The United States Constitution[1] provides the basis for patent laws; it says "Congress shall have 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" (emphasis added). It is a common refrain that the Constitution is a living, breathing document, and the afore-mentioned clause is perhaps an exemplar. Even though our founding fathers are unlikely to have contemplated artificial intelligence ("AI") inventors, it is noteworthy that nothing in the clause itself states that "authors and inventors" have to be individuals[2]. Therefore, allowing AI inventorship is unlikely to run afoul of the plain language of the Constitution.
Title 35 the United States Code is the statutory basis for the federal patent system. Section 100[3] of Title 35 defines the term "inventor" to mean "the individual or, if a joint invention, the individuals collectively who invented or discovered the subject matter of the invention." Various other sections require acts that only an individual can be expected to perform, at least for now. For example, Section 112[4] specifies the requirements for a written description:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
(emphases added). If we were to unpack some these requirements, one could ask how a patent draftsperson can "set forth the best mode contemplated by the inventor" if an AI model is the inventor. Could we seriously argue that the AI model contemplated the invention? One definition of "contemplate" is to "think profoundly and at length." Does the electrical and computational activity of a computer rise to the level of profound thinking? Philosophers and neuroscientists may be more equipped to delve into the intricacies of this question.
Another definition of "contemplate" is to "to view or consider with careful and thoughtful attention." Proponents of AI inventorship may argue that some neural network architectures (e.g., transformers, long short-term memory networks (LSTMs)) are designed to include "attention mechanisms" precisely to mimic human cognitive attention. For example, the recent public interest in chatbots, such as ChatGPT and Bard, highlight the power of the transformer architectures.
Therefore, a simple statutory term such as "contemplated" invokes semantic, linguistic, philosophical, and technological references, all intricately intertwined with the notion of inventorship.
The aforementioned Section 112 also requires that the written description contain the "manner and process of making and using [the invention]." This is an important provision in the statute since the patent is a bargain between the inventor and the public. The public grants the inventor a temporary monopoly of twenty years from the filing date of the patent application, with the understanding that the technology will be available to the public beyond the expiry of the patent. If the written description does not provide the manner and process of making the invention, then the inventor side of the bargain is not fulfilled. As every patent draftsperson knows well, details about the invention are elucidated from the inventors during a disclosure meeting. How would one elucidate such information from an AI model? Can an AI model know how to use an output of the model?
Again, proponents of AI inventorship may argue that the details about the invention may be extracted from the AI model. For example, some AI models may be configured to output data about how predictions or outputs are generated, and such data could be used to fulfill the written description requirement. Such data is generally referred to as "explanations" or "interpretations" of the AI model's decision making process. For example, "feature importance" can identify the most important input features that the AI model used to make its predictions. An image recognition model may be able to indicate which parts of the image the model focused on to recognize an image. Another technique is "local explanation" that involves generating explanations for individual predictions or outputs of an AI model. For example, a local explanation may indicate why a given input led to a particular output. Similarly, a "global explanation" can be an indication of an overall summary of how the AI model functions. Also, "counterfactual explanations" can involve generating alternative input values that may result in a different output. This can be used to determine which input values were primarily influential in driving the AI model's prediction.[5]
Transformer-based model architectures have a highly interpretable structure, and internal weights and activations of the model may be measured. The transformer architecture consists of a series of stacked self-attention layers, where each layer applies a set of weights to the input sequence to compute a new representation of the sequence. The self-attention mechanism involves computing attention weights that indicate how much each position in the sequence should be weighted when computing the representation. The weights used for computing the attention and the representations at each layer can be extracted and analyzed. This can provide insights into how the model is processing and integrating information, and which parts of the input are most important for making predictions. There are several tools and techniques available for visualizing and interpreting the weights and activations of transformer models, including attention heat maps, saliency maps, and activation maximization. These methods can help to identify patterns and relationships in the input data that are most relevant for the model's predictions.
However, even if the written description requirement were to be fulfilled, as described above, the patent statute relies on other aspects of individuality, such as an inventor's oath or declaration, an assignment, a person of ordinary skill in the art, and so forth. Inventors also appear in depositions or as trial witnesses in patent disputes. However, such issues are statutory and/or requirements of a judicial process. The United States Patent and Trademark Office[6], the Federal Circuit[7], and the Supreme Court[8], have made it clear that under the current law, an inventor is an individual. The authority to decide if an AI model can be an inventor lies with Congress. The U.S. Constitution delegates such authority to Congress, and not to the courts. Therefore, if Congress were to decide, in its infinite wisdom, that the term "authors and inventors" can include AI inventors, then so it shall be, and statutory provisions and/or judicial processes can be suitably modified and adapted.
