When Artificial Intelligence Systems Produce Inventions: An Alternative Model for Patent Law at the 3A Era

Currently, robots, Artificial Intelligence, and machine learning systems (hereinafter referred to collectively as AI or AI systems) can create inventions, which, had they been created by humans, would be eligible for patent protection. This Article addresses the patentability of these inventions created by AI systems. We argue that traditional patent law has become outdated, inapplicable, and irrelevant with respect to inventions created by AI systems. We call on policy makers to rethink current patent law governing AI systems and replace it with tools more applicable to the new (3A) era of advanced, automated, and autonomous AI systems. Our argument is based on three pillars: the features of AI systems, the Multiplayer Model, and the irrelevance of theoretical justifications concerning intellectual property. In order to fully convey the ability of AI systems to create inventions, the Article explains, for one of the first times in the legal literature, what AI systems are, how they work, and what makes them (so) intelligent. This understanding is crucial to any further discourse about AI systems. We identify eight crucial features of AI systems: they (1) are creative; (2) unpredictable; (3) independent and autonomous; (4) rational; (5) evolving; (6) capable of data collection and communication; (7) efficient and accurate; and (8) freely choose among alternative options. We argue that, due to these features, AI systems are capable of independently developing inventions which, had they been created by humans, would be patentable (and able to be registered as patents). The traditional approach to patent law in which policy makers seek to identify the human inventor behind the patent is, therefore, no longer relevant. We are facing a new era of machines “acting” independently, with no human being behind the inventive act itself.

The second pillar of our argument is the Multiplayer Model, which characterizes the long process through which inventions are created by AI systems. The Multiplayer Model, which is also almost absent in the current legal publications, describes the multiple participants and stakeholders, both overlapping and independent, involved in the process, including software programmers, data and feedback suppliers, trainers, system owners and operators, employers, the public, and the government. The model conveys that the efforts of traditional patent law to identify a single inventor of these products and processes are no longer applicable.

The third pillar of our argument is the irrelevancy of theoretical justifications, such as personality and inventiveness/efficiency to inventions created by AI systems. In contrast to other scholars, we argue that traditional patent law is irrelevant and inapplicable to these situations, that these inventions should not be patentable at all, and that other tools can achieve the same ends while promoting innovation and public disclosure. These other, non-patent incentives include commercial tools such as electronic and cyber controls over inventions, first-mover market advantages, and license agreements. This proposal serves a gatekeeping function and is superior to a revision of the non-obviousness standard used by other scholars to afford patent protection to inventions by AI systems. In maintaining the traditional patents system by hunting for a “real” human inventor, policy makers exhibit a misunderstanding of advanced technology and AI systems features. We conclude with a discussion of the implications of our analysis for different legal regimes, such as tort, contracts, and even criminal law.

Introduction

The creation of Artificial Intelligence (AI) will be “either the best, or the worst thing, ever to happen to humanity.”1 In this statement by Stephen Hawking during a lecture at the University of Cambridge, he reflects the worries of an unknown future controlled by advanced technology in general and specifically by AI. Elon Musk, the CEO of SpaceX, also warned against the threat of AI, stating, “we are summoning the demon.”2 This Article approaches AI from a different perspective. The AI industry has rapidly become part of our everyday lives and is expected to grow into an estimated seventy-billion-dollar industry by 2020.3 We should face this new realm not fearfully or judgmentally, but with awareness that current laws need reevaluation and new solutions,4 not a continuation of inapplicable frameworks.5

In this Article, we focus on AI and patent law from a new point of view.6 Here, we address the question of whether inventions created by AI systems should be patentable at all, and, if not, what mechanisms can be used instead.

We already live in an era of self-driving cars, autonomous weapons, drug synthesis, disease identifications, medical symptom analysis, and investment advisory tools.7 Language translation, face recognition, answering machines, automated submission of legal appeals and opinions (i.e., automated lawyers), and automated therapists, among other services, have all become part of our everyday routines.8 All of these depend on AI systems to various extents.

This is also true of intellectual property products and processes. AI advanced systems are replacing and sometimes improving human activity and functionality in creating intellectual property products. Though it sounds like science fiction, AI systems already write newspaper articles, create and author stories, produce paintings, create musical compositions, write software, generate other AI systems, and even design inventions.9 AI systems create a wide range of innovative, new, and non-obvious products and services, such as medical devices, drug synthesizers, weapons, kitchen appliances, and machines, and will soon produce many others that, had they been generated by humans, might be patentable inventions under current patent law.10

Massive investments, mostly from big private sector firms such as IBM, Google, Amazon, and Facebook target AI development.11 AI advanced systems are becoming capable of creating unpredictable, innovative outcomes independently, rather than merely by following digital orders. Such inventions by AI systems are the focus of this Article. Torts liability and even criminal law, which have been at the forefront of legal discourse, do not necessarily address these new challenges of creative AI related to intellectual property law. A rethinking of traditional legal tools is required.12

The legal system, including intellectual property law, needs to change significantly to keep pace with recent developments in these technologies.13 Previous literature expressing concerns about AI has focused mainly on workplaces, misuse of AI systems, and general liability issues. There has been little discussion of regulating AI,14 and even less of intellectual property protection for inventions created and developed by AI systems. Musk’s comment about the need for new regulations gives rise to an interesting, more general question: how to strike the balance between regulation and the intellectual property regime in general, specifically between automated and creative AI systems and incentives to innovate.15

In 1998, John Koza, one of the pioneers of AI genetic algorithms, developed an algorithm to create simple circuit designs.16 As he continued to expand his work, Koza ultimately built a cluster of 1000 personal computers that generated seventy-six “human competitive” designs.17 Surprisingly, Koza obtained patents on the automated invention system as well as on the inventions generated by the AI itself.18 While we agree with the granting of the patent in the first case, we challenge the granting of patents to inventions by AI systems.

In this Article, we inquire as to who owns the rights to patentable products and processes produced by AI systems. Traditional patent mechanisms seek to identify conclusively the owner of a patent and who must be within the scope of patent laws (e.g., a corporation, operator, or inventor).19 We analyze AI systems as autonomous, creative, unpredictable, rational, and evolving systems, and argue that these characteristics make justifications such as personality theories and incentive/efficiency arguments irrelevant. We conclude that one cannot conclusively determine an owner for these rights within the scope of patent law. Therefore, the rights fall outside the scope of traditional patent law.

So far, the few proposals suggested by other scholars all attempt to exercise and implement current laws with regard to inventions by AI by focusing on the definition of eligible patent matter, and particularly on non-obviousness standards.20 Some voices call for considering machines as inventors and granting them patent rights.21 Unlike other scholars, we argue here for abolishing patent protection of inventions by AI altogether.22 Further, we suggest promoting innovations and public disclosure of inventions by AI through alternative tools such as, for example, first-mover advantages, social recognition of AIs, and alternative technologies that prevent infringement of rights, rather than relying on traditional intellectual property law to accomplish these goals.23

Part I considers the intelligence of AI systems as a first step in determining who owns the patent rights to inventions created by such systems. We identify and describe eight features of AI systems—including autonomy, creativity, and unpredictability—that establish the intelligence of these systems. Part II describes patent law’s refusal to recognize nonhumans as inventors and the issues that have resulted from it. Part III introduces the Multiplayer Model, a characteristic of AI systems. Part IV addresses different aspects of the interaction between AI systems and intellectual property regimes, and of the AI Multiplayer Model, specifically through a theoretical legal and economic analysis, a Lockean labor analysis, and a personality analysis. Part V discusses the legal hurdles, within intellectual property laws, that need to be overcome in order to alter the process by which the owner of inventions by AI systems is identified, particularly in intellectual property law. It addresses the theoretical justifications for intellectual property with an emphasis on law-and-economics theory and describes current U.S. intellectual property law in the context of AI systems. Finally, Part VI proposes our new alternative Model and Part VII briefly addresses international tools for implementing them.

