Industry Analysis: Artificial Intelligence in Patent Law

By Ella Dorfman '28

---

Definition and Context

Patent law falls under intellectual property (IP) law, which encompasses works of the mind, including copyrights, trademarks, and trade secrets. (1) Patent holders are granted exclusive rights for a period of time until their invention, concept, or process enters the public domain. (2) Utility patents include inventions and processes.

Patent attorneys and agents construct claims (language that defines the scope of a patent) through searching for prior art, negotiating with the United States Patent and Trademark Office (USPTO), and protecting IP from infringements. (3) Throughout these processes, they consult with the inventor and potential patent holder about the invention so that it can be described in accurate and specific detail. The patent examination process continues from when attorneys appeal rejections from USPTO examiners until the Patent Trial and Appeal Board provides approval. (4) At this time, inventors’ fees average into the thousands. (5)

The advancement of international technology and acceleration of the creation of novel inventions, in addition to the patent examination and prior art process (public information before a patent’s filing date), put strain on IP workers (attorneys and agents), thus calling for the automation of repetitive tasks. While patent attorneys and agents can conduct these tasks manually, artificial intelligence applications assist in claim construction, prior art searches, and predicting patentability in this field. In turn, they optimize the lengthy and costly patent examination process. (6) Natural language processing (NLP) can identify and compare long documents using technical context, machine learning can predict the patentability and potential infringements of a submission, and LLMs can write in patent jargon and classify a type of invention. (7) Human-in-the-loop systems integrate human reinforcement with LLMs to ensure the information LLMs synthesize is accurate and does not compromise legal and ethical standards. (8)

When deciding whether to approve a patent, examiners consider elements referenced as novelty, usefulness, non-obviousness, and proper subject matter. Usefulness refers to an invention or method possessing practical applications and benefits, and non-obviousness means that whatever is being patented would not be obvious to an ordinary person in the art. (9) As for proper subject matter, an abstract idea or discovery, such as a natural law of the universe like gravity, cannot be patented. (10)

Current State of the Concept

Natural language processing and machine learning have generative and analytical capabilities that have been overlooked because of the nature of patent law syntax and vocabulary. Patent law spans wide fields from engineering to life sciences, each requiring an attorney with a high level of specialization and education. (11)

Institutions such as the United States Patent and Trademark Office and law firms are encouraging employees to use AI with the goal of increasing patent examination speed and reducing costs. NLP, ML, and LLMs are aiding in the prior art searches, claim construction, and patentability evaluations. Last year, the American Bar Association released its formal ethics guide on how attorneys ought to use generative AI while protecting clients. (12) Major points included ensuring fees aligned with the use of AI and fairly representing the client to the attorney's best abilities. (13) This guide highlights that while AI can improve workflows, attorneys are responsible for verifying outputs and maintaining client confidentiality. 

The most mature aspect of AI-aided patent examination is in prior art search tools, which is one of the first steps in a patent application.

Prior Art Search

Traditionally, patent agents and attorneys have used manual and specific keyword queries to look for public information that would increase the likelihood of rejection and to evaluate the patentability of their submission. Software such as ESPACENET from the European Patent Office and TOTALPATENT by LexisNexus were the primary tools, but keywords using these software had issues. (14) Differing universal terms across disciplines and restrictions in length, nuance, and scope of words created limitations of keywords that restrict patent agents and attorneys from finding documents aligning with the exact invention they are attempting to patent.

To improve the prior art search for complex and accessible information, the USPTO has launched AI search engine tools such as Patents End-to-End (2022), the Similarity Search tool, and the Artificial Intelligence Search Autopilot Program (2025). (15) Patents End-to-End and the Similarity Search tool can help examiners find documents closely related to their submissions, and the Artificial Intelligence Search Autopilot Program allows users to prioritize ten prior art problems and weigh their claims against existing information. (16)

Claim Construction

         While NLP is currently used in prior art search engines, research into LLMs’ ability to analyze claims is being conducted. Knappich et al. accumulated a dataset of 14,000 patent claims from the United States with reasons for ambiguity and integrated them with LLMs in a system called Patent Definiteness Examination Corpus (PEDANTIC). (17) The goal of PEDANTIC was to evaluate the definiteness of patent claims, which is how specific and well-constructed claims are to an ordinary person in the arts. (18) While patent agents and attorneys can decipher if a claim is indefinite, they need to understand why it is, which is the main reasoning behind PEDANTIC. (19) LLMs analyzed office actions from the USPTO and provided annotations related to definiteness. Following, real human patent examiners provided feedback. (20) After this whole process, researchers used an LLM to judge both PEDANTIC and examiners’ answers. 

