The incorporation of automation tools can significantly enhance efficiency and accuracy in a wide range of applications. There are complex issues regarding the use of AI for legal work that are beyond the scope of this article. Practitioners should review the U.S. Patent and Trademark Office Guidance on Use of Artificial Intelligence-Based Tools and ABA Formal Opinion 512 regarding the use of AI in the practice of law.
This article considers the usefulness of such tools in the context of patent drafting. While GPTs deserve attention, it is important to recognize that earlier technologies are effective and often better than GPTs for generating content for patent applications. Understanding the different tools enables patent practitioners to choose the most suitable tool for specific tasks.
Types of Text Generation Tools
The primary tools in use today can be categorized as follows: rules-based, rules combined with AI, and GPTs.
Rules-Based Text Generation
The tool modifies text based on preset programming (or rules) like search and replace, if Y then X, etc. Because modifications are based on straightforward rules, this technology is reliable and predictable but offers limited applicability.
Rules Plus AI Text Generation
By layering AI on top of rules-based automation, a tool gains advanced language understanding. Nongenerative AI tools may be used to facilitate the processing of text to improve the performance of rules. For example, nongenerative AI can perform tasks such as parsing sentence structure to extract clauses and phrases and the relationships between them. This broadens the technology’s applicability by allowing for more sophisticated rules with the information provided by nongenerative AI tools and also maintains the reliability and predictability.
GPT Text Generation
GPTs have generated a lot of hype—much of which is deserved. The workflow is simple. You provide a prompt, and a GPT will return natural-sounding text that it predicts will answer your prompt. It seems powerful on the surface, but it is too imprecise or unpredictable for most sections of a patent application.
This lack of precision stems from the way GPT is trained. GPT has been fed vast amounts of information. When prompted to provide the capital of France, GPT can reliably generate “Paris” because that information was encountered repeatedly in its training. On the other hand, when prompted to describe a new invention, GPT will generate text that sounds like it was written by a human, but this text likely will not describe the invention in the manner that is needed by the patent attorney.
The reason is that GPT generates what we describe as “average” text. The power of GPT comes from the vast amount of training data and parameters in the model. For example, when you ask GPT to write a job description for a bookkeeper, GPT will generate a job description that is, in a sense, an average of all the job descriptions in its training data. When writing a job description, this is often a great first draft to start from.
To allow GPT to deviate from the average, you can provide more information in your request. For example, you could ask GPT to provide a job description for a bookkeeper with at least 10 years of experience in manufacturing. GPT will then provide you with an average job description meeting those requirements.
With GPT, there is a tradeoff between the desired precision of the output and the amount of instructions you need to provide to GPT. When an average job description is perfectly fine, then minimal instructions are needed. For a job description for a very specific job, you will need to provide more extensive instructions. For a highly precise output, the instructions to GPT may be as extensive as the GPT output, negating the utility of GPT.
When describing the details of a patentable invention, specifics are really important. To obtain text describing the specific invention, detailed instructions would be needed.
Human-Generated Text
There is one final tool to mention before diving into the use cases: the human brain. Compared to automation tools, humans are expensive and slow. However, no technology existing today can match our ability to think critically or our years of experience drafting patents.
Using Text Generation Tools to Draft Patent Applications
With the above summary of tools that may be used for patent drafting, we will now look at four specific sections of a patent application and evaluate the usefulness of the tools for generating those sections.
Claim Summary
For most practitioners, the claim summary section is a rewrite of the claims as prose text instead of as claim language. The claim summary section provides literal support for the claims to reduce the risks of claims being invalid for not being described by the patent. For example, a claim that starts with “An automobile engine comprising one or more cylinders . . . ” may be rewritten as “In some aspects, the techniques described herein relate to an automobile engine including one or more cylinders . . . ” In addition to adding the broadening language to the beginning of the claim, the word “comprising” has been changed to “including,” which is a common practice among patent attorneys.
The claim summary has an important legal purpose, but rewriting 20 claims in the manner described above is tedious and error-prone. When practitioners write a claim summary section, they apply rules such as the one above. For many practitioners, these rules are simple enough that software may follow them and automatically rewrite the claims into the claim summary section without any AI.
Other practitioners may use a combination of rules and their own reasoning (e.g., changing sentence structure) when writing a claim summary. For these practitioners, rules plus AI may be used to automatically generate a desired claim summary section. For example, AI may be needed to parse the text of the claims (e.g., extract clauses and relationships) or to change verb tenses. Rules or rules in combination with AI are thus excellent tools for generating a claim summary section.
What about using GPTs to generate a claim summary section? GPTs can produce text that will look like a claim summary section, but it is possible that GPTs will not follow the rules precisely. Where GPT changes the text of your claims in some way, the GPT-written text may fail to meet the literal support requirement of a claim summary section.
There are two characteristics of a claim summary generation. First, the generated text must be very specific and precise. Rules and rules plus AI tools are excellent at generating specific text, while GPTs are not. Second, the generated text is a simple modification of existing text. There is no need to use a complex tool when a simple tool will do the job.
Abstract
A patent application is required to have an abstract. The abstract provides a short summary of the patent application to allow a reader to quickly understand the subject matter of the patent. For many practitioners, the abstract is a summary of the claims.
GPTs are excellent tools for generating an abstract. With a carefully crafted prompt, a GPT can be instructed to rewrite patent claims so that the subject matter is expressed in the style of a patent abstract. GPT training data includes large amounts of patent text, so GPTs contain information about the style of patent abstracts. Further, GPT prompts may use few-shot learning to help guide GPTs as to the desired style of the output text.
