of corporate limited liability, the corporate veil may be pierced and the shareholders held personally accountable.113

As with corporations, some AI systems may be registered and granted legal personality while serving as a mere façade for the covert operations of their true owners. Where courts find that an AI person has no independent agency of its own or is operating to conceal the true identity of fraudulent actors, they may decide to pierce the AI’s “digital veil” and pursue the entities behind the AI’s actions directly. The extent to which entities will hide behind the digital veil is currently unknown, and it will be for the courts to determine the rules governing the piercing of an AI’s digital veil and the assignment of liability to other actors.

6.7Prohibition of AI Personality in Critical Fields

The risk of fully autonomous AI persons operating in certain critical fields, such as military or nuclear operations, may be too great for governments to be willing to bear.114 Governments could, and probably should, therefore choose to prohibit the granting of legal personality to AI systems in fields in which mistakes could have catastrophic consequences, such as the outbreak of war or a nuclear disaster. This would be akin to adopting a modified version of the “human oversight” requirement in the EU AI Act, under which the requirement would apply not to every high-risk system but only to a very specific set of highly critical systems in which the risk posed by AI autonomy is too great.

6.8Responsibilities and Rights Allocated to AI Persons

6.8.1Responsibilities

Once recognized as legal entities, AI persons would probably bear responsibilities resembling those of other legal entities, such as humans and corporations, although the extent to which these responsibilities would apply to AI persons might differ. At the most basic level, the following responsibilities could be envisioned for an AI person:

  1. Compliance with laws and regulations: AI persons will have to abide by the law of the land to the extent that it applies to them. This would include all laws and regulations, from the offences in the Criminal Code to privacy and data protection laws. Failure to comply with the law will expose AI persons to punishment.
  2. Civil and criminal liability: An AI person would be liable for any harm or damage it causes as a result of its actions. As legal entities, AI persons would have a duty to obey the laws of the jurisdiction in which they reside, and they could be sued like any other entity.
  3. Contractual obligations: An AI person would be bound by the terms of the contracts it enters into, such as licensing agreements or service contracts. Failure to adhere to the terms of a contract would allow the other party to sue the AI person in court.
  4. Ethical considerations: Governments may wish to oblige AI persons to abide by specific rules in order to uphold ethical standards. These could include, to the extent possible, transparency and accuracy requirements in the decision-making process.

6.8.2Rights

Since AI persons would be capable of being held liable for damages, governments must allocate certain rights to them so that they can discharge those responsibilities. In particular, AI persons will need some form of compensatory scheme in case they are found liable for their actions. To achieve this, governments may decide to grant AI persons rights similar to those of corporations, including, for example, the right to sue and be sued in all courts and to participate in any actions or proceedings, and the rights to purchase, own, and sell property, to enter into contracts, and to lend money.115 Governments could also choose to extend to AI persons rights such as freedom of expression, which US courts have recognized as extending to corporations.116 However, governments must also remain cognizant of the evolving nature of AI and recognize that these rights and responsibilities may have to change in the future.

7.Conclusion

The increasing presence of AI in everyday life is inevitable. As AI assumes diverse roles within society, its interaction with humans is likely to result in incidents. Once AI systems become sufficiently intelligent to operate autonomously, the liability regimes that currently govern them, including strict liability and non-strict product liability, will become inadequate for regulating their actions. This is due to the unforeseeable nature of AI actions and to the black-box algorithms that make it difficult for users or creators to predict every outcome of an AI system. Governments must therefore begin considering innovative ways of regulating AI systems capable of complete autonomy once such systems become available, and they must do so expeditiously, as the arrival of these systems may leave it too late to regulate them effectively. An alternative approach would be to recognize AI as a legal entity, similar to a corporation, and to assign liability to the AI itself. Under this model, governments can establish the criteria an AI must satisfy to be granted legal personality. Once approved, an AI person must be able to hold rights such as owning, purchasing, and selling property and assets, and entering into binding contracts, so that it can be held liable for its actions. Governments must also consider imposing certain rules and restrictions on the operation of AI persons, such as a mandatory recording requirement obliging the AI to record its entire operation, and restrictions preventing AI persons from operating in critical fields such as the military and nuclear sectors. Any AI that does not meet the criteria for personhood can be regulated under a strict liability regime, under which the creator or user can be held liable for the AI’s actions. The advantage of this model for governments is that it can be implemented immediately without granting legal personhood to any current AI system. All such systems can be regulated under the strict liability regime until they satisfy the criteria set by governments for personhood eligibility.

