Providing the USPTO With Input About AI Inventors

Current law and court rulings have confirmed that a patent application cannot list an AI system as an inventor.  If the law were to change, how could it be implemented (e.g., how are patent rights allocated or assigned)? And under current law, what about an invention created jointly by people and an AI system (e.g., is the portion invented by the AI system part of the invention that can be patented)?

The comments expressed herein are solely the views of the individual authors, and are not necessarily the views of Morgan, Lewis & Bockius LLP or Berkeley Law. Participants included:

• David V. Sanker, Ph.D., Morgan, Lewis & Bockius LLP
• Jianbai Wang, Ph.D., Morgan, Lewis & Bockius LLP
• Alex Choi, law student at Berkeley Law
• Sandeep Stanley, law student at Berkeley Law
• Jin Cho, law student at Berkeley Law
• Vernon Espinoza Valenzuela, law student at Berkeley Law
• Kriti Rathi, law student at Berkeley Law
• Michelle D’Souza, law student at Berkeley Law
• Sophia Goepfert, law student at Berkeley Law
• Grace Pan, law student at Berkeley Law
• Nina Zhang, law student at Berkeley Law
• Kathy Bui, law student at Berkeley Law
• Kristy Lam, law student at Berkeley Law
• Ellen Jin, law student at Berkeley Law
• Sanika Nayak, law student at Berkeley Law
• Edlene Miguel, law student at Berkeley Law
• Eric Ahern, law student at Berkeley Law
• Harshini Malli, law student at Berkeley Law

• New drug discovery. See, e.g., “AI is dreaming up drugs that no one has ever seen. Now we’ve got to see if they work.” MIT Technology Review. This article points out that “all told, around two dozen drugs (and counting) that were developed with the assistance of AI are now in or entering clinical trials.”

  Current methods for screening potential antibiotics are expensive and time-consuming. Researchers at DeepMind and Google have addressed this with their AI software, AlphaFold. The AI model has been able to accurately predict protein structures from amino acid sequences, and researchers are hopeful they can use the AlphaFold structures to find new antibiotics that bind to specific bacterial proteins. Researchers at MIT have used another machine-learning algorithm to discover halicin, a promising new antibiotic compound to address growing antibiotic resistance. This algorithm was trained using the structures of known antibiotic molecules to predict the properties of thousands of unresearched molecules. Halicin, which has a different chemical structure from known antibiotics, was unlikely to be discovered without AI assistance. See article.

• New materials. See “Scientists use machine learning to accelerate materials discovery.” Argonne National Laboratory. This article describes “an automated process for identifying and exploring promising new materials by combining machine learning (ML) — a type of artificial intelligence — and high performance computing.” For example, using the single element carbon, the algorithm was able to predict how the atoms can order themselves in a range of temperatures and pressures. The algorithm then constructed a series of “phase diagrams” - a map that helps guide researchers in their discovery of new and useful matter. In other cases, machine learning has been used to reduce the energy consumed in manufacturing processes used to produce melt-blown non-woven materials. See article..

• According to current U.S. practice, the contributions of these systems would be considered inventive if performed by humans.

It is useful to consider an analogy with two scenarios that only involve humans. In a first scenario, an ingenious person conceives of an invention, and describes it for others to carry out. In this simple scenario, only the person who conceives of the invention is an inventor. In a second scenario, a person (such as a product manager) writes up a business problem in the form of a requirements specification and gives the set of requirements to a development team. The development team devises a solution to the problem. In this second scenario, the person who wrote up the requirements is generally not an inventor, because the requirements do not specify an implementation or a concrete solution. Only the members of the development team are inventors (and only those who actually contributed to conceiving of the solution).

Using a technical tool (including an AI tool) versus an AI inventor largely mimic these two scenarios. In the first scenario, a person defines a specific task for a tool, and the tool performs the instructions. In the second scenario, one or more humans provide a high level set of requirements to an AI system, and the AI system conceives of the solution.

An important distinction between these scenarios is who is in control of the inventive process. For example, a person may write an automated test script to evaluate a large array of potential solutions. In these cases, the human inventors are still in control of the process. This contrasts with the usage of an AI inventor, where the AI inventor is in control of the process.

A second important distinction is the level of detail provided to the AI system. When a person provides all of the details or writes a specific set of program instructions to perform, the person has done the “hard” work. But when a person just specifies a goal with a limited set of inputs, the AI system is doing the “hard” work.