So why do GPTs do so well here when they were such a poor choice for the claim summary section? The key factor is removal of the requirement of precise text. As long as the abstract generally describes the subject matter of the claims, the precise text of the abstract is not of great importance. As with any use of GPT, the patent practitioner must review the abstract for accuracy, but in many instances the generated abstract may be usable without any changes or with small modifications.
By contrast, rules and rules plus AI are not great choices for generating an abstract. These tools could be used to generate a grammatically correct abstract that essentially repeats the language of the claims. Such an abstract would be awkward to read (because claim language is awkward legalese), and it would thus not be useful for quickly conveying the subject matter of the patent. Further, an abstract must be no more than 150 words, and rules and rules plus AI are not sophisticated enough to determine what subject matter to leave out when the claims are longer than 150 words.
Detailed Description
The detailed description is the longest section of a patent application. It must meet many legal requirements, such as providing a description of the invention to enable a person to implement the invention. Fortunately for practicing patent attorneys, no technology is yet able to generate an entire detailed description of sufficient quality. A detailed description has many different portions for different purposes, and it may be that some portions of the detailed description are amenable to automated generation.
The two factors discussed above are relevant here: whether (1) precise and specific text is required, or (2) the text to be generated is a simple modification of existing text. Consideration will now be given to a couple of common drafting tasks for the detailed description.
Frequently, a patent application needs to describe background technology that is generally relevant to the subject matter of the invention. For example, a patent application relating to an electronic circuit may need to describe how capacitors work. Although the invention is not a new capacitor, the patent application may need to explain how to build the invention using different types of existing capacitors. GPTs are excellent tools for generating a description of background technology relevant to the subject matter of an invention. A key factor is that almost any text describing capacitors will be useful, and we do not need a specific or precise description of capacitors.
Other portions of the detailed description have text that looks similar to claims. For example, a patent application may have a “method” claim that recites a sequence of steps. With a method claim, there should be a flowchart diagram to illustrate the sequence of steps, and the flowchart will need to be described in the detailed description. The text describing the flowchart will be very similar to the text of the method claim. GPTs are not good tools for generating text describing a flowchart. First, the flowchart must precisely describe the invention of the method using the specific text of the method claim. As discussed above, GPTs are not good at these kinds of generation tasks and may use average language instead of precise language. Second, when describing the flowchart, a practitioner generally fills in many gaps that are included in the claim text. Where GPTs fill in the gaps, they will use average language that is unlikely to be useful.
Here are a couple of examples. First, claims are necessarily short and leave out details that may aid in understanding the invention or how to implement the invention. Where GPTs try to fill in such details, they will use average text instead of specific text relevant to the invention. Second, a claim may describe one possible implementation of the invention, but it may be desired to describe other possible implementations that may be claimed in future applications. GPTs may add other implementations, but they are likely to be average implementations (e.g., something seen in its training data) rather than the needed novel implementations.
People have proposed providing additional inputs to GPTs to help GPTs overcome the flaws described above. For example, an inventor may provide an academic paper describing the invention or an invention disclosure call between the patent attorney and the inventor may be recorded and converted to text. While this kind of information is very valuable to a patent practitioner, it remains to be seen whether GPTs are able to use this kind of information. The key problem is that GPTs are trained to generate average text from their training data in response to prompts. GPTs are not programmed to have “reasoning”; they are programmed to generate likely sequences of words given prompts and their training data. This is not to say that it is not possible for generative AI to use these kinds of inputs effectively, but that the current technology is not able to do so.
Rules or rules plus AI have some utility for generating text describing a flowchart, but that utility is limited. Rules are only able to rewrite the text of the claims, but this rewriting can still save time for a practitioner drafting a patent application. For example, rules may be used to automatically draft a template for describing a flowchart. The claim text may be rewritten to look like flowchart text, and placeholders may be inserted to indicate where the practitioner should provide additional information.
Rules and GPTs thus have some utility for automating the detailed description section and decreasing the amount of time required to draft a patent application. The patent practitioner, however, still needs to do most of the writing.
Claims
Patent claims precisely define the invention and indicate the rights of the patent owner to prevent others from using the invention. Drafting claims requires the nuanced understanding and expertise of a skilled patent attorney. Given an invention disclosure document, is it possible to automatically draft a valuable claim covering that invention?
Rules, even with AI, are clearly too simple for such a task.
People are already using GPTs to generate patent claims from an invention disclosure. The generated claims will look like patent claims, but are they valuable claims? Will the claims focus on the key point of novelty and be broad enough to be valuable but also narrow enough to avoid the prior art? It seems the tendency of GPTs to generate average text means they are very unlikely to generate valuable claims. GPTs will generate claims based on other patent claims that they have seen, but different inventions have very different claim requirements, and it is unlikely that GPTs will do this well.
A patent attorney could use a GPT to generate an initial version of the claims and then rewrite the claims to fix the flaws in the GPT-generated claims. It may, however, take just as long to rewrite the claims as to write them from scratch. Additionally, the GPT’s influence on the claim language may lead the attorney down the wrong road intellectually and impact their thinking around the client’s invention. Either way, attorneys are better off just writing the claims from scratch. There really is no match for the power of the human brain—at least not today.
A Collaborative Future?
While automation tools, including GPTs, have shown promise in patent drafting, their utility is highly dependent on the specific section of the patent application. Rules-based tools and those combining rules with AI excel in tasks requiring precision and adherence to specific guidelines, such as generating claim summaries. GPTs, on the other hand, are more effective for generating text that does not require high precision, like patent abstracts or descriptions of background technology. As of now, human expertise remains indispensable, particularly for drafting sections like claims and detailed descriptions that require nuanced understanding and critical thinking. The ongoing evolution of these tools, however, suggests that the future of patent drafting will likely be a collaborative effort between human practitioners and advanced AI technologies.