I. What Is So Intelligent About AI Systems?

In order to convey the challenges of patentable inventions created by AI, we begin with an explanation of AI systems and how they produce innovative and unexpected products and processes which, had they been developed by humans, might qualify as patentable inventions.

Defining AI systems is not an easy task. There are numerous definitions of AI and many types of AI systems.24 John McCarthy, who coined the term “artificial intelligence,” did not provide an independent definition for it, while Russel and Norvig suggested almost ten different ones.25

Definitions vary as different aspects of AI systems are emphasized.26 An AI system can be defined, based on its features, as one capable of performing tasks that normally require human intelligence, such as recognition, decision-making, creativity, learning, evolving, and communicating.27 AI can also be described as an instrument that makes existing solutions more efficient by using all of the data within reach of the AI system.28 Definitions also differ in various contexts (i.e., medical treatments or chess strategies). For our purposes, we focus on definitions most relevant to the patent system and adopt Scherer’s somewhat evasive definition: “machines that are capable of performing tasks that, if performed by a human, would be said to require intelligence.”29

We claim that there are eight important features of AI systems that create new challenges to intellectual property law.30 Some AI systems include some or all of these features, whereas others are more similar to computer software systems. Those with all eight of these features, including robots operating alongside AI systems, are being used not only to solve complex problems across an ever-increasing number of industries—smart vehicles, consumer devices, health, and pharmaceutical technologies—but also to create and produce products and processes themselves.31 These eight features are related to each other and sometime overlap. However, each one focuses on a different characteristic. We state that they are the main stones building the 3A era of advanced, automated, and autonomous AI systems.

Creativity. AI systems create new products and processes, and significantly improve existing ones. They are capable of processing and reproducing other products, processes, and available data in order to create new outcomes.32 For example, AI systems can draw, create designs, and even produce inventions such as drugs and technical devices.33 This feature is crucial in considering the intellectual property realm and, in particular, patentable inventions.

Unpredictable Results. AI systems are based on algorithms capable of incorporating random mutations that result in unpredictable routes to the optimal solution, and hence to unpredictable solutions (from software programmers’ points of view).34 AI systems are goal-driven; they process data and take action in order to generate products, data, and processes that cannot be predicted by programmers, operators, or any other entities involved.35 For example, an AI system that creates paintings is generating an unpredictable product, rather than simply copying an existing work. AI systems that work on developing new and innovative antibacterial drugs can process data from a large volume of microorganisms (i.e., bacteria), “break” the data into tiny (sometime nano) components and find similarities and patterns that the human involved has not observed and cannot identify, resulting in new and unexpected structural information for drug development.36

Independent, Autonomous Operation (t-autonomy). This feature is one of the most important to understand in order to grasp AI systems in general and their departure from the framework of current patent law. Although the definition of autonomous AI system might vary according to the specific industry and from one system to another, we can identify some common characteristics.37 Degrees of independence and creativity are both relevant. We can say that a device is independent and therefore autonomous to the extent that it accomplishes a high-level task on its own, without external (human) intervention.38 Human intervention can occur in many phases of the process—observation, orientation, deciding, and acting—resulting in different levels of independence.39 An unmanned aircraft and automated pilot can operate independently when needed (e.g., during a communication breakdown).40 Autonomy level is influenced by interaction with other features.

A second dimension of autonomy concerns cognitive ability. The larger the cognitive task assigned to the AI system, the more it can be considered autonomous. The autonomy of an AI is evident in data searches,41 where the algorithms of the AI system may work independently without human intervention beyond defining goals.42 We argue that the main idea behind these AI systems is their capability to identify similarities and patterns when processing data that even programmers and operators themselves were not aware of and many times do not completely understand. Despite this lack of human intervention, however, Koza has patented inventions generated by such advanced AI genetic algorithms.43 Unlike industrial robots that assemble and replicate circuits designed by humans, the AI system in Koza’s patent actually designs new circuits.44 In this way, the AI systems replace the engineer, autonomously choosing, ordering, and assigning strengths to various circuit components to achieve predetermined performance parameters.45

Rational Intelligence. An “intelligent machine” means a rational system that perceives data from the outside world and decides which activities to engage in or avoid in order to maximize its probability of success in achieving a certain goal.46 These AI systems mimic human perception and cognitive functions such as learning and problem solving, thereby imitating intelligent human behavior.47

Evolving. AI systems continue to evolve and change according to new data. This feature also contributes to the unpredictability mentioned above. AI systems may produce results that differ from the initial plan of the programmers or operators of the system.48 For example, an AI system that synthesizes a drug based on bacterial structures will produce new outcomes as new bacteria are processed.

Capable of Learning, Collecting, Accessing, and Communicating with Outside Data. A significant feature of AI systems is their ability to actively “search” for data in the “outside” world. Based on the data gathered, an AI system can continue the process by receiving feedback and then improving the results.49 Siri of Apple and Google Translate of Google serve as simple examples of these features. However, a new generation of autonomous, network-centric applications can collect data incessantly from different sources.50 Driverless cars are collecting and processing data from the outside world (e.g., other cars, obstacles, and traffic signs) and autonomous weapons are processing data in order to identify targets.51

Efficiency and Accuracy. AI systems can process vast volumes of data accurately, efficiently, and rapidly, well beyond the capacity of the human brain.52 Although less sophisticated computer software possess this feature, it also exists in complicated AI systems.

“Free Choice” Goal Oriented. This feature focuses on the capability of the AI system to choose between alternatives in order to achieve the best outcome.53 Automated weapons decide, for example, which targets should be attacked according to the surrounding data.54 Specific AI systems implemented in driverless cars process data in order to choose from different alternatives and decide on routes, speed, and accident avoidance.55

All of these eight features characterize, to a certain degree, different AI applications. We can summarize the main ones as creating the 3A era (of advanced, automated, and autonomous AI systems). These features allow AI systems to create and invent products and processes which would be worthy of patent protection had they been developed by humans. Human ownership over these products of AI is, therefore, questionable.56 Once we understand the features of AI systems and that AI systems create outcomes independently, we realize that humans alone are not entitled to the rights to these products. Thus, traditional patent law is not applicable in the 3A era.

This has become more obvious as technology advances and as AI systems, when embedded with the features listed above, become increasingly capable of mimicking the functions that we consider to symbolize the human mind, creating new products and processes. AI systems have become valuable for solving specific problems and now promise to improve specific human skills—not only accuracy, velocity, and capacity to process vast amounts of data but also creativity, autonomy, novelty, and other features that establish patentable innovations. Moreover, facing the 3A era, AI systems will soon be able to develop inventions without significant guidance or instructions and even create, complete, and submit unlimited number of patent applications themselves.57

Responsibility for such outcomes is usually attributed to the human or entity behind the process of invention. While tort and liability inquiries are at the forefront of scholarly discourse on these processes, we call for discussion of the implications of these technologies for intellectual property in general and, more specifically, for patent law. It still merits consideration, though, whether AI systems own the products they generate. Unlike other scholars, we think they cannot.58

The next Part begins by addressing this question.

II. Legal Implications of Artificial Intelligence Systems

In this Part, we will discuss why, in our view, current patent law is inadequate to regulate technological developments in automated AI systems producing inventions. We start the discussion by accepting the fundamental assumption that AI systems can create inventions that traditionally were created only by human beings, as this is already part of our reality.

A. Intellectual Property Laws Face New Challenges

U.S. patent law (35 U.S. Code § 101) explains who may obtain a patent and what constitutes a patentable invention: “Whoever invents or discovers any new and useful process, machine, manufacture or composition of matters, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.”59 Is an AI a “who”? Can an AI system be entitled to patent protection?

