Patenting the Output of Autonomously Inventive Machines

By Ryan B. Abbott

©2017. Published in Landslide, Vol. 10, No. 1, September/October 2017, by the American Bar Association. Reproduced with permission. All rights reserved. This information or any portion thereof may not be copied or disseminated in any form or by any means or stored in an electronic database or retrieval system without the express written consent of the American Bar Association or the copyright holder.

An innovation revolution is on the horizon. Artificial intel ligence (AI) has been generating inventive output for decades, and now the continued and exponential growth in computing power is poised to take creative machines from novelties to major drivers of economic growth. A creative singularity in which computers overtake human inventors as the primary source of new discoveries is foreseeable.

This phenomenon poses new challenges to the traditional paradigm of patentability. Computers already are generating patentable subject matter under circumstances in which the computer, rather than a human inventor, meets the requirements to qualify as an inventor (a phenomenon I refer to as “computational invention”).1 Yet, it is not clear that a computer could be an inventor or even that a computer’s invention could be patentable. There is no statute addressing computational invention, no case law directly on the subject, and no pertinent US Patent and Trademark Office (USPTO) policy.2

These are important issues to resolve. Inventors have ownership rights in their patents, and failure to list an inventor can result in a patent being held invalid or unenforceable. Moreover, government policies encouraging or inhibiting the development of creative machines will play a critical role in the evolution of computer science and the structure of the R&D enterprise. Soon computers will be routinely inventing, and it may only be a matter of time until computers are responsible for most innovation.

This article addresses whether a computer could and should be an inventor for the purposes of patent law as well as whether computational inventions could and should be patentable. It argues that computers can be inventors because although AI would not be motivated to invent by the prospect of a patent, computer inventorship would incentivize the development of creative machines. In turn, this would lead to new scientific advances.

This article also examines some of the implications of computer inventorship for other areas of patent law. For instance, computers are a natural substitute for the person having ordinary skill in the art (PHOSITA or, simply, the skilled person) used to judge a patent’s inventiveness. The skilled person is presumed to know of all the prior art (what came before an invention) in a particular field—a legal fiction that could be accurate in the case of a computer. Substituting a computer for the skilled person also suggests a need to expand the scope of prior art, given that computers are not limited by human distinctions of scientific fields. This would make it more challenging for inventions to be held nonobvious, particularly in the case of inventions that merely combine existing elements in a new configuration (combination patents). That would be a desirable outcome, although the new test would create new challenges.

Creative Computers and Patent Law

Computers Independently Generate Patentable Results

Computers have been autonomously creating inventions for decades. In 1994, computer scientist Stephen Thaler disclosed an AI system he called the “Creativity Machine,” a computational paradigm that “came the closest yet to emulating the fundamental neurobiological mechanisms responsible for idea formation.”3 The Creativity Machine is able to generate novel ideas through the use of a software concept referred to as artificial neural networks—essentially, collections of on/off switches that automatically connect themselves to form software without human intervention. The Creativity Machine independently generated the subject matter of a patent granted in 1998.4 Dr. Thaler listed himself as the inventor on the patent and did not disclose the Creativity Machine’s involvement to the USPTO.

Similarly, software modeled after the process of biological evolution, known as genetic programming (GP), has succeeded in independently generating patentable results.5 Evolution is a creative process that relies on a few simple processes: “mutation, sexual recombination, and natural selection.”6 GP emulates these same methods digitally to achieve machine intelligence. As early as 1996, GP succeeded in independently generating results that were the subject of past patents.7 In 2005, a patent was issued for subject matter independently generated by a GP-based AI named the “Invention Machine.”8

The Creativity Machine and the Invention Machine may be the earliest examples of computer inventors, but others exist.9 Moreover, the exponential growth in computing power over the past dozen years combined with the increasing sophistication of software should have led to an explosion in the number of computational inventions. Indeed, it is likely that computers are inventing more than ever before.

