Can AI Prompts Be Patented? Don’t Be Too Quick to Dismiss this Question

Recent debates about AI patentability have focused on whether the outputs of an AI system, such as a neural network, can be patented. Such debates have been spurred not only by recent general advances in the power of AI but also by Dr. Stephen Thaler’s “Artificial Inventor” project’s attempts to obtain patents on devices generated using his neural network-based DABUS software. (I interviewed and quoted Dr. Thaler extensively about his previous Creativity Machine software in my 2009 book, The Genie in the Machine: How Computer-Automated Inventing is Revolutionizing Law and Business).

If you thought that whether an AI-generated output can be patented was a cutting-edge question, then consider whether an input to an AI system, such as a prompt to a large language model (LLM), can be patented. A patent on a prompt, such as “Describe how to make and use a wide variety of hammers having shape X,” could potentially be much broader than a patent on a specific product described by the output of an LLM or other AI model. The prompt and corresponding patent claim could take a wide variety of forms. For example, the prompt might merely describe the patentable features of a device or process (e.g., without the leading “Describe how to make and use” in the example above), if the AI system is preconfigured to process such a prompt to generate descriptions of how to make devices/processes that have those features. A corresponding patent claim could correspond directly to the prompt (e.g., “A hammer having shape X”) or could incorporate the processing of the prompt by a large language model into the claim (e.g., “A hammer described by the output of a large language model in response to the large language model being provided with the following prompt: …”). The specification of the patent could go into detail about how appropriate AI models (e.g., LLMs) could be configured and used to process prompts to generate descriptions of products/processes that are covered by the claims.

If you’re curious about whether such an input to an AI system may actually be patentable, then read on.

Understanding the Basics: What Makes Something Patentable?

Before diving into the specifics of whether a language model prompt can be patentable, let’s demystify some core patent law concepts. To be patentable, an invention typically must fulfill four primary requirements:

1. Novelty: The invention must be new.
2. Nonobviousness: The invention must not have been obvious to someone skilled in the art (field) of the invention at the time it was made.
3. Utility or Industrial Applicability: The invention must have a practical use.
4. Enablement: The patent application for the invention must provide sufficient detail for a person skilled in the art to make and use the invention without engaging in “undue experimentation.”

Prompts as Descriptions of Classes of Inventions

The broad nature of a prompt might cause one to be skeptical about its patentability. However, in traditional patent claims, such breadth is not only common but often intentional. Broad patent claims aim to encompass an entire class of inventions (or to “embodiments” of a single invention).

Understanding prompts within the framework of patent law, one can view a prompt as describing a class of embodiments of an invention. For illustration, the prompt, “Provide a detailed description of how to make and use a hammer having a handle and a head attached to one end of the handle,” doesn’t pinpoint a specific hammer design but paints a broader picture of hammers having the features described by the prompt.

For purposes of the following discussion, let’s concede that a prompt describing a class which encompasses even one existing embodiment cannot be patented, as it would fail the novelty requirement of patent law. For example, the hammer prompt above would not be novel because at least one hammer (in fact, many hammers) falling within the scope of such a prompt already exist. Thus, the following discussion only relates to prompts that describe classes of products/processes in which every instance is novel.

Now consider a previously unbuilt and undescribed hammer head shape—let’s call it “shape X.” A prompt saying, “Provide a detailed description of how to make and use a hammer having a handle and a head attached to one end of the handle, where the head has shape X,” with X being a detailed description of this unique shape, could then arguably be novel.

Given this conceptual similarity between prompts and traditional patent claims, the generality of prompts shouldn’t be a fatal objection to their patentability per se. The focus should instead pivot to assessing individual prompts against all patentability requirements on a case-by-case basis, rather than objecting to the patenting of all prompts in principle.

A Spectrum of Potentially Patentable Prompts

The very suggestion that LLM prompts might be patentable may initially evoke skepticism and confusion. To demystify this and make my proposals clearer, I’ll explore a spectrum of potential LLM prompts, ranging from concrete and straightforward to abstract and contentious.