According to a well-known phrase of patent law, anything under the sun, which is made by man, qualifies as patentable subject matter. This statement expresses one of the main challenges of patenting creations by sophisticated AI systems.60 U.S. patent laws take only human inventors into account, defining “inventor” as “the individual or, if a joint invention, the individuals collectively who invented . . . the subject matter of the invention.”61 “Joint inventor” and “co-inventor” “mean any one of the individuals who invented or discovered the subject matter of a joint invention.”62 The law does not consider the possibility of a nonhuman inventor.

Precedents in related intellectual property issues, addressing copyright laws, have not considered nonhumans such as machines and animals to be creators within copyright law. Naruto, a six-year-old crested macaque, for instance, was deemed unable to own the copyright to photos he took of himself.63

Given the treatment of nonhumans within copyright law, then, inventions created by AI systems, although belonging to another branch of intellectual property, may not be included within the scope of patent law. Their status remains unclear.

The human-centric, traditional approach to AI inventions focuses on identifying the human behind the invention in order to assign to that person patent rights and protection.64 In some cases of AI producing intellectual property, however, we argue that this traditional approach might be misleading and wrong. AI systems can produce a surprisingly large number of inventions, write and submit numerous patent applications, and even evaluate (or monitor) the risk of patent claims.65

There are few prior conditions that statutory subject matter must meet before receiving patent protection: novelty, non-obviousness, usefulness, and written description of the subject matter to be included in the application.66 As long as these conditions are fulfilled, an AI system or anyone on its behalf may obtain a patent.67 In other words, a patent is granted when the applicant demonstrates that the legislative eligible subject matter is new, useful, non-obvious, and might contribute to the public welfare.68

Because AI systems, thanks to their fundamental features, can easily meet and fulfill all of these conditions, producing new, non-obvious, and useful inventions, AI systems might have been entitled to patent rights to their inventions had the law not been intended for human inventors alone.

The Multiplayer Model defies (and sometimes undermines) the long-established, but outdated (from our perspective), traditional paradigm.

III. The Multiplayer Model

The question of ownership and AI-produced inventions can also be considered using the Multiplayer Model—though, we argue, the contributions of the many players tussling for rights based on their indirect and insignificant involvement in AI systems that produce inventions do not meet the threshold of inventorship.

A. The Multiple Players of the Model

The new AI realm consists of multiple stakeholders with varying interests, some of whom are at odds with each other, making traditional patent protection less applicable.69 We claim that there are at least ten entities among the many possible stakeholders who are only partially, indirectly, insignificantly, or temporarily involved in the invention process. The categories of stakeholders can overlap (e.g., the programmer can be the owner and the trainer) or remain separate and distinct.

The Software Programmers. An AI system is first developed as a software program. The programmer of the software itself or any entity on his or her behalf (e.g., an employer) undoubtedly owns the copyright to the software. Rather than focusing on the copyright in the software itself, however, we choose to examine the patent ownership of the new, creative, unpredictable, and autonomous intellectual property produced by the AI system. The programmers create the AI software program algorithms but do not necessarily target the final goal of the AI system. AI systems based on identifying similarities and recognizing patterns—such as in stroke prediction, facial recognition, or drug synthesis—can operate with a variety of data, possibly with some modifications (which might be programmed by others). The software program itself or one of its human programmers may generate the specific goal-seeking steps or specific use.70

The Data Suppliers. Usually (but not always), the next phase in the invention process is “exposing” the AI system to data that the system exploits to “learn” how to function and to achieve its goal efficiently.71 With facial recognition, for example, the data supplier can provide the system with millions of pictures of faces in different forms and facing in different directions.72 The system may use existing data as well (such as that from social networks), however, without the programmers actively providing the data or even knowing which websites the AI system will pull the pictures from and when. The Google translate program, for example, uses the users as data suppliers. The data may be open or “closed,” public domain or owned by a different entity.73

The Trainers/Feedback Suppliers. The trainers check the AI system’s results and correct them when necessary, playing an important role in establishing the system’s capacity.74

The Owners of the AI Systems. The AI system’s owner can be the first or successive owners, firms, or individuals.

The Operators of the Systems. This can be an entity that licensed the AI system from the owner or those working with the owner as service providers.

The New Employers of Other Players. These were previously mentioned. They can be stakeholders if they change employment.

The Public. If neither the AI system nor any of the stakeholders are entitled to the rights, the invention might be owned by the public.

The Government. When no one else is entitled to property rights, including intellectual property rights, the government possesses them by default.75 In cases where government investment is made, the government is entitled to all or part of the rights.76

The Investor. This person sponsored the development of the AI system or any other player.

The AI System. The Artificial Intelligence itself is an autonomous entity. Other scholars have already argued that, based on the paradigm of corporate ownership, which enables intangible nonhuman legal entities to retain responsibility for legal dispositions and to hold rights including IP rights, AI systems can be entitled to patent rights as well.77

This Multiplayer Model raises many questions when applied to AI systems. Any of the ten players listed above (or more) can claim ownership over the invention, thereby raising the question of how to identify the actual inventor and the player entitled to the patent rights. Assuming that the programmer of the software might have the copyright to the software does not mean that the owner thereby owns the patent rights to the new, unpredictable, and evolving inventions created by the autonomous AI system.78 Should the rights holder be the person who developed the AI system itself, the person who “discovers” or holds the invention, or the person who selected and provided the training data during the first stage (before the AI develops inventions)? In these situations, almost all employees or contractors might have the contractual obligation to assign the invention to the company,79 but we argue that the initial legal question of who literally invented the patentable subject matter remains unsolved. We further argue that, if none of these players qualifies as an inventor according to the current legal definition, does any other entity or any company hold the rights to a patent for an AI invention? Can all of them be considered co-inventors?

To qualify as an inventor or at least a joint inventor, one must contribute significantly to the conception of the claimed invention. The inventor contributes directly and significantly to the inventive process, which originates the invention according to the eligible patent matter.80 Might the programmer and operator also have a claim if the AI they developed or operated creates an invention as a result of their activities? Does the answer change when inventions by AI are created as a result of scanning data in cyberspace, such as by downloading data from social media?

What happens when the AI system is an autonomous robot that is able to leave a confined setting and maneuver in a public space, where it gains information that contributes to the invention? What if an otherwise immobile AI system “crawls” the internet for training data from multiple sources?

The players who provide a significant part of the training data might have a claim to inventorship if they can prove that the training phase sufficiently contributed to the end result of a patentable invention by the AI system. The owner might also have a claim as the entity that initially launched the AI with particular guidelines or training. If the initial training did not produce inventions and the training materials provided by the trainer turned out to be relatively insignificant, do the trainers still have a claim to ownership? If the AI system learns autonomously from being in a public setting in which no individual actions significantly contribute to the creation of the inventions, who is the inventor? In such situations, could the rights belong to the public? If a firm brings the AI back onsite every evening for a data dump and analysis and makes adjustments based thereon, would it be entitled to a claim of inventorship?81 Is setting the end goals and parameters for the AI system rather than obtaining the resulting data significant enough to establish patent rights on creations made by AI?82

Indeed, those seeking enjoyment of the innovations made by AI can be seen as having conflicting interests with those seeking profits or protection for the invention. We assume here that all players pursue their goals rationally. Admittedly, this assumption could fail—“cognitive biases,” for example, could lead players to “systematically overvalue their assets and disparage the claims of their opponents . . . .”83

The role of operators of AI might be minor if they do not contribute any data. As the technology of AI develops and more players become involved, identifying the inventive task becomes more complicated. No entity makes a significant contribution to the inventive process by AI systems. The programmers do create the AI system itself and therefore, gain the rights to the copyright.84 However, granting the programmers the copyright on the AI system does not necessarily entail granting them the rights to the products and processes developed by AI systems as well. The next Section will focus on this issue.