Human and Computer Involvement in Computational Inventions

Requirements for Inventorship

For an individual to be an inventor, he or she must contribute to an invention’s “conception.” Conception refers to “the formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention as it is thereafter to be applied in practice.”10 It is “the complete performance of the mental part of the inventive act.”11 After conception, someone with ordinary skill in the invention’s subject matter (e.g., a chemist if the invention is a new chemical compound) should be able to reduce the invention to practice. That is to say, he or she should be able to make and use an invention from a description without extensive experimentation or additional inventive skill. Individuals who simply reduce an invention to practice—for example, by describing an already conceived invention in writing or by building a working model from a description—do not qualify as inventors.

The Role of Computers in Inventive Activity

Although computers are commonly involved in the inventive process, in most cases, computers are essentially working as sophisticated (or not-so-sophisticated) tools. In such instances, a computer may assist a human inventor in reducing an invention to practice, but the computer is not participating in the invention’s conception. Even when computers play a more substantive role in the inventive process, such as by analyzing data in an automated fashion, retrieving stored knowledge, or recognizing patterns of information, the computer still may fail to contribute to conception. Computer involvement might be conceptualized on a spectrum: on one end, a computer is simply a tool assisting a human inventor; on the other end, the computer independently meets the requirements for inventorship. AI capable of acting autonomously such as the Creativity Machine and the Invention Machine fall on the latter end of the spectrum.

The Role of People in Inventive Activity

Just as computers can be involved in the inventive process without contributing to conception, so can people. For now, at least, computers do not entirely undertake tasks on their own accord. Computers require some amount of human input to generate creative output.

Yet, simply providing a computer with a task and starting materials would not make a person an inventor.12 Imagine I give my engineer friend some publicly available battery schematics and ask her to develop an iPhone battery with twice the standard battery life. If she succeeds in developing such a battery, I would not qualify as an inventor of the battery by having instructed her to create a result.

People are also necessarily involved in the creative process because computers do not arise from a void—humans have to create computers. Once again, that should not prevent computer inventorship. I would not exist without my parents contributing to my conception (pun intended), but that does not make my parents inventors on my patents. If a computer scientist creates an AI to autonomously develop useful information and the AI creates a patentable result in an area not foreseen by the inventor, there would be no reason for the scientist to qualify as an inventor on the AI’s result. An inventor must have formed a “definite and permanent idea of the complete and operative invention” to establish conception.13 The scientist might have a claim to inventorship if he developed the AI to solve a particular problem and it was foreseeable that the AI would produce a particular result.

Combining Human and Computer Creativity

A computer may not be a sole inventor; the inventive process can be a collaborative process between person and machine. By means of illustration, suppose a human engineer provides a machine with basic information and a task. The engineer might learn from the machine’s initial output, then alter the information that he or she provides to the machine to improve its subsequent output. After several iterations, the machine might produce a final output that the engineer might directly alter to create a patentable result. In such a case, both the engineer and the machine might have played a role in conception. Leaving AI aside, invention rarely occurs in a vacuum, and there are often joint inventors on patents. In some of these instances, if a computer were human, it would be an inventor. Yet, computers are not human, and as such they face unique barriers to qualifying as inventors.

Barriers to Computer Inventorship

The Legal Landscape

Congress is empowered to grant patents on the basis of the patent and copyright clause of the Constitution. That clause enables Congress “[t]o 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.”14 It also provides an explicit rationale for granting patent and copyright protection, namely to encourage innovation under an incentive theory. The theory goes that people will be more inclined to invent things (i.e., promote the progress of science) if they can receive government-sanctioned monopolies (i.e., patents) to exploit commercial embodiments of their inventions. Having the exclusive right to sell an invention can be tremendously lucrative.

The Patent Act, which here refers to United States patent law as a whole, provides a few challenges to computers qualifying as inventors under the patent and copyright clause. First, the Patent Act requires that inventors be “individuals.”15 This language has been in place since at least 1952 legislation that established the basic structure of modern patent law. The “individual” requirement likely was included to reflect the constitutional language that specifically gives “inventors” the right to their discoveries as opposed to other legal entities such as corporations that might assert ownership rights. Such language would help to ensure that patent rights were more likely to go to individual inventors than to corporate entities where ownership was disputed. Legislators were not thinking about computational inventions in 1952. Second, patent law jurisprudence requires that inventions be the result of a “mental act.”16 So, because computers are not individuals and it is questionable that they engage in a mental act, it is unclear whether a computer autonomously conceiving of a patentable invention could legally be an inventor.