Notes

  1. Stephen Hawking, Brief Answers to the Big Questions, (New York: Bantam Books, 2018) at 184. ↩
  2. Geoffrey A Fowler, “ChatGPT can ace logic tests now. But don’t ask it to be creative.” (18 March 2023), online: The Washington Post <https://www.washingtonpost.com/technology/2023/03/18/gpt4-review/>. ↩
  3. “Artificial intelligence (AI) market size worldwide in 2021 with a forecast until 2030” (3 March 2023), online: Statista <https://www.statista.com/statistics/1365145/artificial-intelligence-market-size/#:~:text=The%20market%20for%20artificial%20intelligence,nearly%20two%20trillion%20U.S.%20dollars.>. ↩
  4. EC, Proposal for a regulation of the European Parliament and of the council laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts, 2021/01-6 (COD) at art 6, Annex III [EU AI Act]. ↩
  5. Ibid at art 16. ↩
  6. Ibid at art 21. ↩
  7. Ibid at art 29. ↩
  8. Ibid at art 14. ↩
  9. EC, Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation), arts 7, 17 [GDPR]. ↩
  10. Ibid at art 22. ↩
  11. Bill C-27, An Act to enact the Consumer Privacy Protection Act, the Personal Information and Data Protection Tribunal Act and the Artificial Intelligence and Data Act and to make consequential and related amendments to other Acts, 1st Sess, 44th Parliament, 2022 (second reading completed 24 April 2023) [Bill C-27]. ↩
  12. Ibid at cl 4. ↩
  13. Ibid at cl 5. ↩
  14. Ibid at cl 39. ↩
  15. Ordonnance n° 2021-443 du 14 avril 2021 relative au régime de responsabilité pénale applicable en cas de circulation d'un véhicule à délégation de conduite et à ses conditions d'utilisation, 15 April 2021, 36. ↩
  16. The Act defines it as operating in a mode in which it is not being controlled and does not need to be monitored by an individual. See Automated and Electric Vehicles Act (UK), 2018, s 8 [UK EVA]. ↩
  17. Ibid at s 2. ↩
  18. Ibid at s 5. ↩
  19. Road Traffic Act (Germany) at s 7. ↩
  20. Patrick Ayad, Katharina Göpferich & Susanne Schuster, “Germany takes the lead with a new law on autonomous driving and update” (10 August 2021), online: JD Supra <https://www.jdsupra.com/legalnews/germany-takes-the-lead-with-a-new-law-7746782/>. ↩
  21. EC, Draft Report with recommendations to the Commission on Civil Law Rules on Robotics, (2015/2103(INL)) [EU CLA]. ↩
  22. Ibid at recommendation 31(f). ↩
  23. Ibid at recommendation 26. ↩
  24. EC, European Civil Law Rules in Robotics; Study for the JURI Committee [EU CLRR]. ↩
  25. Ibid. ↩
  26. Ibid. ↩
  27. Ibid. ↩
  28. See eg Nick Bostrom, Superintelligence Paths, Dangers, Strategies, (Oxford: Oxford University Press, 2014). ↩
  29. Manuel Alfonseca et al, “Superintelligence Cannot be Contained: Lessons from Computability Theory” (2021) 70 J of Artificial Intelligence Research. ↩
  30. Jacob Turner, Robot Rules: Regulating Artificial Intelligence, (Switzerland: Springer Nature Switzerland, 2019) at 29–30 [Turner]. ↩
  31. EU AI Act, supra note 4 at art 14. ↩
  32. Ibid. ↩
  33. Ibid at art 14.1. ↩
  34. Ibid at art 14.4(d)(e). ↩
  35. “What is GDPR, the EU’s new data protection law?”, online: GDPR.EU <https://gdpr.eu/what-is-gdpr/>. ↩
  36. EU AI Act, supra note 4 at art 14.4(a). ↩
  37. See Turner, supra note 30 at 29–30. ↩
  38. Joel González Castillo, “Products Liability In Europe And The United States” (2019) 39:2 Revista Chilena de Derecho 277 at 286; Peter E Herzog, “Recent Developments in Products Liability in the United States” (1990) 38:3 The American J of Comparative L 539. ↩