Researchers found that certain agents in the system underperformed in logistic regression classification performance (when AI calculates the probability of something belonging to a certain classification), but some were able to correctly note explanations for indefiniteness that human examiners identified. (21) PEDANTIC demonstrated where LLMs underperform in logistic regression, but where it equals human ability in patent reasoning. (22)

Detailed AI Examples in Patent Law

Patentability Characteristic Analysis: Large Language Models

Besides assisting with prior art searches and claim construction, the core principles of patentability, novelty, utility, and non-obviousness remain clear. Without these characteristics, an invention fails to be patentable. 

In research by Ikoma and Mitamura, LLMs examined prior art and claims to conclude patent novelty, mirroring a real patent examiner. (23) The classification models fell short, while generative models did better at examining non-novel texts than nuanced texts. (24) Both models’ output improved after being prompted with explanations, demonstrating that LLMs can predict the novelty of inventions but lack the human nuance and reasoning that experienced patent examiners possess.

Human Reinforcement of Large Language Models

As can be seen, human feedback increases the effectiveness of LLMs. Not only does the output of AI systems improve after reinforcement, but having a human perspective ensures that legal ethics are not compromised in the process. Bui investigated reinforcement learning from human feedback (RLHF), where LLMs were trained to align their outputs with human expectations and legal standards. (25) Researchers applied RLHF to train LLMs in patent tasks while supplying a dataset with issued and pre-issued patents, and the results showed that LLMs could create more patentable claims with the RLHF. (26) RLHF continues to be necessary when training artificial intelligence to best serve those in IP.

Future Evolution of AI in Patent Law

AI’s presence in patent law will evolve as both technological and legal frameworks adapt to each other. NLP, LLMs, and RLHF are already reshaping the patent examination process. However, the next decade will see AI become a fully integrated assistant throughout the entire patent lifecycle. Instead of remaining in the drafting and examination stages as discussed above, AI may be involved in concept generation (ideating novel inventions or technical concepts). Ren et al. (2025) proposes a LLM that can parse and understand scientific documents so that it can generate initial concepts with the same expertise as a person skilled in the arts. (27) This means that if prompted to produce an idea for an invention, the LLM will generate claims in patent language and understand the full science of that invention. (28) In the study, Ren et al. called this LLM PatentGPT, which was successful on patent benchmark tests and showed potential for AI’s further application in the field. (29)

As AI becomes more deeply integrated into the creative aspects of inventing, it will raise regulatory questions. When AI contributes significantly to a unique concept, the question of inventorship attribution comes to mind. Currently, U.S. law is shifting to question AI’s rights as an inventor, which will change the IP landscape. (30) In 2022, in Thaler v. Vidal, the Federal Circuit held that an inventor must be human and if AI aided a human in the creation of their invention, the human must make a large contribution to satisfy inventorship rules. (31)

The future of AI in patent law is not just about optimizing task speed; it is also about questioning the fundamental definition of an inventor and invention. As NLP and LLMs become collaborators, the legal system must refine disclosure standards and laws around how AI contributions are credited and owned. There must be a balance in taking advantage of AI’s analytical and generative abilities while maintaining patent law and order.

Endnotes

1. Sarah Burnstein, Sarah Wasserman Rajec, and Andres Sawicki, Patent Law (Self-published, 2021).

2. Burnstein, Wasserman Racjec, and Sawicki, Patent Law, 17.

3. Burnstein, Wasserman Racjec, and Sawicki, Patent Law, 33.

4. Valentin Knappich, Annemarie Friedrich, Anna Hätty, and Simon Razniewski, “PEDANTIC: A Dataset for the Automatic Examination of Definiteness in Patent Claims,” arXiv, https://arxiv.org/abs/2505.21342.

5. United States Patent and Trademark Office, “USPTO Fee Schedule,” 2019, https://www.uspto.gov/learning-and-resources/fees-and-payment/uspto-fee-schedule.

