Emilio Tovilla, Contributing Member 2023-2024
Intellectual Property and Computer Law Journal
I. Introduction
Artificial Intelligence (AI) has rapidly developed into a prevalent tool used in various industries, ranging from academia to finance to healthcare.[1] AI programs have also started to integrate in the legal field, leaving mixed impressions. On one hand, AI is used to enhance common legal databases such as Westlaw and Lexis.[2] However, not all impressions have been positive considering a New York attorney was sanctioned in 2023 for using fictitious cases generated by ChatGPT, an AI chatbot.[3] One issue regarding the intersection of law and AI not thoroughly addressed is whether AI can be used to analyze patents.
When a patent application is filed, an invention must clear multiple requirements.[4] The analysis of some requirements must be done through the lens of the “person having ordinary skill in the art,” better known as the PHOSITA.[5] The nature of the PHOSITA can be tricky because certain characteristics that determine the level of the person with ordinary skill may change depending on the claims and field surrounding the invention sought to be patented.[6] Integration of AI could potentially create a flexible tool that analyzes patent applications, however, because many programs are recently developed, limitations are bound to exist.
This article will discuss whether AI, specifically language models, can be used as tools to represent the PHOSITA in the analysis of a patent. Part II will provide background on what the standard for the PHOSITA is and the process of patent claim analysis, as well as how language models work. Part III discusses whether AI language models could fit within the scope of the PHOSITA and what limitations exist. Finally, Part IV concludes by making suggestions on how AI could be used as an effective tool for patent analysis.
II. Background
Under 35 U.S.C. §101, a claimed invention must satisfy four requirements for patent issuance by the United States Patent and Trademark Office (USPTO): (1) subject matter eligibility, (2) novelty, (3) utility, and (4) non-obviousness.[7] The PHOSITA is relevant in the analysis of the novelty and non-obviousness requirements. Both elements help establish whether an invention is new and distinguishable from prior art; however, each is regulated by different sections. Typically, prior art is understood as the legally available technology and information that the claimed invention will be compared with to determine whether the invention is patentable.[8] Use of AI language models could potentially perform the analysis of either element; however, because the analysis is difficult, limitations in the technology must be considered.
PHOSITA
The PHOSITA is a hypothetical person statutorily mandated by 35 U.S.C. §103. It is assumed to have the necessary knowledge to understand the field of technology of an invention and recognize prior art.[9] There is big emphasis on the “ordinary skill” component of the PHOSITA, requiring that the hypothetical person have more than common knowledge but not reach an extraordinary level of expertise.[10] This is because if the PHOSITA had a low understanding of technology, most inventions would be non-obvious. However, if the PHOSITA was an expert with a remarkable level of understanding, most would be considered obvious. It is important to recognize that the PHOSITA is never the inventor of the claimed patent. This is because the question always considers whether the invention was obvious to a person ordinarily skilled in the art, not whether it was obvious to the inventor.[11]
One of the most difficult concepts for patent attorneys to grasp is that the PHOSITA is not the representation of an actual person.[12] Rather, the PHOSITA is an objective standard, such as the reasonable person in tort law. This is a standard of patentability created by Congress and is not determined by the USPTO.[13] Depending on the scope and type of invention, the PHOSITA changes. This is because the PHOSITA’s skill level and understanding varies depending on how advanced the technology is.[14] For this reason, identifying the proper PHOSITA should include relevant factors and knowledge that appropriately correspond with the field and technology of the invention.[15]
Novelty
For an invention to be patented it must be considered new.[16] Novelty is established under 35 U.S.C. §102, asserting that an invention must avoid anticipation to be patentable.[17] Multiple provisions are indicated under 35 U.S.C. §102 and if one or more are triggered, an invention could be deemed anticipated and therefore, unoriginal. Under §102(a), provisions include prior patents on the invention, printed publications, or the invention being placed in public use, on sale, or available in some capacity to the public before the effective filing date.[18]
There is an exception under §102(b) that asserts if disclosure of a claimed invention is made by the inventor within one year the patent application is filed, then it shall not be considered prior art for the purpose of anticipation.[19] Should a public disclosure that qualifies as prior art be made, a detailed analysis must be applied to the facts and law.[20] This analysis is where the PHOSITA is crucial because it is through the lens of the proper PHOSITA that prior art and the patent application is examined.