  39. David S Morritt & Sonia L Bjorkquist, “Product Liability in Canada: Principles and Practice North of the Border” (2000) 27:1 Wm Mitchell L Rev 177 at 178–79. ↩
  40. S M Waddams, “Canadian Law of Products Liability” (1999) 34:1 Tex Int'l L J 119 [Waddams]. ↩
  41. Ibid. ↩
  42. Ibid. ↩
  43. See eg Sale of Goods Act, RSO 1990, c S.1, ss 13, 51; Consumer Protection Act, 2002, SO 2002, c 30, Sched A, ss 6–13. ↩
  44. Lambert v Lastoplex Chemicals, [1972] SCR 569, 1971 CanLII 27 (SCC). ↩
  45. Waddams, supra note 40 at 119. ↩
  46. Donoghue v Stevenson, UKHL 100, SC (HL) 31, AC 562. ↩
  47. Cooper v Hobart, 2001 SCC 79 [Cooper]. ↩
  48. Anns v Merton LBC, [1977] UKHL 4, [1978] AC 728. ↩
  49. Cooper, supra note 47 at paras 21–39. ↩
  50. Clements v Clements, 2012 SCC 32. ↩
  51. See eg Bernhard A Koch, “Product Liability 2.0 – Mere Update or New Version?” in Sebastian Lohsse, Reiner Schulze & Dirk Staudenmayer, Liability for Artificial Intelligence and Internet of Things, (Bloomsbury Publishing, 2019) 99 at 115. ↩
  52. See Shawn Bayern et al, “Company Law And Autonomous Systems: A Blueprint For Lawyers, Entrepreneurs And Regulators” (2017) 9:2 Hastings Science and Technology LJ 135 [CLAS]; see also Francesca Lagioia & Giovanni Sartor, “AI Systems Under Criminal Law: a Legal Analysis and a Regulatory Perspective” (2019) 33 433 at 460 [Lagioia]. ↩
  53. See eg Ron Harris, “A New Understanding Of History Of Limited Liability: An Invitation For Theoretical Reframing” (2020) 16 J of Institutional Economics 643 at 644 [Harris]; Timur Kuran, The Long Divergence, (Princeton University Press, 2014) at 98 [Kuran]; Ron Harris, Industrializing English Law: Entrepreneurship and Business Organization, 1720–1844, (Cambridge: Cambridge University Press, 2000) at 39 [Industrializing English Law]. ↩
  54. Kuran, supra note 53 at 100; Phillip Lipton, “The Utilisation of Evolutionary Concepts in Legal History: Company Law as a Case Study” (2020) 46:1 Monash U L Rev 58 at 77 [Lipton]; Naomi R Lamoreaux & William J Novak, Corporations and American Democracy, (Harvard University Press, 2017) at 6 [Lamoreaux]. ↩
  55. Lipton, supra note 54 at 77; Christopher May, A Research Agenda for Corporations, (Edward Elgar Publishing, 2020) at 20 [May]; Lamoreaux, supra note 54 at 6; Kuran, supra note 53 at 102. ↩
  56. Industrializing English Law, supra note 53 at 40. ↩
  57. May, supra note 55 at 21. ↩
  58. Harris, supra note 53 at 645; Lipton, supra note 54 at 78. ↩
  59. Harris, supra note 53 at 646. ↩
  60. Ibid. ↩
  61. Ibid at 649. ↩
  62. Ibid. ↩
  63. Ibid at 647. ↩
  64. Ibid at 651. ↩
  65. Ibid at 651; Paddy Ireland, “Capitalism Without The Capitalist: The Joint Stock Company Share And The Emergence Of The Modern Doctrine Of Separate Corporate Personality” (1996) 17:1 The J of Legal History 41 at 43 [Ireland]. ↩
  66. Kuran, supra note 53 at 139. ↩
  67. May, supra note 55 at 22. ↩
  68. Kuran, supra note 53 at 121. ↩
  69. Harris, supra note 53 at 652; Lamoreaux, supra note 54 at 12. ↩
  70. Lamoreaux, supra note 54 at 12. ↩
  71. Joint Stock Companies Act 1856 (UK), s 3 [Joint Stock Companies Act]; Companies Act 1862 (UK), s 6 [Companies Act]. ↩
  72. Harris, supra note 53 at 655–58. ↩
  73. Companies Act, supra note 71, s 3(2). ↩