1. Specific Product Design Specifications: Imagine a prompt that describes intricate specifications for a unique mechanical device that is new, useful, and nonobvious. The prompt is provided as input to an LLM which, in response, outputs detailed engineering designs and/or 3D models of the device. Such a prompt, acting as a blueprint for an inventive product, could form the basis for a patent specification and/or patent claims, and is not too far removed from the kind of description that often is used in mechanical patent applications, especially in the claims. As a result, it shouldn’t be too controversial to propose that such a prompt could be patentable.
2. Pseudo-code for a Patentable Algorithm: Now consider a prompt comprising pseudo-code for a new, useful, and nonobvious algorithm. Assume that the prompt, when provided to an LLM-based tool (such as ChatGPT), causes that tool to output code, in a real programming language, that can be executed to carry out the algorithm. If the underlying algorithm itself is patentable, then its pseudo-code representation, serving as a prompt for an LLM to generate usable code, may also be patentable, at least in the sense that it is similar to the kind of description that often is used in the specification and claims of software patents. This example sits firmly in the realm of conventional software patentability and, as a result, should be both easy to understand and not any more controversial than the patentability of software generally.
3. Process or Methodology Description: Here, the prompt describes a new technological process, such as an efficient method for manufacturing drugs or a method of eco-friendly building construction in a new, useful, and nonobvious way. The LLM expands this into a comprehensive set of descriptions of procedures encompassed by the prompt. The patentability of such a prompt hinges on its ability to outline an innovative process that is not obvious, showcasing its potential in guiding AI to articulate a valuable and novel method.
4. General Class of Devices or Solutions: Venturing into even further abstract territory, consider a prompt outlining a broad class of devices, such as energy-efficient engines or self-healing materials, without specifying their exact designs. An example of the latter is: “Describe a class of self-healing materials suitable for use in electronic devices, capable of repairing themselves from physical damage like cracks or breaks without human intervention.” The patentability of such prompts is likely to be more controversial, and may depend on the level of detail present in the prompts and the ability of LLM-based tools to generate concrete product descriptions based on those prompts. Debate over the patentability of such prompts mirrors the longstanding controversy in patent law over the patentability fo broad claims that potentially stifle innovation by overreaching in their scope.
5. Open-Ended Innovative Concepts: At the most abstract end, a prompt might pose a question or concept, such as “ways to use nanotechnology in medical diagnostics” or “innovative technologies for large-scale environmental cleanup, addressing issues like ocean plastic, air pollution, and soil contamination.” Such prompts don’t specify a particular device or process, but instead open the door to a range of possibilities. Although prompts at this end of the abstraction spectrum are not likely to be patentable for a variety of reasons, it’s interesting to consider how far AI technology might push patentability in this direction as it becomes increasingly able to transform abstract descriptions of problems to be solved into concrete specifications of products and processes for solving those problems.

Traversing from concrete examples to abstract concepts illustrates the diverse potential of LLM prompts as potentially patentable subject matter, or at least as the basis for patent specifications and claims. I hope that this exploration and the examples I’ve provided make the concept of patenting prompts seem a bit less crazy and a bit more like a natural stage in the evolution of patenting abstract product and process descriptions that can be transformed into concrete blueprints using computer technology.

Making the Case for Patenting Prompts

As you can see, if a novel prompt guides an LLM, such as by using ChatGPT, to generate an output that describes how to make and use a potentially patentable invention, we’re progressing in the right direction. Such a prompt, if it’s genuinely unique, meets the novelty criteria.

Similarly, there is no reason why, at least in principle, a prompt could not satisfy the nonobviousness and utility/industrial applicability requirements. For example, if a prompt is not obvious to a person having skill in the relevant art (which might either be the art of prompt writing or the art of the inventions described by the output of the LLM when provided with the prompt as an input), then the prompt should satisfy the nonobviousness requirement. And if the products or processes described by the LLM’s output satisfy the utility/industrial applicability requirements, then so should the prompt itself.

Where the rubber really meets the road is with the enablement requirement, which requires that the specification of a patent describe, in sufficient detail to patent law’s hypothetical Person Having Ordinary Skill in the Art (PHOSITA), how to make and use the invention without needing to engage in undue experimentation. In lay terms, as applied to LLM prompts, this asks: Does the prompt offer a sufficiently clear pathway for actionable and reproducible results?

