Listen to this post

This article was first published by ALM / Law.com in The Intellectual Property Strategist

I. Introduction

Organizations across all industries are adopting generative AI systems as critical components of their business strategy. These systems often take the form of hosted or on-premises pretrained large language models (LLMs), both proprietary and open source. Organizations acquiring access to pretrained LLMs from a small but growing list of providers can apply various customization techniques. Once customized, LLM usage by an organization can potentially result in an output that constitutes an invention like those on which thousands of US patents are granted every year. As just some examples, a suitably customized LLM could generate a technique to determine a navigation plan consistent with an ODD associated with an autonomous vehicle, an algorithm to predict disease onset based on clinical and environmental factors, or computer code to detect malware by overcoming dynamic obfuscation attempts.

Typical license provisions vest ownership of intellectual property rights in such output in the organization as user of the LLM. A statutory predicate to the contractual outcome regarding ownership of patent rights is the requirement of a sufficient contribution by a natural person in the effort that yielded the output. The issues implicated by this requirement are one development among more to come as patent law and policy try to catch up to proliferating AI technology.

II. Human Inventorship

Pursuant to the “Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” (October 30, 2023), the US Patent Office recently provided guidance regarding inventorship requirements for AI-assisted inventions. While the US Patent Office guidance by its own terms does not have the effect of law, it nonetheless sets forth current agency policy regarding interpretation of legal requirements governing inventorship. The US Patent Office guidance relies on Federal Circuit reading of the patent statute to remind that conception as the “touchstone” of inventorship requires formation in the “mind of the inventor” of a definite, permanent idea of the complete, operative invention. Under the Pannu factors and exercise of some discretion, the US Patent Office guidance qualifies a natural person as inventor through a “significant contribution” to conception. Without human inventorship so characterized, the US Patent Office will refuse the invention – at least for now.

The US Patent Office guidance provides an inclusive set of “guiding principles” involving AI-assisted inventions along an apparent continuum from absent to present human inventorship. Near one end, the US Patent Office guidance indicates that mere ownership or oversight of an AI system that is used to create an invention does not amount to a significant contribution to conception of the invention. Nor does a request from a natural person for an AI system to solve a recognized problem or pursue a general goal. Near the opposite end, the US Patent Office guidance perhaps unremarkably recognizes as an inventor a natural person who takes an AI generated output and makes a significant contribution thereto to create an invention – e.g., an inventive system of which the output is an element. With respect to the more interesting circumstance in which an LLM output itself is the invention, the US Patent Office guidance provides that a significant contribution can be found by the manner in which a natural person constructs a prompt in view of a particular problem to effect a “particular solution” generated by an AI system. In somewhat circular fashion, the US Patent Office guidance likewise informs that a natural person who designs, builds, or trains an AI system in view of a particular problem to elicit a “particular solution” can qualify as an inventor “where the designing, building, or training of the AI system is a significant contribution to the invention created with the AI system.” Although it is not clear, the language and context of the US Patent Office guidance could be interpreted to suggest qualification of a natural person as an inventor in certain circumstances where an AI system substantially performs the traditional act of conception. This suggestion warrants clarification.

The US Patent Office guidance is not definitive. It seeks public comment that may prompt changes in the near future. Notwithstanding, given the ongoing surge in generative AI interest, the principles addressed in the US Patent Office guidance still provide a basic, albeit likely dynamic, framework to assess patent prospects for inventions generated by LLM customization.

III. LLM Customization

Different techniques are being utilized to customize LLMs. Some examples include prompt engineering, retrieval augmented generation, fine tuning, and parameter-efficient fine tuning. There are many variations among the foregoing techniques. A multitude of factors can inform adoption of a particular customization technique including, for example, the level of available technical expertise, the availability of computing resources, the need to protect sensitive data, and budget considerations. Under the guiding principles set forth by the US Patent Office, different implementations of these techniques will yield different answers to the question of human inventorship – i.e., answers will be fact specific with no bright line test.

