Can AI be an inventor in materials discovery?

In this article, three faculty members at the University of Toronto in the fields of energy materials discovery for a sustainable future, machine-learning-aided biomedical science, and corporation law, finance, securities, regulation, and venture capital ask and discuss the question visible in many aspects of our lives: “In the process of discovery, can AI be an inventor on a patent and an owner of intellectual property?” In this article, three faculty members at the University of Toronto in the fields of energy materials discovery for a sustainable future, machine-learning-aided biomedical science, and corporation law, finance, securities, regulation, and venture capital ask and discuss the question visible in many aspects of our lives: “In the process of discovery, can AI be an inventor on a patent and an owner of intellectual property?” The legal definition of an inventor on a patent encompasses an individual or a group of individuals, human in nature, responsible for making an original contribution to at least one of the claims outlined in the patent. Inventorship on a patent necessitates more than just incremental technical enhancements to existing concepts and principles; it must transcend the obviousness benchmark threshold for individuals skilled in the art of the relevant subject matter, and it must also be original and distinctive from what exists in the related prior art in that field. This tried-and-true premise is now under pressing scrutiny due to the question of whether artificial intelligence (AI) can legitimately qualify as an inventor on a patent. This query bears implications for the idea-to-innovation pipeline and the legal landscape concerning ownership and use of intellectual property (Figure 1). The history of intellectual property legislation can be traced to the Venice Patent Statute drafted in 1474. It was expanded in the Paris Convention for the Protection of Industrial Property of 1883 and later in the World Trade Organization agreement on Intellectual Property Rights of 1994. These treaties established the foundation and mandated the standards of the international patent system, but AI has now disrupted how science is conducted and has changed the way inventions are made; it has shaken the footing upon which patent law is grounded, as the inventor is an intelligent machine. Within this context, the inquiry into whether software can be an inventor has been the recent focus of some high-profile cases in the United States judicial system. A notable instance revolves around the US Court of Appeals for the Federal Circuit’s ruling last year in the case of Thaler v. Vidal. In that case, the US Patent and Trademark Office (PTO) rejected two patent applications filed by computer scientist Stephen Thaler on behalf of his AI system in which the AI system was listed as the sole inventor. Thaler sued the PTO, arguing that the rejection of AI-created patents restrains the ability of the current patent system to effectively stimulate innovation, facilitate technological advancement, and attract investment, all of which ultimately benefit both the economy and society at large. The Court of Appeals for the Federal Circuit, however, held that US patent law unambiguously requires an inventor to be a natural human being, which AI systems are not, despite their impressively advancing prowess in computer intelligence, creativity, and intuition. In this case, Stephen Thaler, a computer scientist, president, and chief executive of US-based AI firm Imagination Engines, has petitioned the US Supreme Court to make a final judgment on the verdict that only human inventors and not AI systems can be the legal creator of generated inventions. He posits that the rejection of AI-created patents restrains the ability of the current patent system to effectively stimulate innovation, facilitate technological advancement, and attract investment, which ultimately benefits the economy and society of a nation. In practical terms, as AI machines increasingly acquire skills akin to human intelligence, creativity, and intuition, it is true that they can rapidly explore massive material chemistry and materials science data libraries more efficiently than humans to predict new structures and compositions that optimize specific properties of the known materials dataset. One can envision a scenario where AI inventors outsmart and outpace humans to the point where they overwhelm examiners and the system with patent applications. This raises the question: should we have AI forensic patent examiners? It would be a kind of judicial “Turing test” appraisal of listed AI inventors on a patent to determine whether they were clever enough to be mistaken for a human by exhibiting the ability to think with a mind of their own. Imagine a futuristic AI patent litigation courtroom scene with an AI robot judge at the bench, an AI jury pool in the jury box, AI lawyers at their council table, and the AI plaintiff and defendant closely following the action. If this happens, it seems inevitable that AIs will ultimately become the litmus test of non-obviousness (i.e., that which is not obvious to someone skilled in the relevant field of technology). In that case, human inventors may be entirely frozen out. With AI’s access to all knowledge in the subject matter of the field, this situation would speed things up but also potentially make the whole process rather complex and non-transparent. Also, while we tend to think of AI as one thing, it most certainly is not. Different AI programmers will inevitably design different AI algorithms, potentially yielding inconsistent answers on the question of patentability. Disagreements can arise in matters such as what is obvious, what is a practitioner of average skill in the art, and so forth. Similarly, different AIs may make different decisions on what merits inclusion in the underlying database. By what process would the programmatic parameters be determined? Would an AI judge participate in that process? Under current patent law, the absence of human involvement in the computational materials discovery process does not justify an inventive claim within a patent, nor would it withstand scrutiny in a legal dispute. Moreover, if an AI-driven materials discovery was subsequently reduced to practice in an AI-enabled and known materials synthesis protocol, then the same argument against patentability would apply (Figure 2). On the other hand, should the need arise for human creativity and ingenuity to intervene, revealing a novel synthesis pathway due to the inability of AI to accomplish the task, then the legal definition of an invention would be validated and its patentability warranted. Take, for example, a novel material for CO2 photocatalysis that was unveiled in a machine