Non-obviousness
Whether a PHOSITA would find a patent obvious is a question of law based on underlying findings of fact.[21] Under 35 U.S.C. §103, if the comparison between prior art and a claimed invention would result in the subject matter as a whole to be obvious to the PHOSITA at the time the invention was made, an invention is no longer patentable.[22] The statute further requires that the claimed invention be a non-obvious improvement of prior art.[23] When conducting this extremely difficult analysis, the USPTO examiner inspects previous patent documents, aiming to find any patents and published patent applications that are like the claimed invention.[24] Whether the examiner finds a single document with all the same claims, or combines two or more prior patents to find all claims, the invention could be deemed obvious.[25]
To clarify the overall standard for non-obviousness, the Supreme Court set out four factors known as the Graham test.[26] These factors are the underlying factual determinations to every non-obvious analysis and include (1) level of ordinary skill in the art (2) scope and content of the prior art (3) difference between the claimed invention and the prior art; and (4) secondary considerations.[27] Once the factual determinations are made, it is decided whether based on the evidence a PHOSITA would have found the invention obvious at the time of creation.[28]
Language Models
AI language models can fulfill a variety of tasks, including the production of content, question answering, and text summarization to name a few.[29] Generally, language models are created to analyze, understand, and generate human language and use information from a large data base to create content.[30] While language models have been fascinating, the early phase of this technology naturally brings defects. Some limitations of language models include lack of common-sense knowledge, inability to understanding abstract concepts, and inability to make inferences from incomplete information.[31]
In November of 2022, OpenAI released the widely popular language model known as ChatGPT.[32] This chatbot technology has been impressive by showing how AI can generate answers, whether in a few sentences or entire essays, to a wide array of questions present by the user. ChatGPT is trained on a large dataset of text and then uses patterns learned from the dataset in creating responses to questions.[33] Use of the chatbot in the legal field has been done for a variety of common tasks like legal research and document drafting.[34] However, the outcomes have not all been positive. In June 2023, a New York attorney was sanctioned after submitting a brief that relied on ChatGPT.[35] The submitted brief cited to multiple fake cases generated by ChatGPT, which the attorney believed to be real case law.[36]
In addition to generating false information, a limitation of ChatGPT is that the technology may not have the capacity to answer certain questions because it was trained with information from 2020 and before.[37] This leaves nearly three years’ worth of information that could make a huge difference in accuracy while performing research.
III. Discussion
Considering AI language models are trained with high volumes of information and generate responses based on the same information, the technology has potential to become a tool in patent application analysis. Review of prior art is done through the perspective of the PHOSITA to determine whether any similarities exist with the claimed invention. Since this is an objective standard that uses factual information, a language model trained with information pertaining to fields within patent statutory subject matter is a possibility in the legal field. Of course, limitations must also be considered.
Language Models in the scope of the PHOSITA
AI language models are created to generate, analyze, and understand human language.[38] During its creation, a language model is trained with high quantities of information that create patterns, which are then used to produce content and answer questions.[39] In the analysis of a patent application, to determine if an invention is new and non-obvious, prior art is reviewed and compared to the claimed invention through the perspective of the hypothetical PHOSITA. The PHOSITA is not intended to represent a real person but creates a standard of what a person with ordinary skill in the respective field and technology of the patent would know.[40] Based on the mechanisms of how language models are created, it seems like a tool that has potential to fit within the scope of the PHOSITA for the purposes of analyzing novelty and non-obviousness.
The best possible way for AI language models to be used effectively in the analysis of patents is by creating language models geared specifically for those purposes. Taking ChatGPT as an example, a patent law version of the chatbot could be created with information pertaining to all fields that fit within the statutory subject matter element of 35 U.S.C. §101. Statutory subject matter refers to the types of inventions that are patentable, which include those that fall within the categories of processes, machines, articles of manufacture, and compositions of matter.[41] A ChatGPT trained with such information could be asked to compare a claimed invention to prior art and determine if there are any similarities, resulting in anticipation or obviousness. AI’s inability to formulate its own thoughts or bias could be beneficial because it allows for an objective approach, which is required for the PHOSITA. A patent law ChatGPT could also allow a user to research the field of the claimed invention and find all the available pieces of prior art to make a full and accurate analysis. This being said, there are limitations in the use of AI language models.