B. Ownership of AI Software Versus Ownership of AI System Inventions

Patent protection is not the only way to promote innovation. Software can be protected by another regime of intellectual property law: copyright law. Software innovation, including AI software itself, has rapidly developed and flourished without the aid of patent laws.85 The Constitution delegates to Congress the power to grant exclusive rights to “Authors and Inventors” in their respective “Writings and Discoveries.”86 Therefore, the two constitutional inquiries about intellectual property concern: (1) who can be counted as an Author or an Inventor, and (2) what can be counted as Writings and Discoveries?

The AI system itself triggers copyright protection because software code is one of the “Writings” protected by the Constitution.87 However, the discourse about ownership with regard to AI software is totally distinct from the question of ownership in the inventions made by AI systems. Under certain circumstances, inventions by AI might deserve copyright protection. For example, AI-generated art might be regarded as proper “works of authorship” pursuant to § 102 of the Copyright Act by virtue of AI’s sufficient connection to human creativity.88

Nevertheless, patents offer much broader and more easily enforced intellectual property rights than copyrights do, for three main reasons.89 First, the copyright owner of the AI system itself cannot exclude others who independently invent substantially similar inventions to those produced by his or her own system. To the contrary, all who independently develop inventions may be subject to patent rights and therefore, be excluded.90 Second, under copyright’s “fair use” doctrine, others can reproduce copyrighted inventions for “criticism, comment, news reporting, teaching . . . , scholarship, [and/]or research . . . .”91 Third, patents are being registered and, thus, are easier to enforce, rather than copyrightable materials, of which the obligation to register them is subject to submitting claims.92

On the one hand, we do not challenge the eligibility of the programmer to be entitled to ownership according to copyright laws governing the software she or he develops. This is also true with regard to AI software. On the other hand, this entitlement does not automatically result in ownership over the products and processes created by AI systems. We argue that the stakeholders over a piano, a brush, a camera, a computer, a printer do not hold the rights over the rhythm, the painting, the photo, or the story created by those instruments. This conclusion brings us back to our initial question: who owns the inventions of AI systems?

The next Part will address this question from a theoretical perspective in general and as related to the Multiplayer Model, typical of AI systems, in particular.

IV. Theoretical Approaches

The U.S. Constitution grants Congress the power to “promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.”93 Discourse concerning the theoretical justifications for intellectual property tends to focus on three main substantive theories: (1) law-and-economics theory, a utilitarian approach that examines intellectual property rules according to their cumulative efficiency and ability to promote total welfare; (2) personality theory, which focuses on the personality of the creators and inventors; and (3) Lockean labor theory, which justifies the property interest as the fruits of the creator’s labor.94 Today, U.S. intellectual property law is based primarily on utilitarianism95 and secondarily on John Locke’s theory of labor.96

A. Law and Economics

1. Law, Economics, and Intellectual Property Law

The law-and-economics approach focuses on promoting the production (and distribution) of scientific and cultural goods via utilitarian laws designed to promote economic efficiency. In general, this approach aims to maximize the total social welfare of the public from an economic perspective.97 The law-and-economics approach attempts to solve the problem of intellectual property assets as intangible market products, which “free rider users,” who enjoy the product without paying properly or being given permission to do so, can easily copy without rewarding the authors or inventors. Copying by free riders costs less than the investment necessary to create and develop products; the lack of consequences for free copying therefore threatens to deter authors and inventors from enriching our world by generating intellectual property products.98 Consequently, one of the purposes of intellectual property laws is to incentivize creators and inventors with exclusive rights to intellectual property products, preventing others from using their products without permission and without paying for them.99 According to Richard Posner, the public, authors, and inventors have (theoretically) “signed” a social contract in which the public (society) gives authors and inventors exclusive rights to their works for a limited duration, which provides enough incentive for them to create and develop.100 However, once the exclusivity period expires, the rights are transferred to the public and become part of the public domain.101 The market price of the product reflects its social value.102

Although the law-and-economics approach to intellectual property is dominant in the United States, many scholars have found its prevailing influence troublesome.103 Professor Amy Kapczynski, for example, not only refers to the presumed efficiency of intellectual property law as alleged, but also considers intellectual property using only the utilitarian-efficiency approach. Kapczynski claims that the price of intellectual property products gives us a decentralized way to link social welfare to the production of information. She further claims that, by looking beyond economic justifications of intellectual property, we discover different institutional approaches to scientific and cultural production that are no less efficient.104

Under classic law-and-economics theory, “the ultimate goal of the patent system is to bring new designs and technologies into the public domain through disclosure.”105 In other words, patent protection serves two functions: to incentivize innovations and to ensure public access to knowledge. Indeed, U.S. patent law seeks a “careful balance between public right and private monopoly to promote certain creative activity . . . .”106 As such, the patent “monopoly” (exclusive rights) is often understood not to secure the inventor’s natural right in his or her discoveries but rather to induce new knowledge.107 Therefore, the exclusive rights granted under U.S. patent law for a limited time are simply a means to an end.108

To effectively incentivize innovations, patent law provides rewards, which should be high enough to promote innovation109 as well as cover the inventor’s research and development (R&D) costs, the probability of failure, and unexpected obstacles.110 Unfortunately, since the trade-off between reward and cost is heavily fact-dependent, a general utilitarian theory provides only limited guidance.111 Today, as new technologies such as AI advance, the R&D cost of some inventions has decreased tremendously. General AI software is available to be used for several purposes.112

Compared with the incentive to innovate, the public disclosure function of patent law is much easier to achieve. Forced public disclosure under U.S. patent law ensures that today’s inventors stand on the shoulders of those who came before them.

With regard to AI systems, the concept of incentive as a human trait is commonly considered to be meaningless. We claim that autonomous machines do not need any incentive—that incentive is relevant only to people and entities until machines, robots, and AI systems start producing; during system maintenance in the productive process; and during distribution and accessibility implementation after the inventions are made (which are processes different from creating intellectual property). The following Sections will discuss several issues regarding the second relevant meaning of incentives.

2. Transactional Costs, Cumulative Innovations, and Outcomes

Classic utilitarian theory works best when transaction costs are low enough to be neglected.113 However, transaction costs, such as, for example, for patent searches, will inevitably increase as the number of AI inventions increases. Companies large enough to absorb increasing transaction costs will have the advantage over smaller companies and may even be able to afford to patent inventions that they do not plan to implement.114 Further, if transaction costs increase too much, the incentives to innovate may not offset them.115 Under these circumstances, innovation will stall.

To adapt to AI’s Multiplayer Model as previously discussed, classic utilitarian theory should include a discussion of cumulative innovations.116 Here, AI offers two distinct benefits to society: its own present value117 and its value as a platform to generate more inventions.118 Under utilitarian theory, inventors would proceed whenever all benefits exceed all costs. An economically efficient patent system must allocate profits and losses, and maximize efficient investment among all the players described above.119

We argue that within the AI Multiplayer Model, the more players involved, the less efficient the process becomes. We argue that with multiple players, the question of ownership usually obstructs rather than facilitates the process; ownership of AI inventions based on an IP theoretical justification, therefore, is likely to be inefficient and obstructive. Incentive is still an important factor in the inventive process. Incentivizing the AI software programmers occurs by granting them copyright protection. Nevertheless, as mentioned previously, granting the programmers copyright does not mean they are entitled to ownership on inventions produced autonomously, by automated advanced AI systems. Trainers, operators, distributors, and sellers—they all need incentives as well. However, they are not the inventors of the inventions (being produced by AI systems), hence, they are not eligible for the patent rights and ownership rights. The only one who does not need incentives to invent (as incentive is understood according to the law-and-economics approach) is the AI system itself while producing inventions. The discourse about incentives also supports the conclusion that patent law has become irrelevant and inapplicable when AI systems produce inventions. We argue that incentives should be based on commercial tools, not on IP theoretical justifications in general or on patent law specifically.

The next Section will focus on the third theoretical justification for IP: the Lockean labor theory.