learning search but that required the intelligence and experience of a human chemist to devise a successful synthetic pathway to realize the material in practice. Furthermore, if a non-obvious application of a materials technology emerges, such as a novel product or process capable of an industrial application, through human initiative, it would likely pass the benchmark for patentability within a legal context. Today, the patent system is based on inventions by humans, the legal owners or assignors of a materials discovery patent and the just reaper of the rewards. One of the big questions being hotly debated currently focuses on the rightful ownership of the rewards derived from AI-generated materials discovery patents. There are lots of potential actors who merit consideration. These include the coder(s) of the AI software, the proprietor of the computer that ran the AI, and the creators of the database on which the AI was trained. Nor should the software user be overlooked. After all, the end user is instrumental in figuring out what questions are put to the AI software—or more likely, what sequence of questions—in order to extract a creative and useful answer from the AI. There is certainly skill in that process that is not to be overlooked, unless the AI is writing its own codes and programming its own software, which raises new issues. The issue of rightful ownership is inextricably tied to—and indeed, is arguably the same thing as—the vital question of creating appropriate incentives to innovate. The incentive question is complicated, as it not only involves figuring out which actors need to be incentivized, but also by how much. It may be that these questions can be resolved in a relatively straightforward fashion by using our best guesses as to how different parties in the creative process should be rewarded (and hence incentivized) and embodying these in a default set of rules that can be contracted around. That way, if the default rules are sub-optimal, private parties are free to depart from them and to agree to whatever configuration works best for them. This not only employs market forces in allocating rewards, but also allows for different sets of private actors to arrive at their own custom-made solutions. For example, suppose that the default rule specifies that the AI programmers earn 1/3 of the lucre from a patented invention. This default rule, if sub-optimal, could be departed from in various ways. For example, one could imagine a corporation paying staff programmers a fixed salary (perhaps augmented by either formulaic or discretionary bonuses) to design AI software, the rewards from which then reside, by contract, entirely with the programmers’ corporate employer. Alternatively, independent AI programmers might sell their AI programs holus-bolus to a corporate end-user, thus capturing via the sale price their share of the expected value generated. Many other ownership configurations are possible, allocating rewards in a myriad of ways. There is also the question of how long an AI-generated patent should last. The current 20 years seems much too long. It is based on the assumption that the time, cost, and effort expended in producing novel inventions is very great. Thus, a lengthy period of protection is necessary to properly incentivize inventive activity. That mold ill fits AI-spawned inventions, which can almost certainly be produced much more quickly and at much less cost. Granted, with technological progress as fast as it has become, the period of protection becomes more academic, as new patents cease to have commercial value much quicker than was once the case. Nonetheless, where AI-spawned inventions are concerned, a protective period of perhaps two to four years might be more appropriate. In the future, this AI paradigm of focusing on solving specific problems with a lower intelligence level than humans might undergo transformation with the emergence of artificial general intelligence (AGI) systems. These AGI systems are engineered to tackle any challenge that a human can with equivalent or potentially superior intelligence, marking the pinnacle of artificial superintelligence (ASI). This leveling of the intelligence playing field of humans and machines has the potential to reshape the legal landscape governing patent rights and intellectual property ownership. It is conceivable that the arrival of ASI could lead to the development of specialized AI machines tailored to address specific problems, thereby introducing a new dimension to patentability. These are all complex legal issues that will likely be hard to enforce by existing patent law. One way to resolve the problem would be the creation of an international treaty of bespoke intellectual property law that serves to protect inventions created by artificial intelligence. These laws should be designed to follow standardization principles and ensure efficient resolution of disputes, as delineated in a recent commentary.1George A. Walsh T. Artificial intelligence is breaking patent law.Nature. 2022; 605: 616-618Crossref Scopus (5) Google Scholar And, as noted above, there is a good argument that the rules should be default rules from which private actors may choose to depart if they choose. The ramifications of a legal moratorium on patented inventions conceived by AI would have profound consequences, such as the effect of demotivating funding agencies and businesses to pursue worthwhile research with perceived limits on the return on investment. A legal restriction on AI inventors could deprive society of innovations that, for example, enable vaccines and antibiotics to save lives, renewable energy systems to ameliorate climate change, and fertilizers to provide sufficient food for a growing global population. Or perhaps it will just motivate inventors to use AI systems to do the bulk of the work and then pass off the resulting inventions as their own. On a final note, the advent of AI is poised to usher in a plethora of opportunities. Materials chemists, scientists, and engineers will still exist, but they will do things very differently. For example, some of them will develop AI algorithms; some of them, with a profound comprehension of these AI tools, will adeptly wield them in order to advance their research and development endeavors; and some, with all the resources they have and a bit of luck, will be fortunate enough to unearth truly disruptive innovations to start the next industrial revolution and even garner a Nobel Prize. In envisioning the future, imagination is our limit. G.A.O. acknowledges support from the Natural Sciences and Engineering Research Council of Canada (NSERC). C.Q. acknowledges support from Ontario Institute for Cancer Research (OICR) as an OICR Early Career Investigator. The authors declare no competing interests.