Limitations
One limitation of language models is the inconsistency of generated answers. Language models such as ChatGPT are still in earlier phases and will develop into more reliable technology as time progresses. Errors made by ChatGPT have misled practicing attorneys by making up case law that was ultimately used in submitted briefs to the court.[42] While it is imperative that attorneys always check their work for error, the possibility of incorrect information and need to double check everything when using language models could require more work rather than reducing it.
Another limitation for language models is the inability to make inferences based on incomplete information.[43] Language models are unable to think for themselves and can only generate answers with the information provided to it during its creation. A vast number of fields are constantly evolving, and innovation consistently changes the course of what is current. In fact, one of the purposes of patents is to incentivize the conception of inventions to further the development of society. AI language models would require constant updates to ensure that prior art and information is as current as possible. Under 35 U.S.C. §102(b) it could be argued that because analysis of patents requires all prior art to be a year or prior from the patent filing date, the constant updates would not be necessary. However, updates would be inevitable at a fast rate even under 35 U.S.C. §102(b). This would ensure all patent applications were analyzed as accurately and as fairly as possible, but at an inconvenience.
Finally, a third limitation is the language model’s inability to understand abstract concepts.[44] As outlined in the Graham test for the analysis of non-obviousness, one of the factors is to determine the level of ordinary skill in the art.[45] This factor helps determine the level of skill of the PHOSITA. Because comprehension of technology varies, with some requiring only a bachelor’s degree and other a doctoral degree, determination of ordinary level is quite challenging, even for patent experts. To determine this, various aspects, such as the educational level of active workers in the field and the sophistication of technology used in the invention are considered.[46] Due to these requirements, it would be difficult to determine whether language models would have the capacity to answer such crucial questions because of the hypothetical nature. Additionally, the PHOSITA must be a person with ordinary skill, not extraordinary levels of expertise.[47] Having immense information on all types of technology yet being unable to understand abstract concepts, like what a person with ordinary skill is, might make it impossible for AI language models to correctly identify the proper PHOSITA.
IV. Conclusion
The use of AI in law is in a gray area. However, AI programs are being developed for the purpose of fulfilling legal work, such as the chatbot coined Lawadroid copilot.[48] This AI tool is programed to perform tasks like legal research, document drafting, and even brainstorming blog ideas.[49] While the technology shows promise, it has the potential for defects considering other chatbots like ChatGPT ran into issues. These AI technologies will improve over time with trial and error. The question then is whether the legal community should risk errors to reap benefits at a future time. These errors clearly raise concern over ethics and ensuring that legal proceedings stay true to the rules and fair legal system.
With language model technology in its current state, defects can lead to distribution of misleading information with the potential for major repercussions in the legal field. A suggestion for the integration of AI is to have creators of language models incorporate attorneys and legal scholars in the experimentation stage of the technology. This will maximize the success of law focused AI programs because legal professionals can best explain what is necessary in the AI work product to abide with expectations and rules in the field. Additionally, creators of language models should develop a method that allows the programs to comprehend abstract concepts. Many concepts of law, including the PHOSITA, are comprised of abstract ideas and therefore should be understood by AI for efficient use by attorneys.
The analysis of patents is a complicated area of law as is. While at the moment, the technology does not possess all the qualities necessary for reliable use in patents, AI language models do show potential to make the process of patent claim analysis simpler. This could be revolutionary and potentially pave a path for other areas of law to benefit from AI.
[1] Bernard Marr, 15 Amazing Real-World Applications Of AI Everyone Should Know About, Forbes (May 10, 2023, 2:51 am), https://www.forbes.com/sites/bernardmarr/2023/05/10/15-amazing-real-world-applications-of-ai-everyone-should-know-about/?sh=7f69e86685e8.
[2] Paul Riermaier, ChatGPT and the Law, University of Pennsylvania Law (Feb. 6, 2023), https://www.law.upenn.edu/live/news/15538-chatgpt-and-the-law https://www.law.upenn.edu/live/news/15538-chatgpt-and-the-law.