LLM outputs resulting from a customization technique such as prompt engineering and retrieval augmented generation will not establish inventorship by a natural person simply because an engineer supervises operation of the related LLM. Likewise, a bare prompt merely stating a problem or expressing a goal for an LLM to produce a solution is insufficient.

However, as mentioned, when a customization technique is leveraged to output a “particular solution” to a particular problem, human inventorship is possible. While it provides fact-specific examples that apply the guiding principles, the US Patent Office does not define the term “particular solution” to inform contexts beyond the examples. The term suggests that prompting of the LLM should substantively and meaningfully dictate or otherwise constrain some aspect of the LLM output or its generation as determined by a natural person. For example, a constraint might be a specified attribute of the input or output or a specified methodology to generate the output, such as a required input set, technical domain or sub-domain, feature, component, utilization of an elemental principle or technique – or perhaps a required omission thereof. Presumably, the more a prompt is constrained to demarcate the boundaries or characteristics of the output, and perhaps especially novel portions thereof, the more likely the occurrence of a particular solution.

A “particular solution” may require more than an output reflective of the routine entry of tone, style, token length, mode, or other format type through an output indicator of a prompt. Other configurations relating to operational settings might potentially bear on the existence of human inventorship, such as temperature, Top-K, Top-P, and the like. For example, these types of settings could be supportive of a particular solution if selection of their values is a key factor in yielding the invention as opposed to other values that result in non-inventive outputs.

Other customization techniques, such as fine tuning and parameter-efficient fine tuning, may demonstrate human inventorship with more ease. These techniques are characterized by adaptation of pretrained LLMs and adjustment of model parameters through human effort. As such, they implicate the guiding principle involving designing, building, or training of generative AI models by natural persons that may well amount to a significant contribution to an invention.

The US Patent Office guidance does not appear to definitively address whether human effort to generate a particular solution should also be claimed when a customized LLM output as the particular solution would be otherwise patentable on its own.

IV. Supporting Evidence

The US Patent Office or a court later on may require evidence of inventorship by a natural person. The US Patent Office guidance expressly instructs its personnel to scrutinize facts from the “file record or other extrinsic evidence” when determining the question of significant contribution by a natural person. Accordingly, relevant records may become essential to overcome rejections under sections 101 and 115 relating to defective inventorship. Further, although the US Patent Office guidance currently indicates that the duty of candor will implicate relevant records only in a rare instance, precautionary record retention seems wise.

To this end, consider maintaining all potentially relevant records and data that demonstrate the nature and extent of human effort in inventive LLM outputs. Prepare contemporaneous records of LLM customization and usage that engendered the invention to be patented. One example might be a log of few-shot or chain-of-thought prompting with expressed focused on a specific problem to achieve a particular solution. Another example might be evidence of a knowledge database retrieval technique that is optimized to generate a tailored prompt for a particular solution. For comparison, also keep records of inputs provided to and associated outputs generated by the LLM before customization. An input in this regard might be a rudimentary, untailored prompt without specified constraints. An output might be a result in the prior art that compares unfavorably with the invention resulting from LLM customization. Further, if successful efforts to adapt or fine tune a pretrained LLM itself are undertaken, maintain records of impacted layers or newly adjusted parameters, as applicable. Again, such records should contrast favorably over other records reflecting usage of the LLM without such adaptation.

V. Conclusion

Many questions about extant human inventorship and pursuit of patent rights remain unanswered. More will be known as the US Patent Office judges the merits of an increasing volume of model generated inventions. Apart from inventorship, the US Patent Office acknowledges that AI-assisted inventions raise other issues of subject matter eligibility, obviousness, and enablement that may need tackling soon. We should expect more clarity from further US Patent Office and governmental pronouncements in response to evolving law and policy. Notwithstanding, the current US Patent Office guidance on human inventorship is helpful to imperfectly inform patent strategies for customized LLM outputs.