Avoiding Disclosure of Artificially Intelligent Inventors

Given that computers are functioning as inventors, and likely inventing at an escalating rate, it would seem that the USPTO should be receiving an increasing number of applications claiming computers as inventors. That the USPTO has not suggests that applicants are choosing not to disclose the role of AI in the inventive process. That may be due to legal uncertainties about whether an AI inventor would render an invention unpatentable. Without a legal inventor, new inventions would not be eligible for patent protection and might enter the public domain after being disclosed.

There is another reason why computers might not be acknowledged: a person can qualify as an inventor simply by being the first individual to recognize and appreciate an existing invention. That is to say, someone can discover rather than create an invention. Uncertainty (and accident) is often part of the inventive process. In such cases, an individual need only understand the importance of an invention to qualify as its inventor. Thus, assuming that a computer cannot be an inventor, individuals who subsequently “discover” computational inventions by mentally recognizing and appreciating their significance would likely qualify as inventors. So, it may be the case that computational inventions are only patentable when an individual subsequently discovers them.

In Support of Computer Inventors

Computers Should Qualify as Legal Inventors

Arguments Supporting Computer Inventors

Preventing patents on computational inventions by prohibiting computer inventors, or allowing such patents only by permitting humans who have discovered the work of creative machines to be inventors, is not an optimal system. In the latter case, AI may be functioning more or less independently, and it is only sometimes the case that substantial insight is needed to identify and understand a computational invention. Imagine that I give my AI some publicly available battery schematics and instruct it to develop an iPhone battery with twice the standard battery life. The AI could produce results in the form of a report titled “Design for Improved iPhone Battery”—complete with schematics and potentially even preformatted as a patent application. It seems inefficient and unfair to reward me for recognizing the AI’s invention when I have not contributed significantly to the innovative process.

Such a system might also create logistical problems. If I created an improved iPhone battery as a human inventor, I would be its inventor regardless of whether anyone subsequently understood or recognized the invention. If I instructed my AI to develop an improved iPhone battery, the first person to notice and appreciate the AI’s result could become its inventor (and prevent me from being an inventor). One could imagine this creating a host of problems: the first person to recognize a patentable result might be an intern at a large research corporation or a visitor in someone’s home. A large number of individuals might also concurrently recognize a result if access to an AI is widespread.

More ambitiously, treating computational inventions as patentable and recognizing creative computers as inventors would be consistent with the constitutional rationale for patent protection. It would encourage innovation under an incentive theory. Patents on computational inventions would have substantial value independent of the value of creative computers; allowing computers to be listed as inventors would reward human creative activity upstream from the computer’s inventive act. Although AI would not be motivated to invent by the prospect of a patent, it would motivate computer scientists to develop creative machines. Financial incentives may be particularly important for the development of creative computers because producing such software is resource intensive. Though the impetus to develop creative AI might still exist if computational inventions were considered patentable but computers could not be inventors, the incentives would be weaker owing to the logistical, fairness, and efficiency problems such a situation would create.

There are other benefits to patents beyond providing an ex ante innovation incentive. Permitting computer inventors and patents on computational inventions might also promote disclosure and commercialization. Without the ability to obtain patent protection, owners of creative computers might choose to protect patentable inventions as trade secrets without any public disclosure. Likewise, businesses might be unable to develop patentable inventions into commercial products without patent protection. In the pharmaceutical and biotechnology industries, for example, the vast majority of expense in commercializing a new product is incurred after the product is invented during the clinical testing process required to obtain regulatory approval for marketing.

Arguments against Computer Inventors

Those arguments reflect the dominant narrative justifying the grant of intellectual property protection. That account, however, has been criticized, particularly by academics. Patents result in significant social costs by establishing monopolies.17 Patents also can stifle entry by new ventures by creating barriers to subsequent research.18 Whether the benefit of patents as an innovation incentive outweighs their anticompetitive costs, or whether patents even have a net positive effect on innovation, likely varies between industries, areas of scientific research, and inventive entities.

