Attention Is All You Need: Prior Art in the Age of AI

In recent years, the rise of artificial intelligence (AI) has precipitated both excitement and apprehension, sparking conversations regarding the far-reaching effects of AI on the intellectual property landscape. In April 2024, such conversations prompted the U.S. Patent & Trademark Office (USPTO) to issue a Request for Comments (RFC) regarding the impact of the proliferation of AI on prior art, the knowledge of a person having ordinary skill in the art (PHOSITA), and determinations of patentability. Among the 15 questions posed by the RFC, several (paraphrased below) were directed to AI-generated prior art:

• Can AI-generated content be recognized as “prior art” under current U.S. patent law, or is the generation of prior art limited to human creators?
• Should we treat AI-generated disclosures differently from those generated without AI, considering the potential for AI to produce inaccurate information (e.g., hallucinations) and how might this influence their status as prior art?
• At what point could the volume of AI-generated prior art be sufficient to create an undue barrier to the patentability of inventions? What if this volume is sufficient to detract from the public accessibility of prior art (i.e., a PHOSITA exercising reasonable diligence may not be able to locate relevant disclosures)?

We examine the potential for AI-generated content to flood the corpus of prior art, as well as the effects AI-augmented PHOSITAs would have on the eligibility of printed publications as prior art under Title 35 of the U.S. Code § 102(a)(1). We argue that although traditional frameworks for evaluating the eligibility of printed publications built by a long and rich history of case law remain generally robust against AI-generated content, these frameworks may not be appropriate in a future world of AI-armed PHOSITAs. Instead of attempting to draw a line in the sand between human- and AI-generated content (which is increasingly a gray area), we should return to the fundamental principles that underlie patent law. In particular, we posit that prior art eligibility criteria for printed publications should not distinguish between human- and AI-generated works, but should instead consider whether a publication has received sufficient attention from a human audience.

‍

What is Accessibility?

The traditional standards for evaluating the eligibility of printed publications as prior art are enablement and accessibility. Here we focus on accessibility. Two seminal cases, In re Hall (Fed. Cir. 1986) and In re Cronyn (Fed. Cir. 1989), both centered around student theses. Though both were technically accessible to the public, Hall was indexed by subject matter in the university’s main library, whereas Cronyn was merely indexed by author name on cards stuffed away in a shoebox in a departmental library. The Federal Circuit determined that the former was considered to be sufficiently accessible to PHOSITAs, while the latter was not. In Jazz Pharm. v. Amneal Pharm. (Fed. Cir. 2018), the court explained that “[a] reference is considered publicly accessible ‘upon a satisfactory showing that such document has been disseminated or otherwise made available to the extent that persons interested and ordinarily skilled in the subject matter or art, exercising reasonable diligence, can locate it,’ quoting In re Wyer (C.C.P.A. 1981). The Jazz court continues, quoting Constant v. Advanced Micro-Devices (Fed. Cir. 1988): “‘If accessibility is proved, there is no requirement to show that particular members of the public actually received the information.’”

When the Internet became widespread in the 1990s, inventors and patent examiners faced rapidly expanding databases of online content. Since then, numerous court cases challenged the eligibility of online content as prior art, focusing on the “publicly accessible” prong. For example, in Voter Verified v. Premier Election Solutions (Fed. Cir. 2013), the court determined that being indexed by a search engine was “not a necessary condition” for an online article to be considered prior art. Rather, its appearance in an online periodical likely known to a PHOSITA was sufficient. On the other hand, in Acceleration Bay v. Activision Blizzard (Fed. Cir. 2018), an article was deemed to not be prior art because it was indexed by author and year online but could not be found by keyword search. In Samsung Electronics v. Infobridge (Fed. Cir. 2019), the cited reference was deemed inaccessible to the public because users needed to follow several steps with too many possible options to navigate a website that held the reference.

Scholars have called for refining the accessibility criterion. In a recent paper, “The Scope of the Prior Art,” Prof. Jay Thomas examines how legislative changes and court rulings have broadened what qualifies as prior art, often in ways contrary to innovation policy. He notes that obscure references from anywhere in the world are permitted to invalidate patents, unfairly holding inventors accountable for information that is legally but not practically accessible.

‍

AI-Generated Prior Art

The freedom of the Internet, combined with the increasing accessibility of AI models, means that anyone would be able to generate and publish unlimited prior art references. Indeed, All Prior Art and others are exploiting this possibility by using computers to automatically generate defensive references to prevent others from patenting, thus becoming “public domain trolls.” The problem, therefore, lies in the policy question of whether such content should be considered prior art for the purposes of the novelty requirement.

In their 2015 article, “Patents in an Era of Infinite Monkeys and Artificial Intelligence,” Ben Hattenbach and Joshua Glucoft point out that high-quality AI-generated content may be justified as being prior art, but low-quality AI-generated content is counterproductive. They write: “[p]ublishing masses of nonsense achieves the opposite of what these requirements seek to accomplish—it dilutes the set of actual public knowledge, burying genuinely useful information and leaving society worse off. Rewarding private companies for flooding the Internet with mostly inoperable and irrelevant publications would also impose considerable burdens [...]. Thus, there are good policy reasons to conclude that endless volumes of largely nonsensical, computer-generated text resulting from insufficiently guided processing should not be accorded prior art status.”