  74. Canada Business Corporations Act, RSC 1985, c C-44, s 45; New York Business Corporation Law s 628. ↩
  75. Citizens United v Federal Election Commission, 558 US 310 (2010). ↩
  76. Harris, supra note 53 at 659. ↩
  77. Kuran, supra note 53 at 62. ↩
  78. Ibid at 98. ↩
  79. Ibid at 140. ↩
  80. Harris, supra note 53 at 659. ↩
  81. William P Hackney & Tracey G Benson, “Shareholder Liability for Inadequate Capital” (1982) 43:4 U Pitt L Rev 837 at 840 [Hackney]. ↩
  82. Robert B Thompson, “Piercing the Corporate Veil: An Empirical Study” (1990–1991) 76:5 Cornell L Rev 1036 at 1040 [Thompson]. ↩
  83. Hackney, supra note 81 at 840. ↩
  84. Organization for Economic Co-operation and Development. ↩
  85. James Manyika et al, “A new look at how corporations impact the economy and households” (2021) McKinsey Global Institute Discussion Paper. ↩
  86. Max Radin, “The Endless Problem Of Corporate Personality” (1932) 32:4 Columbia L Rev 643 at 654. ↩
  87. Ibid. ↩
  88. Shawn Bayern, “The Implications Of Modern Business-Entity Law For The Regulation Of Autonomous Systems” (2015) 93 Stanford Technology L Rev 93. ↩
  89. Ibid at 101. ↩
  90. Ibid at 103. ↩
  91. Delaware General Corporation Law s 141(b). ↩
  92. New York Business Corporation Law s 701. ↩
  93. Luciano Floridi & J W Sanders, “On the Morality of Artificial Agents” in Wendell Wallach & Peter Asaro, eds, Machine Ethics and Robot Ethics, (Routledge, 2017) 317 at 325 [Floridi]. ↩
  94. Xavier Oberson, Taxing Robots: Helping the Economy to Adapt to the Use of Artificial Intelligence, (Edward Elgar, 2019) [Oberson]; see also Vikram Chand, Svetislav Kostić & Ariene Reis, “Taxing Artificial Intelligence and Robots: Critical Assessment of Potential Policy Solutions and Recommendation for Alternative Approaches – Sovereign Measure: Education Taxes/Global Measure: Global Education Tax or Planetary Tax” (2020) World Tax J 711 at 730 [Chand]. ↩
  95. Oberson, supra note 94 at 114. ↩
  96. Ibid at 120. ↩
  97. Ibid at 124. ↩
  98. Ibid at 140; Ryan Abbott, The Reasonable Robot: Artificial Intelligence and the Law, (Cambridge University Press, 2020) at 36–49. ↩
  99. Oberson, supra note 94 at 21. ↩
  100. Oberson, supra note 94 at 133. ↩
  101. Ibid. ↩
  102. Halsbury’s Laws of Canada (online), Torts, “Standard of Care and Its Breach 2. Reasonableness (1) Reasonable Person (a) General” at HTO-57 “Definition of ‘reasonable person.’” ↩
  103. Turner, supra note 30 at 199. ↩
  104. EU AI Act, supra note 4, arts 51, 52. ↩
  105. GDPR, supra note 9, art 22. ↩
  106. EU AI Act, supra note 4, art 12. ↩
  107. Road Traffic Act (Germany), s 1g. ↩
  108. See eg EU AI Act, supra note 4, art 12; Road Traffic Act (Germany), s 1g. ↩
  109. Ibid. ↩
  110. GDPR, supra note 9, art 1(f). ↩
  111. See eg Stephen M Bainbridge, “Abolishing Veil Piercing” (2001) 26:3 J of Corporation L 479; Thompson, supra note 82. ↩
  112. José Maria Lezcano Navarro, Piercing the Corporate Veil in Latin American Jurisprudence, (New York: Routledge, 2016) at 48 [Navarro]. ↩
  113. Professor Jeffrey McIntosh, Disregarding Corporate Personality, Course Materials; PowerPoint Presentation (Faculty of Law, University of Toronto, 2021). ↩
  114. Lagioia, supra note 52 at 457. ↩
  115. See New York Business Corporation Law, s 202; Delaware Business Corporation Law, s 122. ↩
  116. Supra note 75. ↩