B. Locke’s Labor Theory

Labor theory is usually credited to John Locke, who wrote in his Second Treatise on Government that “every man has a ‘property’ in his own ‘person.’ This nobody has any right to but himself. The ‘labour’ of his body and the ‘work’ of his own hands, we may say, are properly his.”120 Thus, according to Locke’s labor theory, an inventor has an inherent right to the fruits of his labor as he does to the fruits of his mind and soul.121

There are two limitations, however, on acquiring property rights through labor, and both limitations appear to be fundamentally at odds with patent law.122 The first states that one can acquire property rights only “where there is enough, and as good left in common for others,” meaning that the inventor has not deprived others in the world of necessary goods.123 Current patent law runs afoul of this proviso by granting the original inventor an exclusive right to make, use, and sell his or her invention. Subsequent inventors who independently generate an already patented invention are forbidden by the original patent grant from using or selling it, thereby suffering a significant loss.124 The second proviso demands that one acquire property rights to satisfy one’s needs and no more.125 In reality, many technology companies acquire patents as offensive strategies (acting as so-called patent trolls), not to satisfy their “needs.”126

Although a pure reading of Locke’s labor theory may be inconsistent with patent law in general, especially in terms of granting patent protection to entities (instead of humans),127 the theory can supplement the law-and-economics analysis in evaluating patentable subject matter.128 This Article adopts Locke’s labor theory only to the extent of arguing that inventors should be awarded for the fruits of their labor. Notably, Locke’s labor theory is based on the concern that an inventor be rewarded fairly129—granted exclusive rights for having mixed his labor with the contributions of his mind, soul, and ideas. To do otherwise would inflict harm to others’ claims to the commons, a situation that would run counter to Locke’s mandate that no harm be done to others.130 Therefore, a proper application of Locke’s labor theory to patent law depends largely on defining the suitable “fruits” that could grow from one’s contribution. In terms of intellectual property law, this entails defining the adequate benefit to inventors according to their “donation” and, more specifically, deciding on the size of the rewards that Locke’s labor theory justifies.131

The division of profits among the various players—the AI programmers, trainers, owners, and operators—should reflect their respective contributions to the development of the inventions created by the AI. Here, these inventions would not have been developed without the original programming of the AI in the first place. Therefore, Locke’s labor theory calls for compensating those players, including the programmers, with a portion of the profits from the subsequent inventions created by the AI, in light of the programmers’ and other players’ contributions.132 We argue that, similarly, by exploiting the AI to generate inventions, the operator of the AI should also receive an economic reward. The size of this reward depends on the difficulty and the extent of innovativeness in the setting of the end goals and parameters, which in turn depends on the definition of “operator” and the level of mind and soul (“labor”) contribution to the inventions made by the AI system. The more the players, such as the computer operators, are simply following guidelines, the less they will be entitled to ownership under Locke’s labor theory. The less these players are involved in the inventive process itself, the less their entitlement to the AI systems’ inventions. According to our Multiplayer Model, these players typically do not get too involved in the creative process itself and are more often executing someone else’s ideas or orders; their “contributions” to the inventive process that results in the AI’s products are therefore limited. As in Lockean theory, we conclude that in most cases, they should be considered neither inventors nor co-inventors.

Bearing this in mind, we will explore the third main theory used to justify intellectual property rights: the personality approach.

C. The Personality Approach

Personality theory is based on Hegel’s view that property rights are a means for developing and realizing one’s personality.133 Hegel argues that “an idea belongs to its creator because the idea is a manifestation of the creator’s personality . . . .”134 Consequently, an AI system cannot be entitled to patent rights to its creations and inventions because personality is exclusively attributed to human beings. However, we can examine the personality approach with regard to the Multiplayer Model involved in an AI system’s operation.

Private property rights are justified according to Hegel’s philosophy because acting on external property imposes one’s stamp on the external world, thereby fundamentally contributing to the development and flourishing of individual personhood. Personhood and freedom may be expressed through work with assets.135 Creating and developing intellectual property fosters both the intellectual and emotional components of the human personality. Creators and inventors therefore have a natural right to control the use of their intellectual products, because controlling property is a key component in developing personhood.136

Professor Margaret Radin adds to this theoretical approach by dividing property into fungible and nonfungible assets. Radin considers one’s personhood to be embedded in personal assets that must therefore be protected more vigorously than exchangeable assets, to which one has a weaker connection.137 Drawing on Hegel’s personality approach, Radin explains that “to achieve proper self-development—to be a person—an individual needs some control over resources in the external environment. The necessary assurances of control take the form of property rights.”138

Building upon the insight that every individual is entitled to a minimum amount of property in order to develop his or her personality, Radin recommends a continuum of property protection. At one end of the spectrum, there is “personal property” (e.g., wedding rings, portraits, houses) that constitutes a continuation of one’s entity as a human.139 At the other end of the spectrum, there is “fungible property” that is held for purely instrumental reasons140 (e.g., money, stocks, automobiles in a dealer’s showroom). Due to its unique value to the owner, the loss of personal property cannot be compensated through payment or replacement with another object of similar market value. A fungible object, by contrast, is replaceable.141

Scholars by and large agree that intellectual property rights are closer to the personal property end of Radin’s continuum because an intellectual product reflects the personality of its creator.142 Accordingly, in the arena of copyright law, scholars have used personality theory to call for the strengthening of authors’ rights, including rights of attribution and integrity.143

Despite its application in copyright law, Hegel’s and Radin’s personality theory is less frequently invoked to justify patent law. This is because technological inventions usually embody utilitarian solutions to very specific needs.144 Instead of manifesting the personality of the inventor, patents manifest the inventor’s raw insights in solving a particular problem.145 For example, in inventing the light bulb, Thomas Edison “searched for the filament material that would burn the longest, not a filament that would reflect his personality.”146 From this perspective, personality theories do not justify ownership in inventions in general, nor do they unambiguously justify ownership of inventions made by AI systems.

Nevertheless, one might claim that an inventor still has opportunities to express his or her personality in technological inventions. An inventive process does not deal exclusively with objective facts. In many cases, there are multiple ways to solve a problem. While the problem itself may be impersonal, an inventor chooses to tackle it in a manner that reflects his or her individual personality.147 For example, it may be possible to program an AI in various ways, each one representing a different personal style for accomplishing the task. Moreover, even though a patented invention might not reflect the inventor’s unique aesthetics or emotions, the invention is still a personalized integration of the inventor’s training, education, intellectual skills, and creative spark.148 However, this argument does not change our conclusion regarding ownership of inventions by AI. The personality justification for intellectual property rights is not applicable to the Multiplayer Model, in which players’ roles are more technically oriented than creativity-based.

A patent gives an inventor the “right to exclude others from making, using, offering for sale, or selling the invention” for a limited time.149 We claim that in a Multiplayer Model of AI invention, ownership is questionable. We conclude that granting intellectual property rights to the different stakeholders is not justified under any of the three theories: utilitarian legal and economic theory, labor theory, or personality theory.

The next Part will focus on the normative legal aspect of eligible patent matter in U.S. patent law, whose wordings and goals, we argue, are not designed to handle inventions created by AI. Unlike other scholars who seek a human behind the inventive process or create new legal personalities to whom such ownership rights could be granted, we claim that patent law is simply not applicable to inventions by AI.150

V. The Legal Analysis: Current U.S. Intellectual Property Law and AI

Under current U.S. patent law, an inventor must show that his or her invention is eligible for patent protection because it is useful, novel, non-obvious, and adequately enabled and described.151 Of these factors, subject matter eligibility and non-obviousness are the most flexible.