6. Luong Vu Bui, “Advancing Patent Law with Generative AI: Human-in-the-Loop Systems for AI-Assisted Drafting, Prior Art Search, and Multimodal IP Protection,” World Patent Information 80 (2025): 102341. https://doi.org/10.1016/j.wpi.2025.102341.

7. Bui, “Advancing Patent Law with Generative AI: Human-in-the-Loop Systems for AI-Assisted Drafting, Prior Art Search, and Multimodal IP Protection,” par. 2.

8. American Bar Association, “ABA Issues First Ethics Guidance on a Lawyer’s Use of AI Tools,” 2024, https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/.

9. Burnstein, Wasserman Racjec, and Sawicki, Patent Law, 76.

10. Burnstein, Wasserman Racjec, and Sawicki, Patent Law, 84.

11. Lekang Jiang, and Stephan M. Goetz, “Natural Language Processing in the Patent Domain: A Survey,” Artificial Intelligence Review 58, no. 7 (2025). https://doi.org/10.1007/s10462-025-11168-z. 

12. American Bar Association, “ABA Issues First Ethics Guidance on a Lawyer’s Use of AI Tools,” par. 1.

13. American Bar Association, “ABA Issues First Ethics Guidance on a Lawyer’s Use of AI Tools,” par. 12.

14. Lea Helmers, Franziska Horn, Franziska Biegler, Tim Oppermann, and Klaus-Robert Müller, “Automating the Search for a Patent’s Prior Art with a Full-Text Similarity Search.” arXiv (2019). https://arxiv.org/abs/1901.03136. 

15. United States Patent and Trademark Office, “USPTO Launches New AI Pilot for Pre-Examination Utility Application Search,” 2025, https://www.uspto.gov/about-us/news-updates/uspto-launches-new-ai-pilot-pre-examination-utility-application-search. 

16. United States Patent and Trademark Office, “USPTO Launches New AI Pilot for Pre-Examination Utility Application Search,” par. 3.

17. Knappich et al.,“PEDANTIC: A Dataset for the Automatic Examination of Definiteness in Patent Claims,” par. 3.

18. Burnstein, Wasserman Racjec, and Sawicki, Patent Law, 226.

19. Knappich et al.,“PEDANTIC: A Dataset for the Automatic Examination of Definiteness in Patent Claims,” par. 3.

20. Knappich et al.,“PEDANTIC: A Dataset for the Automatic Examination of Definiteness in Patent Claims,” par. 13.

21. Knappich et al.,“PEDANTIC: A Dataset for the Automatic Examination of Definiteness in Patent Claims,” par. 23.

22. Knappich et al.,“PEDANTIC: A Dataset for the Automatic Examination of Definiteness in Patent Claims,” par. 28.

23. Hayato Ikoma and Teruko Mitamura, “Can AI Examine Novelty of Patents? Novelty Evaluation Based on the Correspondence Between Patent Claim and Prior Art,” arXiv (2025). https://arxiv.org/abs/2502.06316.

24. Ikoma and Mitamura,“Can AI Examine Novelty of Patents? Novelty Evaluation Based on the Correspondence Between Patent Claim and Prior Art,” par. 22.

25. Bui, “Advancing Patent Law with Generative AI: Human-in-the-Loop Systems for AI-Assisted Drafting, Prior Art Search, and Multimodal IP Protection,” par. 4.

26. Bui, “Advancing Patent Law with Generative AI: Human-in-the-Loop Systems for AI-Assisted Drafting, Prior Art Search, and Multimodal IP Protection,” par. 6.

27. Runtao Ren, Jian Ma, and Jianxi Luo, “Large Language Model for Patent Concept Generation,” Advanced Engineering Informatics 65 (2025): 103301. https://doi.org/10.1016/j.aei.2025.103301.

28. Ren, Ma, and Luo, “Large Language Model for Patent Concept Generation,” par. 5.

29. Ren, Ma, and Luo, “Large Language Model for Patent Concept Generation,” par. 6.

30. Hickey, Kevin, and Christopher Zirpoli, “Artificial Intelligence and Patent Law.” Congressional Research Service, 2024, https://www.congress.gov/crs-product/LSB11251. 

31. Hickey, Kevin, and Zirpoli, “Artificial Intelligence and Patent Law,” par. 7.