In a paper published in Nature Biotechnology, three Stanford University-affiliated co-authors, including Stanford Law School Professor Lisa Larrimore Ouellette, tackle an issue of growing importance at the intersection of law and technology: How the use of artificial intelligence to codify and communicate knowledge may influence the usefulness of the technical literature for subsequent researchers and innovators. In particular, they studied how the growing use of AI in the patent-drafting process is exacerbating existing challenges around patent law’s disclosure requirements.

Patent law requires that inventors not only describe their inventions, but do so with enough clarity that someone skilled in the relevant field can replicate the work. This disclosure requirement helps disseminate knowledge about new inventions, and is an important part of what society gets in return for granting inventors exclusive rights to their inventions. Disclosure also ensures that a patent’s scope is limited to what the inventor contributed, according to the authors of How Will AI Affect Patent Disclosures? However, enforcing the disclosure standards has historically been inconsistent, they write. Patent examiners, working under tight time constraints and with sometimes limited training, often prioritize evaluating novelty and obviousness over assessing disclosure quality.

“AI will amplify errors in both improperly granting poorly disclosed patents and improperly denying patents on the basis of earlier AI-generated references that do not actually disclose those inventions,” write co-authors Lisa Ouellette, the Deane F. Johnson Professor of Law; Victoria Fang, JD ’24, a former United States Patent and Trademark patent examiner who collaborated on the study while a student at Stanford Law; and Nicholas T. Ouellette, a Stanford University professor of civil and environmental engineering.

But the rise of AI patent drafting also provides an opportunity to reform the patent examination process and promote innovation, according to the authors.

“The disruption to the patent drafting and review process created by AI gives us an opportunity to improve disclosure across the patent system,” said Lisa Ouellette, the author of a leading patent law casebook and a large body of scholarship that focuses on empirical and theoretical problems in intellectual property and innovation law. “Increasing enforcement of existing disclosure rules and heightening requirements to limit granting patents to those who have performed real inventive work would better equip patent offices to do their work and would improve the patent system overall.”

The USPTO recently issued guidance requiring human review of AI-drafted patent applications, “but this will not be enough,” say the paper’s authors, who undertook a study, described in their article, to evaluate whether current AI tools are capable of generating plausible patent specifications that are unlikely to be rejected by a patent examiner for failure to meet the disclosure requirements.

The authors’ recommendations include raising disclosure standards to require inventors to implement at least some version of their invention before filing a patent application (or meet stricter disclosure requirements for unproven technologies). They also call for improved training of patent examiners, including providing them with examples of inadequate disclosures and emphasizing the importance of technical accuracy.

Putting AI tools to the test

To explore the capabilities of AI in drafting patent disclosures, the researchers designed a study evaluating two patent-specific AI systems (Edge and Vaero) and the general-purpose ChatGPT-4o. They tested the tools with four technologies related to fluid dynamics: two well-documented, existing technologies likely included in the AI’s training data, and two hypothetical technologies.

For the existing technologies, the AI tools produced patent specifications that generally included sufficient detail for a skilled researcher to replicate the inventions. However, the outputs also contained inaccuracies or misleading statements alongside correct information. This suggests that while AI can assist in describing established technologies, its outputs require careful human oversight to ensure reliability.

When the researchers turned to hypothetical technologies, asking the AI platforms how to implement unproven concepts, none of the tools generated workable solutions or accurate technical descriptions for these unsolved problems. But rather than acknowledging their shortcomings, the AI platforms created seemingly credible, but ultimately incorrect or speculative, specifications. 

“Interestingly, the AI-generated specifications for both existing and hypothetical technologies were often sufficient to satisfy the current level of scrutiny provided by patent examiners even if they did not meet the scientific standards,” said co-author Fang. “This points to a mismatch between how patent law standards are applied and the practical utility of patent disclosures.”

“The outputs of AI tools can look convincing at first glance, but can also have some pretty serious flaws. Without careful review, they risk diluting the value of patent disclosures as a source of technical knowledge,” said co-author Nicholas Ouellette. “And if examiners incorrectly grant patents that don’t actually explain how to implement an invention, those patents stand as a legal barrier to researchers who later figure out how to make the invention work.

Toward a better system

“The problem isn’t with these AI patent-drafting tools themselves,” Lisa Ouellette said. “In fact, they show enormous promise for reducing the costs of patenting and inequalities in who has access to the patent system.”

Instead, the authors think AI tools reveal more fundamental problems with patent law’s disclosure standards and how those standards are enforced. They propose several strategies to improve disclosure quality and mitigate the risks of AI-generated patents, including: 

• Requiring inventors to implement at least some version of their invention before filing a patent or to meet stricter disclosure requirements for unproven technologies; 
• Providing patent examiners with better training and resources to evaluate disclosures effectively; 
• Bringing in external experts to review disclosures and enhance the rigor of patent examinations, particularly for cutting-edge or highly technical inventions;
• Using AI itself to improving patent examination by identifying potentially problematic disclosures for closer scrutiny; and
• Extending the window for challenging a patent’s disclosure beyond the current nine-month limit to provide a safety net for addressing issues missed during initial examination.

“By refining disclosure standards, reconsidering examination practice, and using technology wisely, we can ensure that patents are granted only for what the inventor actually invented and disclosed,” Lisa Ouellette said.