A. Subject Matter Eligibility

In § 101 of the Patent Act, Congress explicitly defines as eligible for patent protection any “process, machine, [article of] manufacture, or composition of matter . . . .”152 We argue that even though a patent’s subject matter eligibility is based in statutory law, it has gained a distinctly common-law feel over the years. Despite the apparent breadth of § 101, the Supreme Court has carved out three exceptions as unpatentable subject matter—“[t]he laws of nature, physical phenomena, and abstract ideas.”153 These exceptions embody the “basic tools of scientific and technological work” and therefore must remain in the public domain to ensure that patent rights neither “tie up the use of such tools [nor] inhibit future innovation premised upon them.”154

Among the myriad new developments in subject matter eligibility, the machine-or-transformation test is most closely related to AI. In Bilski v. Kappos, the Court imposed the machine-or-transformation test as a threshold requirement for a process to be patent eligible.155 Under the machine-or-transformation test, a process is patent eligible only “if: (1) it is tied to a particular machine or apparatus, or (2) it transforms a particular article into a different state or thing.”156 The Court essentially denied patent eligibility of “mental processes” based on the assumption that mental processes are necessarily abstract and intangible.157 However, this assumption does not apply to AI, whose thought processes are reduced to either physical transformations or the architecture of the machine itself.158 Therefore, current subject matter eligibility doctrine might be well equipped to analyze AI systems. As the Court itself acknowledged, though not intentionally in relation to AI systems, “[§] 101 is a dynamic provision designed to encompass new and unforeseen inventions . . . .”159

In general, the main drawback of limiting patent scope through § 101 is that it could lead to arbitrary boundaries easily circumvented with “magic words” in the claim language.160 Indeed, throughout the 1980s and early 1990s when software itself was unpatentable, many inventors easily circumvented this barrier by claiming hardware “machines” as software inventions.161

B. Non-Obviousness

Another criterion of patentability is non-obviousness, as set forth in § 103 of the Patent Act of 1952.162 The statute mandates that to be patentable, the invention must not be obvious to a person having ordinary skill in the art (PHOSITA) at the time of the invention.163 This non-obviousness requirement has remained largely unchanged since 1952, except for a minor change in how “prior art” is defined.164

The Supreme Court first established the framework for § 103 in 1966 in Graham v. John Deere.165 The Court identified four factors in evaluating non-obviousness: (1) the scope and content of the prior art; (2) the skill level of a PHOSITA; (3) the differences between the claimed invention and the prior art’s teachings; and (4) any objective indicia of non-obviousness, such as commercial success.166 Graham also established “secondary considerations” that suggest that an invention is non-obvious,167 the non-exhaustive list of which includes “commercial success, long felt but unsolved needs, [and] failure of others.”168

The Court of Customs and Patent Appeals, the predecessor of the Federal Court, subsequently attempted to develop a bright-line rule for non-obviousness known as the teaching-suggestion-motivation (TSM) test. The TSM test required that the prior art contain a teaching, suggestion, or motivation.169 In KSR v. Teleflex, however, the Supreme Court denounced the Federal Circuit’s strict application of the TSM test as too rigid and narrow,170 arguing that it overemphasized “the importance of published articles and the explicit content of issued patents.”171 While it did not reject the TSM test completely, the Court noted that this overly narrow application departed from 35 U.S.C. § 103 and the Graham obviousness framework.172

Advances in AI may require redefining “ordinary skill” and the “PHOSITA” assessment. By far the most important development of the PHOSITA standard also came in KSR, with the Supreme Court transforming the PHOSITA requirement from a mere “automaton” to a person with ordinary creativity levels.173 The Court further clarified that problems with “a finite number of identified, predictable solutions” would likely yield innovations founded on common sense and undeserving of patent protection.174 By this standard, older technology makes many inventions obvious and predictable.175 Inventions by advanced AI systems with creative and non-obvious characteristics, however, have increased processing capacities, widened access to searchable information, and increased efficiency in analyzing information—all of which would merit a patent if a human invented them.

In light of technological advancements such as AI, the non-obviousness hurdle must be set appropriately. “[I]f the hurdle is too high, deserving inventions will become unpatentable,” disincentivizing innovation.176 “[I]f the hurdle is too low, a flood of junk patents may” cause true inventors to face more infringement lawsuits, which also disincentivizes innovation.177 Taking these policy considerations into account, the Court stated in Graham that the non-obviousness analysis must be flexible and “amenable to a case-by-case development.”178 The Court took one step further in KSR and recognized the implications of technological advancement for non-obviousness by modifying the PHOSITA considerations.179 It remains unclear, however, to what extent a PHOSITA can integrate “the ambit of applicable art . . . .”180

We argue, however, that subject matter eligibility can categorically address the broader issues raised by AI, acting as patent law’s gatekeeper and directly addressing the types of inventions eligible for patent protection.181 As the Supreme Court stated, “[t]he obligation to determine what type of discovery is sought to be patented must precede the determination of whether that discovery is, in fact, new or obvious.”182 Therefore, subject matter eligibility serves two necessary functions: “checking the volume of patent applications, and . . . excluding subject matter that . . . [is] too costly to protect.”183

C. The Question of Infringement

As with inventorship, existing laws and precedent appear to rule out a machine or program as infringer.184 If a human becomes the owner of an AI’s invention, however, that person should bear responsibility for infringement. “[W]hoever without authority makes, . . . offers to sell, or sells any patented invention” is committing infringement.185 As the AI learns and modifies its behavior, moreover, it is possible that a resulting product, process, or action may infringe on one or more patent claims. Under current law, the induced infringer is “[w]hoever actively induces infringement of a patent.”186 The Federal Circuit has interpreted this to mean that the alleged inducer must have knowingly aided another’s direct infringement of a patent. When the AI system is operating on its own and not on the premises of the operator, does the owner then become the direct infringer? If the operator is operating the AI under instructions of the owner, then the owner and/or operator might be considered responsible for inducement187—an odd conclusion, given that the infringer typically does not change identity based on location. The issue might turn on the amount of control of the operation that has been given to the owner and the operator, respectively. The task of assigning responsibility for infringement becomes even more problematic if the AI system learns from data within the public domain or from a variety of publicly accessible sources. How does “a company developing an autonomous vehicle or robot, or even software that can run anywhere across a network, . . . safeguard against [infringement]?”188

Proving copyright infringement is also difficult. The plaintiff must prove substantial similarity after identifying and eliminating the elements that are unoriginal and unprotected.189 These tests may be hard to run on sophisticated systems that change data in a creative way.

Scholars suggest measures that might be implemented to reduce uncertainty, such as forbidding certain kinds of AI systems, requiring chips to identify the source of the owner, or even self-defense technology solutions against counterfeiting and copying.190 While reinventing alternative tools to prevent AI systems from copying other works or using protective data, we might think about solutions outside of the legal realm, such as “technology traffic lights” indicating sites that forbid intellectual property–protected materials or “stop signs” for forbidden zones.191 When the data is accessible, we need a “green light” to allow AI systems to enter. In any case, the intellectual property laws are outdated for these advanced technologies systems.

VI. An Alternative Model for Patent Law at the 3A Era—Incentivizing Stakeholders Within the AI Multiplayer Model

A. Rethinking the Incentive Effect of the Current Patent Regime

The patent reward in general is questionable. Scholars argue that patent laws fail to reach their goal of maximizing benefit to society. Granting twenty years of exclusive rights to the inventor or the inventor’s transferee may not significantly incentivize inventors. Surveys show that even CEOs in most industries see patent incentives as relatively unimportant.192 If corporate leaders who are driven by a profit motive do not see the value of patent incentives, then certainly such an incentive is meaningless to the multiple players and cumulative patent models. When determining whether to grant patent rights in uncertain situations, therefore, such as in regard to inventions by AI systems and the multiplayer and cumulative patent models, we should not overemphasize the importance of the patent system compared to other alternatives.

What are the flaws in patent incentives that make them irrelevant to AI? First, patent law only rewards the first inventor, while second comers get nothing. The drive to be first—even by an hour or two193—can force competing owners of AI systems deep into diminishing returns. This is wasteful.

Second, patent law grants exactly the same length of protection for all inventions,194 “regardless of [their] R&D costs and . . . other economically relevant factors.”195 Although this approach is simple, it “grossly overreward[s] some inventions and underreward[s] others.”196

Third, patents are superfluous for products that would be invented anyway.197 For instance, “conventional software developed rapidly even before courts were willing to grant it patent protection.”198 Patent protection serves a limited purpose.