For instance, commentators have argued that patents may not be needed to incentivize R&D in the software industry.19 Software innovation is often relatively inexpensive, incremental, quickly superseded, produced without patent incentives, protected by other forms of intellectual property, and associated with a significant first-mover advantage. Likewise, patents may be unnecessary to spur innovation in university settings where inventors are motivated to publish their results for prestige and the prospect of academic advancement.20

Of course, computational invention patents may not be an all-or-nothing proposition; they may further encourage activities that would have otherwise occurred on a smaller scale over a longer time frame. If patents are not needed to incentivize the development of creative computers, it may be justifiable to treat computational inventions as unpatentable and fail to recognize computer inventors. Yet, whether patents produce a net benefit as an empirical matter is difficult to determine a priori. Even though individuals and businesses do not always behave as rational economic actors, in the aggregate, it is likely that providing additional financial incentives to spur the development of creative computers will produce a net benefit.

Patents for computational inventions might also be opposed on the grounds that they would chill future human innovation, reward human inventors who failed to contribute to the inventive process, and result in further consolidation of intellectual property in the hands of big business. Other nonutilitarian patent policies may similarly fail to support computer inventorship.21

Ultimately, despite concerns, computer inventorship remains a desirable outcome. The financial motivation it will provide to build creative computers is likely to result in a net increase in the number of patentable inventions produced. Particularly, while quantitative evidence is lacking about the effects of computational invention patents, courts and policymakers should be guided first and foremost by the explicit constitutional rationale for granting patents. Further, allowing patents on computational inventions as well as computer inventors would do away with what is essentially a legal fiction—the idea that only a human can be the inventor of the autonomous output of a creative computer—resulting in fairer and more effective incentives.

Computer Inventors Are Permitted under a Dynamic Interpretation of Current Law

Whether a computer can be an inventor in a constitutional sense is a question of first impression. If creative computers should be inventors, as this article has argued, then a dynamic interpretation of the law should allow computer inventorship. Such an approach would be consistent with the founders’ intent in enacting the patent and copyright clause and would further the purpose of the Patent Act. It would not run afoul of the chief objection to dynamic statutory interpretation, namely interference with reliance and predictability and the ability of citizens “to be able to read the statute books and know their rights and duties.”22 That is because a dynamic interpretation would not upset an existing policy; permitting computer inventors would allow additional patent applications rather than retroactively invalidate previously granted patents, and there is naturally less reliance and predictability in patent law given its highly dynamic subject area and struggle to adapt to constantly changing technologies.

Computer inventorship should not be prohibited based on statutory text designed to favor individuals over corporations. It would be particularly unwise to prohibit computer inventors based on literal interpretations of texts written when computational inventions were unforeseeable. If computer inventorship is to be prohibited, it should only be on the basis of sound public policy. Computational inventions may even be especially deserving of protection because computational creativity may be the only means of achieving certain discoveries that require the use of tremendous amounts of data or that deviate from conventional design wisdom.

Implications of Computer Inventorship

Computational Invention Ownership

If computers are recognized as patent inventors, there remains the question of who would own these patents. Computers cannot own property, and it is safe to assume that “computer personhood” is not on the horizon. This presents a few obvious options for patent ownership (assignment) such as a computer’s owner (the person who owns the AI as a chattel), developer (the person who programmed the AI’s software), or user (the person giving the AI tasks). The owner, developer, and user may be the same person, or they may be different entities.

Ownership rights to computational inventions should vest in a computer’s owner because it would be most consistent with the way personal property (including both computers and patents) is treated in the United States and it would most incentivize computational invention. It would encourage owners to promote access to their AI systems with the hopes of having users generate patentable results.

This also reveals an important reason why computational invention works best when the computer is the legal inventor. If computational inventions were treated as patentable but computers could not be inventors, then presumably the first person to recognize a computer’s invention would be the legal inventor and patent owner. That means the computer’s user, rather than its owner, would likely be the patentee as the person able to first recognize a computational invention. To the extent this is an undesirable outcome, the best solution may be to permit computer inventorship.

