Tech Companies Take Note: USPTO Guidance on AI Inventions Will Issue in 2024

Published Date
Jan 12, 2024
To ring in the new year, U.S. IP litigation associate Alan Billharz considers the forthcoming guidance on AI inventions that the U.S. Patent and Trademark Office is set to provide in 2024.

On October 30, 2023, President Biden signed the Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. Technology companies working with AI should take note of two upcoming patent-related deadlines in the Order. By February 27, 2024, the U.S. Patent and Trademark Office (USPTO) must publish guidance addressing inventorship and the use of AI in the inventive process, including illustrative examples in which AI systems play different roles in inventive processes and how, in each example, inventorship issues ought to be analyzed. The USPTO must then, by July 26, 2024, issue additional guidance to address other considerations at the intersection of AI and intellectual property, which could include updated guidance on patent eligibility for AI-related technologies. 

For the past few years, the USPTO has been actively exploring the intersection of AI and IP. Based on the prior statements and decisions of the USPTO, we predict at least some of the guidance that the USPTO will likely include in its official responses to the Order. We also highlight some of the more difficult questions that the USPTO may choose to address in its official response to the Order, or in future guidance to come.

Inventorship and AI

As the Order recognized, a primary issue for patents and AI is the extent to which an applicant may use AI and still qualify as an inventor. Conception is the touchstone of inventorship, and inventorship has traditionally required that a human inventor conceive of the patented invention.1

Furthest out on the spectrum is the case where the applicant admits that an AI device or program has invented the underlying technology completely on its own. An intuitive response to this extreme case is that an unconscious machine cannot conceive of inventions, and therefore cannot invent. The USPTO has firmly agreed with that response, as have the courts. In 2019, Stephen Thaler filed a patent application with an AI device, DABUS, listed as the inventor. Thaler stated explicitly that DABUS had conceived of the invention. The USPTO denied Thaler’s application, determining that the patent statute “consistently refers to inventors as natural persons.”2 The USPTO also relied on precedent from cases where applicants had listed states and corporations, rather than humans, as inventors.3 The U.S. Court of Appeals for the Federal Circuit affirmed the USPTO’s decision, and the U.S. Supreme Court declined to hear an appeal.4

We should expect the USPTO to address more complex scenarios. For example, accepting that an AI cannot be listed as an inventor, what about a human AI practitioner who creates a useful invention while using an AI system? Of course, a practitioner can claim conception with respect to designing the AI itself, but if the AI produces a useful model for, say, predicting the weather, can the practitioner also claim to have conceived that model? Conception requires more than “unrecognized accidental creation.”5 Indeed, “conception requires that the inventor appreciate that which he has invented,” possessing a “definite and permanent idea of the complete and operative invention.”6 If the human AI practitioner fails to appreciate how the AI-generated model operates to achieve a particular result, then the human may not qualify as the inventor of that AI-generated model either. Thus, there is some risk that certain AI-generated inventions may not be patentable at all. 

The use of machine learning and other AI algorithms is unlikely to be a categorical barrier to conception, however. The USPTO has already granted many AI patents, and it has even given an example of a machine learning algorithm as a patentable technology in its issued guidance (though in the context of subject matter eligibility rather than inventorship). Nevertheless, we could see guidance from the USPTO finding a failure of conception where the purported inventor relies on opaque AI algorithms. One potential solution may be to clarify that a practitioner can conceive of an invention generated by an AI “black box” if the practitioner recognizes the benefit of that invention—similar to how a chemical compound patented for one purpose by a first inventor may be recognized as having another patentable purpose by a second inventor.  In this way, AI-generated inventions stretch the bounds of conception, so further guidance in this area from the USPTO could impact research methods and allocations at technology companies. 

To navigate these conception challenges, applicants may avoid disclosing that they used AI in the inventive process. Of course, future litigants may discover inventors’ use of AI in a lawsuit concerning the validity of the patent. However, there is currently no obligation for inventors to disclose their use of AI to the USPTO, and the USPTO has not yet discussed any AI disclosure requirement. (Indeed, the USPTO itself has utilized AI to determine which patents incorporate AI.) That approach is consistent with the lack of any other duty to disclose specific methodologies employed in the process of invention—that is, beyond the general requirements of written description and enablement of the invention itself. Because AI strikes close to the heart of invention, we may see further guidance from the USPTO regarding the disclosure of the use of AI tools.