As a fourth point, patent law fails in the multiplayer and cumulative patent environment characteristic of AI systems. It is not flexible in allocating rewards and is thus economically inefficient. Indeed, patent law only offers four ways to allocate rewards among two parties: “0:0 (both patents invalid), 50:50 (both patents upheld), 100:0 (first patent valid, second invalid), [and] 0:100 (first patent invalid, second valid).”199 Patent law may also impede future technological progress by making it harder for other AI systems to build on earlier inventions.200

Fifth, in practice, patent law would pose difficulties in bringing patent infringement actions against or on behalf of inventions by an AI or against the copying of an AI’s invention. Because of the unpredictable nature of AI, it is very difficult to identify the human that is responsible for the “actions” of an AI system. Sometimes the human involved in the process does not know how an AI system arrives at an invention. In other cases, the AI system can “break” data into electronic nanocomponents and rebuild it in different ways, thus rendering it impossible to establish proof of infringement.201

The inefficiency of the patent system regarding AI inventions forces us to imagine alternative venues to satisfy the needs and goals of stakeholders. These alternatives are described below.

B. Non-Patent Model Within the AI Multiplayer Paradigm

1. First-Mover Advantages

First-mover strategy grants the first entity to expose and sell its products or services to the public a competitive advantage: a head start over rivals and the chance to capture a significant portion of the market.202 First-mover advantages arise endogenously. In the case of AI systems, exposing new and advanced technology generates substantial profit margins, along with several other advantages.203

First, technology leadership. New innovative technology can provide significant cost advantages to first movers, allowing them to maintain leadership in market share. Technological pioneers can protect their research and development through patents. In most industries, however, patents confer only weak protection, are easy to invent around, and have transitory value given the pace of technological change. Patents account for only a small proportion of pioneers’ perceived quality advantages, and patent races can cause the downfall of firms unable to move quickly enough. In the case of AI systems, where, as we argue, patent laws are not applicable, this first-mover advantage plays an important role.204

Second, the first mover captures a monopoly-like status without other competitors with whom to share the market. This status usually means, in economic terms, gaining the monopoly in marginal revenue and cost by selling at higher rates (than the competitive balanced rate) as well as producing a lower quantity of products or services than demand.205 Both mechanisms generate higher profits than the open market.

Third, controlling the resources. In many markets, there is room for only a limited number of profitable firms; the first mover can often select the most attractive niches and strategically limit the amount of space available for subsequent entrants. First movers can establish positions in geographic or product space such that those that follow them find it unprofitable to occupy the interstices or brave the threat of price warfare, which is more intense when firms are positioned more closely. First movers remain committed due to sunk investment costs and enjoy greater advantages in larger economies of scale; they are therefore driven to maintain greater output following entry.206

Fourth, consumers’ loyalty after using certain products or services. Customers may follow the first brand they encounter that meets their demand and functions satisfactorily. Consumers will switch brands only when encountering a significant price differential; if pioneers convince a significant number of consumers to try their products, the products’ attributes may shape standards for the entire product category. Switching costs in order to win over individual customers is considered a burden. The benefits of familiarity with a brand tend to overcome the drive to search for alternatives.207

Fifth, blocking competitors in the future. First movers who capture the market can easily create hurdles for those who come after them—by, for example, reducing prices until the opponent is defeated (a dump process).208

In the AI industry, the invention process as well as product life cycles can sometimes be extremely short. Therefore, regulating the process for obtaining patents may be less important than adjusting the advantages for first movers in the market, particularly for recouping research and development investments shortly after marketing.209

The advantages enjoyed by the first mover sometimes translate into more lasting competitive advantages as well.210 First movers can more easily stay ahead of rivals by continually improving on their inventions.211 Switching costs can also deter existing customers from buying imitations of their products.212

First-mover advantages tend to fade over time and the frequency of cost switching often decreases over the years as buyers become more knowledgeable about competing products.213 Still, research shows that, while the average duration of the monopoly enjoyed by a first mover has declined, the first-mover advantage remains significant.214 Furthermore, the absolute size of sales per time unit increases for the first mover due to this effective monopoly.215

Therefore, we suggest that the relevant players within the Multiplayer Model who bring AI inventions to market will take advantage of being first movers instead of relying on an inapplicable and outdated patent regime.

2. Digital Tools Against Copying and Counterfeiting

The main hazard of nullifying patent law is counterfeiting. While contract laws affect the parties to contracts, patent laws influence the public, including third parties who are, under patent law, prevented from counterfeiting intellectual property even if they have not signed any contract.

We suggest implementing alternative digital tools that we dub “red lights” to prevent copying of protected materials. They will function even more efficiently than submitting court procedurals, which is pricey and time consuming.

Buying a device based on a patent allows the purchaser to “own” that device—to take it home, use it, put in on a shelf, or lend it to a friend. Products produced by AI systems, on the other hand, are capable of infringing on patents that protect data AI systems may find, for example, through independent web-surfing. Technical tools such as firewalls, we argue, may be the solution needed to protect AI systems and stop the counterfeiting of protected products,216 acting similarly to intellectual property law. Vendors of e-books, for example, can digitally delete books from the devices of consumers who have not paid for them, and without warning or explanation. Amazon deleted Orwell’s 1984 from the Kindles of surprised readers several years ago.217 Cloud storage, streaming, e-books, and other digital goods offer users convenience and flexibility, but they have potentially harmful effects on their privacy and other rights.218

These digital tools are already in wide use. The online marketplace eBay, for example, uses digital tools to hamper those infringing on intellectual property rights.219 Its website used to declare:

eBay developed the Verified Rights Owner (VeRO) Program to help protect not only intellectual property, but the consumer as well. Highlights of the program include: [e]xpeditious removal of listings reported to eBay by more than 5,000 intellectual property rights owners[; p]roactive monitoring and removal of listings that violate eBay policies designed to prevent the listing of infringing items on eBay[; a]bility to save searches and have the results emailed to you[; s]uspension of repeat offenders[; c]ooperation with rights owners seeking personal information on alleged infringers.220

Further discussion and regulation should welcome these advanced technology tools that will inevitably become part of our everyday digital lives.

3. Acknowledgement of Stakeholders Within the AI Industry

Society should reward the multiple players involved with AI—including the programmers, trainers, and operators who make contributions to scientific inquiry—by giving them the recognition their contributions deserve, even if neither the AI nor these players can be considered inventors and even when inventions are not patentable. This social recognition would likely encourage the players involved in AI systems to innovate further.

Inventions by AI are unpredictable; AI should be understood, therefore, to contain an inventive intuition, like the human mind. This creative intuition may derive from an AI’s features (“personality”) and may make it receptive to some sort of recognition.