One of the arguments made by proponents of AI inventorship is that if an AI model were not allowed to be an inventor, as is the current state of the law, then many "inventions" may go unclaimed and unpatented. It is argued that this will impede the "promot[ion of] the progress of science and useful arts," which is the fundamental constitutional underpinning of the intellectual property system. To the extent an AI model may predict new outputs not contemplated by the designer of the model, the model architecture is itself man-made. It is an intricate network of nodes and layers that is trained to process certain types of inputs and make certain predictions. Therefore, it is under the control and direction of the designer of the model, and therefore the designer has a substantial legal claim as the individual that conceived any part of that output.
It is conceivable that there are "inventions" that may indeed be unclaimed or unpatented. How can this notion be aligned with the current state of the law? One perspective could be that we are creating new paradigms of an invisible nature. Consider, for example, a rock that is beautifully sculpted by a gushing stream, or a tree that is air brushed to exquisite beauty by a wayward wind. The sculpted rock and the tree would be considered natural phenomena, and neither the stream nor the wind would be allowed to claim a copyright for such objects, or a patent for the process of making them. Perhaps, in the same vein, by designing AI models, we are creating new paradigms of an invisible, hidden, natural world of probabilistic, electro-magnetic "winds and streams," and what we are witnessing are natural wonders from this natural world that dazzle and amaze us. However, from a patentability perspective, these are simply natural phenomena, and are therefore not patentable subject matter. Perhaps as technology advances, such seeming novelties will eventually become mundane, and it could become difficult to distinguish between an output of a general purpose computer and that of a specific purpose computer (like the AI models), thereby raising the bar for patentability.
One consideration is that even if the U.S. legal system is slow to adapt to the notion of the AI model as an inventor, other nations may make this possible, creating new intellectual property rights. For example, Australia came tantalizingly close to this reality.[9] In such cases, U.S. commercial interests may suffer and IP rights may be usurped. One strategy may be to protect the unpatentable subject matter as a trade secret. However, in some cases, the subject matter may be easily reverse engineered or duplicated. In such cases, a defensive strategy may be to make the subject matter publicly available, for example, as a defensive publication.
Certainly, there will come a time, in the not too distant future, when AI models will become self-replicating, independent, unsupervised, decision making sentient beings. Some already claim that certain AI models are sentient. As intellectual property and technology professionals, we need to consider the larger implications of allowing an AI to be an inventor. Such a broadening of the notion of an "individual" will be an earth shattering change to a centuries' long held view about who an "individual" is. It will be a tacit acknowledgement that our Constitution views AI models on an equal footing with individual inventors. This has the potential to open up a Pandora's Box of dilemmas: must AI models be accorded equal rights in society (Voting rights? Moral rights? Life, liberty, and the pursuit of happiness? Right to bear arms?). These are significant considerations that we as a society must ponder. We need to deliberate, prepare, and shape our future by creating laws that can pre-empt and adapt to the fast pace of technological advancement. Perhaps prior to designating AI models as inventors, we need to regulate AI models. John Steinbeck may as well have been talking about intellectual property rights when he wrote:
Ever'body wants a little piece of lan'. I read plenty of books out here. Nobody never gets to heaven, and nobody gets no land. It's just in their head. They're all the time talkin' about it, but it's jus' in their head.[10]
And then again, the AI model may mock us by quoting Shakespeare[11]: "I will keep where there is wit stirring, and leave the faction of fools."
[1] U.S. Const. Article 1, Section 8, clause 8
[2] Several of the founding fathers, including Benjamin Franklin, contemplated a universe with other intelligent beings
[3] 35 U.S.C. § 100(f)
[4] 35 U.S.C. § 112(a)
[5] Certain portions of this paragraph and the next are based on conversations with ChatGPT
[6] In re Application No. 16/524350
[7] Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022)
[8] Denying certiorari in Thaler v. Vidal (No. 22-919)
[9] The Australian Full Federal Court recently reversed a decision by the Federal Court that an AI system known as DABUS can be an inventor under the Australian Patents Act 1990
[10] Of Mice and Men, ch. 4, John Steinbeck
[11] William Shakespeare, Troilus and Cressida - Act 2, scene 1