Patent Eligibility and AI

Another primary issue for patents and AI is the extent to which AI technologies are even eligible for protection. Like other algorithmic technology, there is significant concern that AI technologies could be found unpatentable because they are directed to abstract ideas, similar to mathematical concepts. Yet the USPTO has marked as informative one Patent Trial and Appeal Board decision holding that a machine learning patent was not abstract. Moreover, the USPTO has provided a subject-matter eligibility example indicating that relatively standard applications of machine learning are directed to patentable subject matter. Example 39 is directed to “[a] computer-implemented method of training a neural network for facial detection.”7 The USPTO explains that “some of the limitations may be based on mathematical concepts, [but] the mathematical concepts are not recited in the claims.”8 Similarly, in a 2019 decision,9 the PTAB held that a “neural network device” comprising “a collection of node assemblies” is not a mental process. And in yet another 2019 decision marked informative, the PTAB held a speech recognition system developed using “end-to-end deep learning” was not a mental process because, although the steps of the algorithm could theoretically be performed mentally, they could not performed practically.10

It may be difficult, however, for inventors utilizing “black box” algorithms to satisfy the enablement and written description requirements. Example 39 and issued machine learning patents suggest that these requirements may be satisfied for an AI algorithm, but the USPTO may provide more robust guidance in its responses to the Order. And, because conception is a prerequisite for both inventorship and written description,11 the USPTO’s guidance on either issue may shed light on the other. 

Another issue that may be addressed by the USPTO is whether AI algorithms that evolve over time based on initial conditions and training techniques can satisfy the definiteness requirement for patents. Example 39 includes claim language that could produce infinitely many models depending on the data set used and training techniques—e.g., “training the neural network in a first stage using the first training set.”12 Thus, Example 39 suggests the USPTO may be willing to provide some wiggle room as to how a model is claimed, provided that its purpose is described with sufficient definiteness. That approach probably will not work for some AI algorithms, however. For example, in unsupervised learning algorithms, patterns emerge naturally through the algorithm’s formation of associations among the training data. This technical distinction could be utilized by the USPTO (or by the courts) to draw lines on patentability and inventorship. Unsupervised models could present an example of an invention “conceived by a machine,” or at least not conceived by a natural person. 

Finally, the sharing of data sets and the usage of “foundation models”—general-purpose models that can be repurposed for specific applications—may create patentability and obviousness concerns. As machine learning algorithms become conventional, training them to perform certain tasks may become the next iteration of “do it on a computer,” the traditional argument against patent eligibility for claims directed to performing conventional tasks using generic computers.13 The USPTO may determine that the application of widely known machine learning techniques to widely available data sets is categorically unpatentable or obvious, regardless of the particular application. Again, Example 39 seems to suggest not, but it is worth considering that Example 39 issued five years ago, and the USPTO has since received significant public commentary through its AI and Emerging Technology Partnership. Thus, the USPTO may provide more guidance about the role training data plays in the obviousness or patentability of AI technologies.

Watch This Space

While the USPTO has been engaged in significant examination of the intersection of AI and IP for several years now, many significant questions regarding AI and IP remain—including those set forth above. The USPTO’s prior efforts suggest at least some of the guidance it will issue in response to the Order, but practitioners will want to prepare for many possible outcomes when the USPTO’s guidance issues in February and July of this year. 

Special thanks to law clerk Evan Wainright for his contributions to this article.



1Fina Oil & Chem. Co. v. Ewen, 123 F.3d 1466, 1473 (Fed. Cir. 1997).
2In re Appl. No. 16/524,350, Decision on Petition, at 4 (USPTO Apr. 22, 2020).
3Id. at 4-5 (citing Burroughs Welcomed Co. v. Barr Labs., Inc., 40 F.3d 1223, 1227-28 (Fed. Cir. 1994) (“Conception is a ‘mental act’ that must be performed by a natural person.”)).
4Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), cert. denied, 143 S. Ct. 1783 (2023).
5Invitrogen Corp. v. Cleantech Lab’ys, Inc., 429 F.3d 1052, 1063 (Fed. Cir. 2005).
7Subject Matter Eligibility Examples: Abstract Ideas, 2019 Revised Patent Subject Matter Eligibility Guidance, at 8 (USPTO Jan. 7, 2019).
8Id. at 9.
9Ex Parte Henry Markram, Rodrigo De Campos Perin, & Thomas K. Berger, No. APPEAL 2018-008166, 2019 WL 2244864 (P.T.A.B. May 14, 2019).
10Ex Parte Linden, No. 2018-003323, 2019 WL 7407450, at *5 (P.T.A.B. Apr. 1, 2019).
11Falko-Gunter Falkner v. Inglis, 448 F.3d 1357, 1367 n.13 (Fed. Cir. 2006).
12Subject Matter Eligibility Examples: Abstract Ideas, at 8.
13Alice Corp. Pty. v. CLS Bank Int’l, 573 U.S. 208, 223-24 (2014).

Content Disclaimer
This content was originally published by Allen & Overy before the A&O Shearman merger