Social recognition, such as through social networks, websites, trade journals, or even printing on the AI products themselves could serve as an effective alternative to granting patent rights or inventorship status to AI inventions. Rather than denying or degrading the invention’s utility, social recognition offers a psychological advantage to the people involved in generating the product, such as the inventor and operator, even if they fail to obtain patents.221

Making either the AI inventions or the AI software open source is another possibility. The rise of open source communities demonstrates that human beings are, on the one hand, inherently creative and, on the other, social and generous.222 “[V]olunteers almost always join [such communities] because of softer incentives [than profit, such as] altruism or a desire for education” or social recognition.223 A large disclosure database promotes information sharing.224 The advancement of and increased access to knowledge as well as advancing the welfare of all are fundamental goals of intellectual property law. Databases and voluntary knowledge sharing of open source data can strive toward these goals.225 Such sharing could also prevent non-practicing entities from engaging in rent-seeking behavior by providing a source of prior art.226

4. A Patent Reform Targeted at Inventions by AI

Unlike legal scholars who advocate for targeted patent reforms that apply only to certain cases, we have argued for general patent reform.227 There is long-standing resistance, however, to using technology-specific triggers and differentiated patent awards—including patent filing fees, non-obviousness standards, antitrust policies, and defenses against patent infringement228—to advance certain policies. Such triggers are considered impractical.229 When assessing the strength of patent protection in a particular industry, the government must consider many factors. “[Some] of these factors are hard to assess, and the sheer number of them would make the inquiry unpredictable . . . and vulnerable to manipulation. Moreover, the dividing lines between technologies are highly permeable and tend to shift rapidly as technology changes.”230 While drafting their patent claims, inventors can tailor them to categories that offer greater protection, rendering the technology-specific rules obsolete.231

Our solution—abolishing patent protection for AI inventions—does not solve the one-size-fits-all problem inherent to the U.S. patent system. It does not provide industry-dependent standards that consider dramatically different R&D costs and varying abilities to extract revenue from consumers across different industries. However, considering that the government knowingly overlooks these differences when it awards patent protection, our solution applies only to the challenges of AI and the deficiencies of patent law. This is no worse than current patent law or any alternative solution. As with the drawbacks of targeted patent reforms, many economic factors related to patent strength in a particular industry are difficult or impossible for the government to assess.232 Therefore, taking into account all factors, this Article’s solutions—being first in the market, electronic open source tools, and social recognition—are the most appropriate for inventions by AI.

VII. International Tools That Can Be Used to Harmonize AI’s Patents

Intellectual property laws are governed by international laws which countries are adopting and implementing. Today, almost every country has some form of patent system. Of the 196 countries in the world,233 191 are member states of the World Intellectual Property Organization (WIPO),234 a specialized agency of the United Nations. Since the development of AI is still in an early stage, no enacted statutes or case law in the world directly addresses AI. Nevertheless, many international treaties help provide intellectual property protection across national boundaries. Therefore, any solution regarding patent protection of AI products, including ours, should be consistent with international tools. We briefly discuss three tools as examples.

The Patent Cooperation Treaty of 1971 (PCT) is a multilateral treaty administered by WIPO.235 It facilitates patent registration across the world by making it possible to seek patent protection for an invention in more than one country simultaneously.236 The PCT has more than 142 member states and more than 100 national and regional patent offices.237 “Perhaps the [PCT’s] greatest strength [is its] diversity of legal, linguistic, and national cultures . . . .”238

A PCT filing consists of an international phase and a national phase.239 During the initial international phase, an “International Searching Authority” conducts an international prior art search.240 Next, during the national phase, applicants may pursue their PCT patents in the patent offices of individual countries.241 To do so, applicants must file requests to the PCT, file translations of their applications, and pay the national fees.242 After the national phase, the substantive laws of individual countries apply exclusively.243

In practice, however, this seemingly streamlined PCT filing process has proven redundant. The redundancies stem from the independent prior art search by patent offices of individual countries giving little to no deference to the prior art search already conducted during the international phase of the PCT filing process.244 This means that, to seek worldwide patent protection within the PCT framework, individual inventors are obligated to incur additional costs in both time and money.245

Given that the patent offices of individual countries examine patent applications based on their own standards, the PCT does not promote consistency.246 The Draft Substantive Patent Law Treaty (SPLT), a WIPO initiative, aims to deeply harmonize the patentability requirements set by patent offices of the member states.247 Under the SPLT, member states would agree, among other things, to adopt identical rules about non-obviousness or inventiveness and the requisite amount of information to be disclosed.248 According to Professors Reichman and Dreyfuss, “[a] free-standing instrument, such as the SPLT, would shrink the remaining flexibilities in the TRIPS Agreement with no side payments and no concessions to the catch-up strategies of developing countries at different stages of technological advancement.”249 Nevertheless, there are some departures from U.S. patent law’s non-obviousness standards in the patent systems of other jurisdictions and the international community.

While the European Patent Office (EPO) uses the term “inventive step” instead of “non-obviousness,” the core of the inquiry remains the same. It relies on determining whether or not the PHOSITA could have, in an obvious manner, derived the solution to the technical problem. The standard for non-obviousness is similar to the U.S. standard in Japan as well, as evaluation of the inventive step requires determining if there are suggestions to the claimed invention in the prior art.250

Conclusion

Stephen Hawking stated: “[T]he short-term impact of AI depends on who controls it, the long-term impact depends on whether it can be controlled at all.”251

Are we navigating the unknown, or can we conceive of a one-size-fits-all solution for the near future? We believe that the answer lies between these two scenarios.

Companies have already attempted to implement AI in several areas. General Electric, for example, is already using an AI system in the design of jet engines.252 Engineers in Wisconsin have optimized efficiency and minimized emissions for diesel engines,253 engineers in Virginia designed a satellite communications antenna, car and truck companies are using driverless cars, lawsuits are being submitted by robots, and diseases are being detected—all with AI systems.254 However, there are still those who argue that genetic algorithms have never reached the level of success their proponents envisioned.255

On the one hand, as we have argued, traditional patent laws are no longer adequate or efficient. On the other hand, many open questions (perhaps more than answers) remain. For example, which norm should we apply when there is only one or a few stakeholders involved in inventing and operating the system? Who is responsible for infringement of people’s or entities’ rights by an AI system? Who is entitled to the income stemming from patents developed by AI systems? If ownership is subject to contractual consent, firms need to rethink what the impact of this new realm on their businesses should be and determine whether to modify their policies accordingly. Can new arrangements (re)allocate property rights when the AI’s invention is not subject to patent right protection, as we suggest in this Article?256 If so, what types of new agreements are needed to ensure that inventions resulting from AI are owned by a specific entity? How should joint development agreements be modified to ensure ownership of technology developed by AI? When entities relying on AI systems seek to insure themselves against claims regarding infringements, what kind of insurance should they use? Once data is being used to teach the AI how to determine right and wrong, are license agreements necessary? If so, how should they be modified?

We can wait for scholars and policy makers to decide. Or we can create an AI algorithm trained to reach the best solution to these open questions.257

 


* Dr. Shlomit Yanisky Ravid, Professor of Law; Yale Law School, Information Society Project (ISP), Fellow; Fordham University School of Law, Visiting Professor; Ono Academic Law School, Israel (OAC), Senior Faculty; the Shalom Comparative Legal Research Center, OAC, Founder and Director. Special thanks to Professor Bonnie Kaplan at Yale University for her outstanding contributions and valuable feedback. Thanks also to Yale Law School’s Information Society Program (ISP) and especially to Jack Balkin, Rebecca Crootof, Michael Fischer, Christina Spiesel, Asaf Lubin, and Nimrod Kozlovsky as well as to Fordham Law Center for Law and Information Policy and in particular to Professor Joel Reidenberg, Sivan Saban-Hacohen, and Rivi Cohen for their encouragement, insight, and comments. We are grateful to Michele Woods, Michal Svantner, and Victor Lopez Vazquez at the World Intellectual Property Organization (WIPO) in Geneva, Switzerland, to the Swiss Institute of Comparative Law in Lausanne, Switzerland, and in particular to Lukas Heckendorn Urscheler, Alberto Aharonovitz, Ilaria Pretelli, Karen Druckman, and Sadri Saieb for enabling this Article. We are grateful as well to the participants of the Beyond Intellectual Property Course at Fordham Law School and in Urbino University, Italy as well as to the workshops being held at, among other institutes, the Yale Law School, ISP, on the challenges of Artificial Intelligence, new technologies, and robots for the wonderful feedback on our research. Finally, gratitude to our gifted research assistants and outstanding legal editors Elizabeth Ledkovsky, Esq., Avishai Zur, Sophie Hogan, and Laura Lagone, Esq. Xiaoqiong (Jackie) Liu, is an engineer; Fordham Intellectual Property Institute Fellow; former editor of the Fordham Intellectual Property, Media & Entertainment Law Journal.