[3] Sara Merken, New York lawyers sanctioned for using fake ChatGPT cases in legal brief, Reuters (June 26, 2023, 4:28 am), https://www.reuters.com/legal/new-york-lawyers-sanctioned-using-fake-chatgpt-cases-legal-brief-2023-06-22/.
[4] 35 U.S.C. §101.
[5] Gordon K. Hill et al., Avoiding Confusion of the PHOSITA, Aipla, (Oct. 6, 2023) https://www.aipla.org/list/innovate-articles/avoiding-confusion-of-phosita.
[6] Daiichi Sankyo Co., Ltd. v. Apotex, Inc., 501 F.3d 1254, 1257 (Fed. Cir. 2007).
[7] 35 USC §101: Statutory Requirements and Four Categories of Invention, United States Patent and Trademark Office, (Aug. 2015) https://www.uspto.gov/sites/default/files/101_step1_refresher.pdf.
[8] Hill, supra.
[9] Id.
[10] Id.
[11] Id.
[12] Id.
[13] Id.
[14] Id.
[15] Id.
[16] Janice M. Mueller, Patent Law 241 (Rachel E. Barkow et al eds., 6th ed. 2020)
[17] Id.
[18] 35 U.S.C. §102(a).
[19] 35 U.S.C. §102(b).
[20] Patent Requirements, Bitlaw (October 1, 2023) https://www.bitlaw.com/patent/requirements.html#:~:text=Novelty%20(Newness)%20Requirement,the%20inventon%20have%20been%20made.
[21] Smithkline Beecham Corp. v. Apotex Corp. 2005 U.S. LEXIS 5999, at *26-27 (E.D. Pa. Mar. 31, 2005).
[22] In re NuVasive, Inc., No. 2015-1670, 2016 U.S. App. LEXIS 21748, at *1381 (Fed. Cir. Dec. 7, 2016).
[23] Patent Requirements, supra.
[24] Id.
[25] Id.
[26] Graham v. John Deere Co., No. 11, 1966 U.S. LEXIS 2908, at *21-29 (U.S. Feb. 21, 1966).
[27] Id.
[28] Smithkline, 2005 U.S. LEXIS 5999, at *27.
[29] Language Models, Explained: How GPT and Other Models Work, Altexsoft, (Jan. 18, 2023) https://www.altexsoft.com/blog/language-models-gpt/.
[30] Id.
[31] Id.
[32] Andrew Perlman, The Implications of ChatGPT for Legal Services and Society, Harvard Law School, (Mar./April 2023) https://clp.law.harvard.edu/knowledge-hub/magazine/issues/generative-ai-in-the-legal-profession/the-implications-of-chatgpt-for-legal-services-and-society/.
[33] Reirmeier, supra.
[34] Id.
[35] Josh Russell, Sanctions ordered for lawyers who relied on ChatGPT artificial intelligence to prepare court brief, Courthouse News, (June 22, 2023)https://www.courthousenews.com/sanctions-ordered-for-lawyers-who-relied-on-chatgpt-artificial-intelligence-to-prepare-court-brief/.
[36] Id.
[37] Language Models, Explained: How GPT and Other Models Work, supra.
[38] Id.
[39] Id.
[40] Id.
[41] Patent Requirements, supra.
[42] Russell, supra.
[43] Language Models, Explained: How GPT and Other Models Work, supra.
[44] Id.
[45] Smithkline, 2005 U.S. LEXIS 5999, at *27.
[46] Envtl. Designs, Ltd. v. Union Oil Co., 713 F.2d 693, 696 (Fed. Cir. 1983).
[47] Hill, supra.
[48] Bob Ambrogi, New GPT-Based Chat App from LawDroid Is A Lawyer’s ‘Copilot’ for Research, Drafting, Brainstorming and More, LawSites, (Jan. 25, 2023) https://www.lawnext.com/2023/01/new-gpt-based-chat-app-from-lawdroid-is-a-lawyers-copilot-for-research-drafting-brainstorming-and-more.html.
[49] Id.
The University of Cincinnati Intellectual Property and Computer Law Journal 
Leave a comment