Assignment of computational inventions to a computer’s owner could be taken as a starting point, although parties would be able to contract around this default. As computational inventions become more common, negotiations over these inventions may become a standard part of contract negotiations.

Rethinking the Ultimate Test of Patentability

Considering the case for computer inventorship provides insight into other areas of patent law. Take, for instance, the nonobviousness requirement for grant of a patent.23 Part of the requirement’s evaluation involves employing the legal fiction of a PHOSITA (or skilled person) who serves as a reference for determining whether an invention is nonobvious. Essentially, an applicant cannot obtain a patent if a skilled person would have found the difference between a new invention and the prior art (what came before the invention) obvious. The test presumes that the skilled person is selectively omniscient, having read, understood, and remembered every existing reference from the prior art in the relevant field of invention (analogous art). A federal judge explained that the way to apply the obviousness test is to “first picture the inventor as working in his shop with the prior art references—which he is presumed to know—hanging on the walls around him.”24

Needless to say, no actual person could have such knowledge, but the standard helps avoid difficult issues of proof related to an inventor’s actual knowledge; also, it prevents obvious variations of publicly disclosed inventions from being patented. Stopping obvious variations from being patented is important because that prevents the removal of knowledge from the public domain. Inventions that would have been obvious to skilled persons are already within reach of the public. This raises the bar to obtaining a patent—a result that is desirable because patents should not be granted lightly given their anticompetitive effects. At the same time, creating too high a bar to patentability is undesirable because then patents would fail to adequately incentivize researchers. A balance is needed. Ideally, the system would only issue patents for inventions that would not have been created but for the expectation of obtaining a patent. The importance of the nonobvious requirement to patentability has led to its characterization as the “ultimate condition of patentability.”25

The idea of a PHOSITA understanding all of the prior art in the field was always fictional, but now it is possible for a skilled entity, in the form of a computer, to possess such knowledge. This makes the skilled computer a natural substitute for the hypothetical skilled person. The standard would require a skilled computer rather than a creative computer for the same reason that the skilled person is not an inventive person. A PHOSITA has traditionally been characterized as skilled at repetitive processes that produce expected results. If the skilled person were capable of inventive activity, then inventive patent applications would appear obvious.

Replacing the skilled person with the skilled computer suggests a change to the nonobviousness test. At present, the test takes into account the skilled person’s knowledge of the prior art. Decreasing the universe of prior art would make it easier to get a patent because, with less background knowledge, a new invention would be more likely to appear inventive. Likewise, expanding the universe of prior art would raise the patentability bar. Although it may be unrealistic to expect a human inventor to have knowledge of prior art in unrelated fields, there is no reason to limit a computer’s database to a particular subject matter. A human inventor may not think to combine cooking recipes with advances in medical science, but a computer would not be limited by such self-imposed restrictions. Now that humans and computers are competing creatively, the universe of prior art should be expanded.

This change would produce a positive result. The PHOSITA standard has been the subject of extensive criticism, most of which has argued the criteria for assessing nonobviousness are not stringent enough and therefore too many patents of questionable inventiveness are issued.26 Expanding the scope of prior art would make it more challenging to obtain patents, particularly combination patents. The Supreme Court has particularly emphasized “the need for caution in granting a patent based on the combination of elements found in the prior art.”27 The scope of analogous prior art has consistently expanded in patent law jurisprudence, and the substitution of a skilled computer would complete that expansion.

Of course, the new standard would pose new challenges. With human PHOSITAs, juries are asked to put themselves in the shoes of the skilled person and decide subjectively what that person would have considered obvious. A jury would have a difficult time deciding what a “skilled” computer would consider obvious. They could consider some of the same factors that are applied to the skilled person, or perhaps the test could require a combination of human and computer activity. For example, the skilled computer might be a skilled person with access to a computer’s unlimited database of prior art.


It is important for policymakers to give serious consideration to the issue of computer inventorship. There is a need for the USPTO to issue guidance in this area, for Congress to reconsider the boundaries of patentability, and for the courts to decide whether computational invention is worthy of protection. Doing so and recognizing that computers can be inventors will do more than address an academic concern; it will provide certainty to businesses, afford fairness to research, and promote the progress of science. In the words of Thomas Jefferson, “ingenuity should receive a liberal encouragement.”28 What could be more ingenious than creative computers?