‍

AI-Assisted PHOSITAs

Though AI-generated content would increase the pool of potential prior art, this alone may not materially alter the status quo. Even disregarding AI-generated works, the number of “printed publications” already far surpasses the ability of a human being to parse them. Current case law precedent, as discussed above, would theoretically be sufficient to address AI-generated content. However, AI is a dual-edged sword: while growth in AI-generated content may render relevant prior art more difficult to find, ever-enhancing AI capabilities in search and reasoning may enable PHOSITAs to more easily access hyper-specific prior art. We argue that the latter effect may actually overshadow the former. 

Modern AI-enabled information synthesis engines (e.g., Deep Research) represent a quantum leap in technology over standard search engines. In light of such tools, the 2024 RFC also included questions regarding the impact of AI on a PHOSITA, such as:

• How, if at all, does the availability of AI as a tool affect the level of skill of a PHOSITA as AI becomes more prevalent?
• How, if at all, does the availability to a PHOSITA of AI as a tool impact whether something is well-known or common knowledge in the art?

If we assume the growing consensus that AI should be treated just as any other tool in an ordinary inventor’s toolbox, then it is possible that in the near future, AI-assisted PHOSITAs would have at their disposal a powerful “prior art Deep Researcher” that instantly locates every technically accessible piece of information relevant to a given query. Under such circumstances, a revision of the prior art eligibility criteria should be considered, perhaps inspired by historical patents of importation and unappreciated prior inventions.

‍

Patents of Importation and Unappreciated Prior Inventions

Historically, some patent systems around the world have given special treatment to works that are technically publicly available, but for practical purposes are inaccessible. As noted by Prof. Sean M. O’Connor in his 2015 article “The Lost ‘Art’ of the Patent System,” early patent systems in an era when communication and transportation were difficult “were designed to establish foreign arts into the domestic market… [t]he fact of local novelty meant an art established elsewhere was still an important kind of innovation in the domestic economy.” As explained by Paul Taylor in his 2024 article “Anti-Monology & Pro-Commerce: The Original Frontier Spirit of American Patent Law & Its Implications for Today,” such “patents of importation” were awarded not to the original inventor, but to the first to introduce an existing technology to another country; hence effectively making the innovation known to a new audience. In Edison Elec. Light v. Novelty Incandescent Lamp (3rd Cir. 1909), the court argued that an accidental or unappreciated prior invention gives “nothing to the world… and is thus entitled to no consideration” as prior art. Indeed, it was the lack of public benefit of the unappreciated prior invention that drove the court to reject any inherency argument. 

While we are not necessarily advocating the resurrection of a system that grants patents to non-inventors who introduce technology from elsewhere to a new audience, there is historical precedence for the idea that in order for information to be considered novelty-destroying prior art, it must not only be available but also must actually receive attention from members of the public.

‍

A Human Attention Criterion

In a world where increasingly powerful AI tools are readily available to analyze fast-growing oceans of AI-generated content, it is possible that with sufficiently advanced technology, all information would qualify under the Jazz standard and Thomas’s criterion. Thus, motivated by the policy considerations noted by Hattenbach and Glucoft, we propose abolishing the Jazz and Constant condition that no members of the public need to actually access the prior art reference if accessibility is proved. Instead, we propose that in a future era of AI-dominated information and AI-armed PHOSITAs, a printed publication prior art reference must have received actual human attention by the effective filing date of the patent application in question. This proposal is consistent with the comments offered by the American Bar Association Intellectual Property Law Section in response to the RFC: “we propose that art generated by AI/non-humans does not enter the public domain unless and until a human has appreciated/recognized the innovation disclosed in the document. (Emphasis added.)” In light of rapidly improving AI capabilities, we go further and suggest unifying the treatment of AI- and human-generated prior art under a common paradigm in which any printed publication must have received attention from humans (other than the author) in order to qualify as prior art.

Implementing such a “human attention” criterion would almost certainly require a legislative amendment to Title 35 of the U.S. Code § 102(a)(1), especially to qualify or replace the phrases “printed publications” and “otherwise available to the public.” Although the precise set of conditions for “sufficient human attention” would need to be refined, this criterion would discourage the use of AI to generate a database of prior art solely for the purpose of blocking others from patenting. Furthermore, it would not be difficult to produce digital papertrail evidence that a particular reference did indeed receive attention from human observers. Finally, human attention is naturally limited, which means that only the most significant or technically interesting references would qualify as prior art, sifting out content that gives “nothing to the world.”

‍

Looking Ahead

The problem of information overload, which is at the heart of the challenges posed by AI-generated prior art, has been a familiar challenge in patent law. We’ve argued that the key to dealing with the deluge of AI content lies not in fixating on the differences between AI and humans, but rather in clarifying existing standards in a utilitarian manner. By focusing on an element of human attention as a criterion for prior art eligibility, we can preemptively discourage the use of AI to abuse the patent system, thus fulfilling the goal of promoting human innovation.

‍