In sum, the distinction between using a tool and having an AI inventor involves measuring the role of human activity to input data or parameters that kickstart the invention process, and also measuring the human activity to utilize the generated output. If formulating the specific input to an AI system is essential to the inventive process, then the inventive process is likely based primarily on the human contribution, particularly if the human formulation requires specific knowledge or skill. On the other hand, if the formulation of the question requires no special skill or knowledge (e.g., “build me a better mousetrap”), then the inventive process is based primarily on the AI contribution.

Since AI “inventors” are not recognized as inventors under U.S. patent law, they should not be listed in a patent application. Indeed, because inventors must be people, the term “inventor” inherently means humans under current law, so referring to an “AI inventor” is a linguistic contradiction. Therefore, by listing all of the humans who have contributed to an invention, an applicant has correctly identified the inventors. There is no falsification and no error that would require correction under 35 U.S.C. § 256.

Since current U.S. patent law does not permit AI inventors, they should not be listed. A key question is how this affects patentability when a patent claim has multiple elements and some of those elements where not “invented” by the listed human inventors.

The are three basic options regarding “mixed inventorship”:

• (i) the invention is not patentable;
• (ii) the invention is patentable, subject to the usual standards under 101, 102, 103, and 112; or
• (iii) the claims are analyzed to identify the AI conceived features, and those features are excluded from the patentability analysis. Only the human-invented features are considered for patentability.

It would be difficult to justify option (i). There is at least one human inventor, so what policy would entail rejection without substantive evaluation? This option would appear to be inconsistent with the constitutional goal to “promote the progress of science and useful arts.”

Option (iii) has some initial appeal but has some substantial implementation issues. It is useful to analogize this to two other IP scenarios.

First, compare this to the recent copyright decision for the novel “Zarya of the Dawn.” The author of the text could get a copyright for her work, but the AI generated images were excluded from the copyright. If the author of the text had paid a person to create the artwork in a work made for hire, she would have been able to get a copyright for everything. The author was able to get a copyright for the human-created portion, which emphasizes an important distinction between copyrights and patents. Whereas a copyright on portions not generated by AI is meaningful, patents are granted on claims as a whole. Removing any portion could doom the entire patent protection.

Second, consider patent examination in Europe. An examiner looks at a claim, removes any element that does not have an alleged “technical character,” and then evaluates patentability on what is left.

Trying to apply this “technical character” type model to differentiating between human and AI invented features would create several practical and legal problems. First, unlike “technical character,” which an EPO examiner can discern by just reading the words, there is no way to differentiate between human and AI invented features of a claim without requiring substantial additional work for patent applicants to provide the necessary data. One speaker at the West Coast Listening Session suggested that patent applicants could just write “a couple of pages” in the specification about what aspects an AI system created. Not only is there no statutory basis to require such additional disclosure, but neither patent applicants nor patent practitioners have an incentive to spend substantial additional time and money for the disclosure. Budgets for patent application drafting are less now than they were 20 years ago, so there is no budget for additional pages about AI inventorship.

Further, even if such a requirement were imposed, it would be both difficult and burdensome for applicants to comply. Even with only human inventors, trying to correlate individual inventors to specific claim features would be difficult, and that ignores the fact that claims evolve during prosecution.

Finally, such a methodology would be subject to substantial error and substantial incentive to minimize disclosure about AI and downplay the role of AI.

In the context of disclosure incentives, the USPTO will also need to consider how this applies to inventions that are fully created by an AI system. Although the issue of AI inventors is now settled law in the United States, the law does not change the fact that AI systems are inventing and will continue to do so (and increasing). Since AI systems are inventing and companies want to protect their inventions, there is a built-in incentive to “stretch” the contributions of humans, particularly when an AI system in the primary “inventor.” If a company has a billion dollar invention that it needs to protect, ethical constraints may not be a sufficient deterrent. An important topic to consider is how to incentivize full disclosure without asking patent applicants to abdicate patent rights.

In conclusion, since AI inventors cannot be listed, listing all of the human inventors (and there is a requirement to have at least one) should be enough. In the absence of recognizing AI inventors, option (ii) is the most practical solution.

There does not appear to be a compelling reason to differentiate based on the number of human inventors.

It is useful to ask the same questions in a context where there are no AI inventors. Consider the case of a single human inventor. To conceive of an invention, the human inventor uses a wealth of knowledge gleaned over many years from teachers, professors, colleagues, textbooks, scientific articles, online research, and other sources. These sources have “trained” the human inventor. In addition, the human inventor may use a variety of hardware and software to perform experiments, such as a computer, simulation software, a microscope, or many other tools.