1. Ryan Abbott, Hal the Inventor: Big Data and Its Use by Artificial Intelligence, in Big Data Is Not a Monolith 187 (Cassidy R. Sugimoto et al. eds., 2016).

2. See Ben Hattenbach & Joshua Glucoft, Patents in an Era of Infinite Monkeys and Artificial Intelligence, 19 Stan. Tech. L. Rev. 32, 44 (2015).

3. See What Is the Ultimate Idea?, Imagination Engines Inc., http://www.imagination-engines.com (last visited July 5, 2017).

4. U.S. Patent No. 5,852,815 (filed May 15, 1998).

5. John R. Koza, Human-Competitive Results Produced by Genetic Programming, 11 Genetic Programming & Evolvable Machines 251, 265 (2010).

6. John R. Koza et al., Evolving Inventions, Sci. Am., Feb. 2003, at 52.

7. See Koza, Human-Competitive Results, supra note 5, at 255–56, 265.

8. U.S. Patent No. 6,847,851 (filed July 12, 2002).

9. See, e.g., Daniel Riester et al., Thrombin Inhibitors Identified by Computer-Assisted Multiparameter Design, 102 Proc. Nat’l Acad. Sci. USA 8597 (2005).

10. Townsend v. Smith, 36 F.2d 292, 295 (C.C.P.A. 1929).

11. Id.

12. Ex parte Smernoff, 215 U.S.P.Q. 545, 547 (Bd. App. 1982) (“[O]ne who suggests an idea of a result to be accomplished, rather than the means of accomplishing it, is not a coinventor.”).

13. Townsend, 36 F.2d at 295.

14. U.S. Const. art. I, § 8, cl. 8.

15. E.g., 35 U.S.C. § 100(f) (“The term ‘inventor’ means the individual or, if a joint invention, the individuals collectively who invented or discovered the subject matter of the invention.”).

16. Townsend, 36 F.2d at 295.

17. See Daniel J. Hemel & Lisa Larrimore Ouellette, Beyond the Patents–Prizes Debate, 92 Tex. L. Rev. 303, 314–15 (2013).

18. See Arti Kaur Rai, Regulating Scientific Research: Intellectual Property Rights and the Norms of Science, 94 Nw. U. L. Rev. 77, 133 (1999).

19. See William M. Landes & Richard A. Posner, The Economic Structure of Intellectual Property Law 312–13 (2003).

20. See Mark A. Lemley, Are Universities Patent Trolls?, 18 Fordham Intell. Prop. Media & Ent. L.J. 611, 621 (2008).

21. Abbott, supra note 1, at 196.

22. See William N. Eskridge Jr. & Philip P. Frickey, Statutory Interpretation as Practical Reasoning, 42 Stan. L. Rev. 321, 340 (1990).

23. See 35 U.S.C. § 103.

24. In re Winslow, 365 F.2d 1017, 1020 (C.C.P.A. 1966).

25. See Nonobviousness—The Ultimate Condition of Patentability (John Witherspoon ed., 1980).

26. See, e.g., Michael D. Frakes & Melissa F. Wasserman, Does the U.S. Patent and Trademark Office Grant Too Many Bad Patents?: Evidence from a Quasi-Experiment, 67 Stan. L. Rev. 613 (2015).

27. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 415 (2007).

28. Diamond v. Chakrabarty, 447 U.S. 303, 308–09 (1980) (quoting 5 Writings of Thomas Jefferson 75–76 (Henry A. Washington ed., 1871)).

Ryan B. Abbott

Ryan B. Abbott is a professor of law and health sciences at the University of Surrey School of Law in the United Kingdom, and an adjunct assistant professor at David Geffen School of Medicine at the University of California, Los Angeles.

This article is adapted from the author’s article “I Think, Therefore I Invent: Creative Computers and the Future of Patent Law,” 57 B.C. L. Rev. 1079 (2016).