Whether a particular prompt enables PHOSITA to make and use the full range of embodiments within its scope without undue experimentation depends on both the prompt itself and the capabilities of the LLM that is used to generate output based on the prompt. Although today’s LLMs may not reliably be able to produce a wide range of instances of the class described by a prompt, this may change over time as LLMs improve and become more powerful. Furthermore, LLMs may be developed that are designed specifically for this purpose, in which case it would become easier over time to use LLMs to generate descriptions of a wide range of instances of the class described by a prompt. In any case, whether a particular prompt satisfies the enablement requirement should be evaluated on a case-by-case basis depending on what is disclosed in the patent’s specification, rather than rejecting all claims to LLM prompts per se.

Unintended Consequences of Prohibiting Prompt Patents

I know that some will object to patenting all prompts, even if a particular prompt satisfies the novelty, nonobviousness, and utility requirements, based on the kinds of concerns raised by the U.S. Supreme Court in its historic opinion in O’Reilly v. Morse (1853), also known as “the telegraph patent case.”

Samuel Morse’s patent in that case included a claim by which Morse asserted that his patent was not limited “to the specific machinery…described in the foregoing specification and claims,” but instead extended to “the use of…electromagnetism, however developed, for making or printing intelligible characters, signs or letters at any distances.” In other words, Morse claimed patent protection for any kind of telecommunications for the life of his patent.

The U.S. Supreme Court invalidated this claim because it would, in the Court’s opinion, be infringed by future inventions that were not disclosed by Morse in his patent and would thereby impede future innovation. I agree with those who conclude that the concerns of the Court in The Telegraph Patent Case are best addressed today by prohibiting patents from being granted if they don’t satisfy the enablement requirement and by limiting the scope of patent claims to the scope of what is enabled in the patent’s specification.

If the Court’s opinion in O’Reilly v. Morse is deemed not applicable to prompt patents, and a prompt patent having claims that are written at the level of the prompt itself satisfies the legal requirements of novelty, nonobviousness, utility, and enablement, then permitting such a patent to be granted would result in the innovative prompt being disclosed directly, clearly, and concisely to the public, thereby promoting the public policy of public disclosure of new inventions.

However, if prompt patents were to be banned, savvy patent applicants could likely obtain a similar result using an innovative workaround that would not involve patenting prompts directly. A prompt writer who crafts a novel, nonobvious, and useful prompt could task an LLM to generate as many instances of the class described by the prompt as feasible. By doing so multiple times, perhaps even across various LLMs, the prompt writer could accumulate a plethora of specific embodiments, which might (depending on the scope of the prompt and the capabilities of the LLMs) approximate the full scope of the prompt itself. The logical next step? Patent these individual embodiments, either in individual patents or in one omnibus patent, instead of patenting the prompt itself.

The end result of the strategy above would approximate the outcome of patenting a broad claim anchored to the prompt directly. Yet it isn’t clear if this promotes the public policy of issuing patents that educate the public about new technologies clearly and concisely. In fact, such a large collection of narrow patents arguably obscures the broader innovation embodied in the generic prompt. In contrast, a clear, simple, and concise claim at the level of the prompt itself might more directly, clearly, and easily convey the nature of the innovation and the scope of the claim than a large number of narrow claims.

A ban on prompt patents might just result in sophisticated prompt writers and their patent attorneys obtaining a dense and confusing mesh of narrower, and arguably less informative, patents on individual embodiments generated using the prompt. The availability of such means to indirectly achieve an effect that is similar to a patent on a prompt, but through a multitude of narrow patents that may not disclose the broader innovation represented by the prompt to the public, should give us pause before supporting a per se prohibition on prompt patents.

Looking Ahead

While not conclusively answering all objections to prompt patents, this article highlights the conceivable patentability of LLM prompts, the value of such patents, and the unintended negative consequences of a per se prohibition on prompt patents. Given the alignment of prompt patents with the spirit of traditional patent claims and their ability to clearly and concisely express the scope of protection conveyed, there exists a compelling case for their patentability in specific contexts. Additionally, the broad scope of protection that prompt patents could confer on their owners gives innovators and their patent attorneys a strong incentive to pursue such patents.

For legal professionals and patent owners, staying attuned to this evolving intersection of AI and patent law is paramount. The decisions carved in the near future could sculpt both the AI and patent realms for years to come.

Image Source: Deposit Photos
Author: denisismagilov
Image ID: 680707436