In the context of a human inventor, there has been no push to grant patent rights to any of the sources of the inventor’s knowledge or to grant patent rights to the owners or creators of any hardware or software tools that the inventor uses.

There does not appear to be any compelling reason to grant patent rights to tenuous indirect sources that helped a person conceive of an invention. If AI inventors are eventually permitted under the law, the same analysis should apply.

In general, if a company spends a lot of money and/or resources for an AI system, the company would expect to own the IP created by the AI system.

This is a policy question that should be guided by the Constitutional provision “to promote the progress of science and useful arts.”

NO COMMENTS.

As discussed above for question 3b, any such requirement would be challenging to implement, and would create an incentive to deviate from a full honest disclosure. In addition, it would be difficult to justify such a requirement under existing patent law.

It is useful to evaluate this question in the context of 35 U.S.C. § 112. Section 112 already requires that the written description must be “in full, clear, concise, and exact terms” and must set forth the best mode for carrying out the invention such that it enables someone skilled in the art to make and use the invention without undue experimentation. This existing language is clear, has survived the test of time, and is thorough.

Under this existing framework, there are two scenarios to consider. In the first scenario, an explanation of the AI contribution is necessary in order to understand the “manner and process of making and using” the invention. In this scenario, patent applicants are already required to describe the AI contribution. This applies, even if the AI system acted as a black box. No patent can be granted if the disclosure does not enable others to practice the invention.

In a second scenario, understanding the AI contribution is not necessary to practice an invention. For example, if an AI system selects a small number of candidate molecules (e.g., from among millions) for further testing as a treatment for a specific disease, it would generally be sufficient to articulate the structure of the candidate molecules, without specifying how the AI system selected them.

NO COMMENTS.

NO COMMENTS.

Although the USPTO is not authorized to change the patent statutes, it is useful to evaluate what would need to change if AI inventors are permitted. It is also important to evaluate this in the context of patent offices in countries around the world.

Two important considerations are assignments and declarations. An AI system cannot, by itself, assign its rights or sign a declaration under penalty of perjury that it actually created an invention. Both of these are core parts of the US patent system. And both of these documents act to prevent stealing an invention from the true inventor. It is very unlikely to build a consensus to eliminate either of these. This creates a practical problem for allowing AI inventors.

If Congress determines that allowing AI inventors “promotes the progress of science and useful arts,” and does not want to eliminate assignment and declaration documents, then a proposed solution is to permit a human surrogate to sign such documents on behalf of an AI inventor. This was first proposed by David V. Sanker and Jianbai Wang in the September 30, 2019 Daily Journal article “Can the US Patent and Trademark Office handle ‘artificial inventors’? This would require specific rules about who can act as a surrogate, but would not change the wording or usage of assignment or declaration documents. In particular, the human surrogate would have to declare under penalty of perjury that the AI system actually created the invention. This would create an incentive for AI inventive systems to be transparent.

The concept of a human surrogate is already in use in specific cases under 35 U.S.C. § 115(d), so expanding this to accommodate AI inventors is not as big of a leap as might be imagined. Specifically, Section 115(d) allows for a Substitute Statement when an inventor has died, is under legal incapacity, cannot be found after diligent effort, or has an obligation to sign by refuses to do so.

Another issue to address if AI inventors are permitted is to precisely define what is considered an AI inventor. For example, should different instances of the same AI system be considered as a single inventor or distinct inventors? Unlike humans, AI systems can be easily cloned and used independently. For example, two or more distinct companies could use instances of the same AI system to create the same or a similar invention. If the instances are considered as separate AI inventors, then each could create prior art to invalidate the other. On the other hand, if the AI system is considered a single inventor, ownership is less clear, particularly if the instances are sharing data, and it is unclear whether prior art from one instance can be used against another instance. This issue becomes even more complex when considering software upgrades and updated training of AI models. If data from multiple instances is used to retrain all of the instances, how would the instances maintain any meaningful identity? In sum, to contemplate allowing AI inventors would require deciding how to handle both cloning and merging of AI system instances.

One good example is in Australia, which has the concept of “Statement of Entitlement” rather than requiring an explicit assignment. This more flexible concept is better suited to accommodate AI inventors. It is not perfect, but illustrates that flexibility can be accommodated

NO COMMENTS.