Applying Generative AI Tools to Patent Law Practice

MBHB Summer technical advisor intern Grace Clopton co-authored this article.

Generative AI tools are poised to transform legal practice and offer unique challenges and opportunities, particularly in patent law. Modern large language model (LLM)-based tools enable faster, more efficient document drafting, research, and review. These AI systems, trained on vast amounts of legal texts and general language, can summarize case law, identify relevant precedents, extract key clauses from lengthy documents, and generate portions of legal documents. However, the integration of generative AI tools into actual patent law firm operations can be difficult to navigate, as ethical, effective use of AI in legal work requires technical competence and proper oversight, while avoiding the inadvertent disclosure of client-confidential information. This article explores some best practices for the implementation and use of generative AI tools in patent law practice including on-prem sandbox installation, proper oversight and competence, and practical guidance regarding prompt engineering and few-shot prompting to reduce the risk of hallucination and irrelevant output from generative AI tools.

On-Premise LLM Installation

Confidentiality is foundational to patent law, particularly because premature disclosure of an invention can undermine potential patent rights. Under 35 U.S.C. § 102, inventors in the United States have a one-year grace period following public disclosure to file a patent application, but most foreign jurisdictions provide no such provision. For this reason, any exposure of an invention prior to filing, whether intentional or not, can destroy novelty and bar patentability. Historically, law firms have safeguarded against this risk by maintaining tight controls over invention-related information. However, the integration of LLMs and other GenAI tools into patent practice introduces new vectors for unintentional public disclosure. The cure for preventing such disclosures could include on-premises, sandboxed hardware installations utilizing open-weight and/or open-source LLMs such as Meta’s Llama 3 and Mistral. These setups allow for the isolation of client information to private local or cloud servers and the customization of workspaces in which to use the information. As patent attorneys and IT staff come up to speed on best practices and workflows for secure LLM use, initial prototyping and testing can be performed in house without fear of inadvertent disclosure.

Confirming Competency Among Legal Experts

While confidentiality is paramount, a proper technical understanding of GenAI tools is also necessary to incorporate LLM output into the patent law workflow and accurately discern hallucinations. One way to gain the requisite LLM expertise is to compel practitioners to enroll in relevant continuing legal education (CLE) courses. In January 2017, Florida became the first state to require technical training as a CLE requirement, mandating that three out of 30 credit hours in each three-year period must be for approved “technology courses.” While no states explicitly require AI-specific CLE training, this seems like a helpful step as patent law firms actively adopt GenAI tools. Apart from earning CLE, attorneys can avail themselves of a huge amount of free information about LLMs and other GenAI tools via YouTube [1], among other websites. Additionally, there can be no substitute for actual hands-on use of ChatGPT and other LLMs to build user experience and intuition about GenAI’s capabilities and limitations.

Ensuring Human Oversight

Active and critical human oversight is essential when prompting LLMs and incorporating the models’ output into legal work product. Although LLM output can appear polished, it can also include subtle (and/or stark) legal and technical errors. Under 37 C.F.R. § 11.18, all submissions to the USPTO must be made with reasonable inquiry into their accuracy. This duty extends to work product generated by AI as well as by human practitioners. Inadequate review could result in sanctions or could jeopardize the patent disclosure or record.

Accordingly, human-in-the-loop, deliberate, line-by-line oversight is needed to review work product from junior associates as well as to review output from GenAI tools.

Proper Prompt Engineering

To circumvent some, but not all, hallucination issues, creative prompt engineering can be applied to LLM operations. Specifically, by crafting input in a well-organized and intentional manner, the LLM will typically provide a more accurate and applicable output. In practice, prompt engineering adds context and focused objectives for the assigned work, and helps define an LLM “persona,” output style, and format. For example, a well-engineered prompt might look like the following: “You are an experienced patent practitioner tasked with developing a strategy and drafting an office action response using [attached template], based on [pending claims] and [attached prior art references].”

Proper prompt writing is an iterative art, and the LLM can learn from various successes and failures, becoming increasingly refined to the objective with more contextual information. OpenAI’s prompt engineering guide for ChatGPT suggests providing, “explicit instructions,” “the objective of the work,” and “’you are’ statements” to accomplish these tasks. Google’s prompt engineering guide for Gemini recommends specifying length constraints (e.g., “give me two paragraphs about…”), providing instructions for response format (e.g., “provide the data in a tabular format”), and utilizing response aggregation (e.g., by combining responses from different agents or expert analyses). Response aggregation is useful when you want the LLM to perform parallel tasks on different portions of data, such as when assigning multiple agents to produce different types of prior art (e.g., issued patents, printed publications, foreign references, and journal articles) for a target patent, as just one example.

Zero-Shot and Few-Shot Inference

Another recommended way to improve GenAI output is to utilize few-shot inference techniques over zero-shot inference attempts. An example of zero-shot inference is a simple input query that doesn’t include examples of the desired output (e.g., “Summarize the following text in one paragraph. [Text]”). In contrast, few-shot inputs provide a desired task along with examples of how an ideal output might be formatted, as well as the length of the output and its contextual content. For example, a few-shot input might be something like: “Summarize the following text in one paragraph. [Text] An ideal output will include a topic sentence introducing the main idea of the text; two to three supporting sentences that provide evidence, explanations, and/or examples; and a concluding sentence that succinctly summarizes the text’s main point.” In patent workflows, few-shot prompting tends to provide better accuracy for tasks like claim drafting or office action response arguments where structure and precise language matter. Meanwhile, zero-shot attempts can work well for simpler tasks like classification or summarization when speed is prioritized over form.

Conclusion

When incorporating generative AI tools into patent law practice, attorneys must prioritize maintaining confidential client information, competence with the AI tools, and human-in-the-loop oversight with respect to both practitioners and the GenAI outputs themselves. As AI technology rapidly evolves, patent law firms would do well to encourage sandboxed experimentation with LLMs – under proper oversight – enabling practitioners to streamline workflows, improve their work product quality, and make more efficient use of their billable time.

[1] See, e.g., https://youtu.be/zizonToFXDs?si=IziCeZlyVEVfAyER (Introduction to Large Language Models)