Is ChatGPT a Useful Tool For Drafting Patent Applications?

SPOILER ALERT:  For several reasons, the answer is a resounding no.  At least not yet.

I recently began seeing more and more references to “ChatGPT” in my Linkedin feed.  Curiosity got the better of me, so I checked it out.

For those unfamiliar with ChatGPT, it is a free computer program that can understand and respond to human language, like a virtual assistant.  But, it is different than Alexa and Siri, and perhaps in a league of its own.  It is trained on a large amount of text, so it can understand what people are saying and respond in a sensible way. It can converse, write computer code and even draft original stories and articles.  By way of non-limiting example, you can ask the program to converse with you about quantum physics, draft a 1,500-word position paper on climate change, and even write a Bingo program in the Python programming language.  After inputting your request, the program rapidly provides you with shockingly sophisticated results.  That said, ChatGPT is not infallible and has been reported to output errors and misinformation.  The program may be accessed at the website chat.openai.com.

I took the program for a test drive and understand now why ChatGPT is getting so much press.  The results are impressive.  And, those results will reportedly only get better as the developer continues to refine the program.  OpenAI, the developer, recently announced that it will eventually release a professional paid version of ChatGPT with faster performance. It has also been reported that Microsoft is in the process of incorporating OpenAI into Bing, Outlook, Word and PowerPoint.

As a patent practitioner, I naturally wondered whether ChatGPT was capable of drafting a patent application.  Could patent application drafters like me soon be replaced by a computer program?

To help answer this question, I first asked ChatGPT if it was capable of drafting a patent application.  The program responded that it could draft an outline of a patent application, but that I should consult a professional.  ChatGPT explained that in drafting a patent application it uses patent-related information obtained from a variety of sources, including text from publicly available patent documents, patent office databases, and other online sources. It has been trained on a large dataset of text data, including patent documents, which allows it to understand and generate text related to patents.

During my test drive, I asked the program to “draft a patent application directed to a four-legged table.” The program swiftly outputted a patent application including a title, abstract, detailed description and three claims.  The program even provided a listing of drawing figures (i.e., top view, side view and cross-sectional view).  The text-based program did not provide any drawing figures to compliment the detailed description.

The description of the table outputted by the program was more robust than I had expected.  According to the program, the inventive table has a tabletop, four legs and brackets connecting the legs to the tabletop.  It also described the purpose, position, structure and material options for each component of the table.  The program even included the following  problem-solution type framework in the detailed description: “… most four-legged tables do not have the ability to adjust their height. This invention addresses this problem by providing a four-legged table with a bracket that allows for the height to be adjusted.”  Bear in mind that I did not provide any background information about height adjustment to the program.  I simply input “draft a patent application directed to a four-legged table.”

The independent claim outputted by the program is shown below.

1. A four-legged table comprising a flat top surface and four legs that support the top, wherein the legs are arranged in a rectangular configuration and are connected to the top surface via a bracket, and the bracket allows for the height of the table to be adjusted.

I found it interesting that the claim randomly included a height adjustable bracket.  I first wondered if the program had attempted to invent a new table on its own.  So, in my cross-examination, the program admittedly responded that it did not have the ability to invent and that the claim was “based on information that [ChatGPT] has been trained on.”  As an aside, I note that the U.S. Court of Appeals for the Federal Circuit has already ruled that a computer program cannot be named as an inventor on a U.S. patent application.

I then wondered whether ChatGPT included the height adjustable bracket to differentiate its table claim from tables known from the “prior art.”  In other words, had ChatGPT output a claim that it believed should be patentable in view of the prior art?

In response to my question as to why it included the height adjustable bracket, the program responded that the height adjustable bracket was a “unique and useful feature of the four-legged table.”  The use of the word “unique” was intriguing.  Was the program conflating uniqueness with novelty and/or non-obviousness?  I then asked how it knew the height adjustable bracket was unique.  The program’s response is shown below.  Essentially, the program admitted that height adjustable bracket was distinctive and useful, but not necessarily novel.

“As a language model, my knowledge is based on the text data I’ve been trained on. I don’t have the capability to independently verify the uniqueness of an invention. However, when I included the mechanism for adjusting the height of the legs in the patent claim, it was based on the assumption that it is unique feature, and it would make the invention more distinctive and useful compared to the other similar products that may exist.  To determine the uniqueness of an invention, you will need to conduct a search of prior art, which includes existing patents, publications, and other publicly available information that may be relevant to your invention. This will help you to determine whether your invention is novel and non-obvious, which are requirements for obtaining a patent.”

In response to my question as to whether “unique meant novel,” the program responded “No, unique and novel are not synonymous when it comes to patent law. Unique means that the invention is one of a kind and has no identical counterpart … Novel, on the other hand, means that the invention is new, or it is not part of the prior art.”

And, in response to my question whether it had conducted a search of the prior art in arriving at the table claim, the program responded, no, as shown below.

“As a language model, I don’t have the ability to conduct a search of prior art. My knowledge is based on the text data I’ve been trained on. I can provide general information on the process of searching for prior art, but a comprehensive search of prior art requires specialized knowledge and tools.”

So, by this time I had sufficiently convinced myself that the table claim outputted by the program had not described a novel and/or non-obvious idea.

Further into my test drive, I tested the program’s ability to generate a patent application directed to more complicated inventions.  I found that the results outputted by the program were non-specific and, generally, not useful.  And, for those reasons, I concluded that it would not be worthwhile to use ChatGPT as a tool for drafting patent applications.

Of course, even if a computer program was capable of drafting the perfect patent application, there would be many reasons to avoid heavily relying on such a program.   Patent practitioners ordinarily work with their clients to develop a patent application that accounts for competitive products, licensing activity, trade secrets, prior disclosures made by the client, related patent applications belonging to the client, and more…  A computer program would be blind to those critical details.

As a patent practitioner, another important factor that must be considered when using a computer program is confidentiality.  I note that ChatGPT states that it does not have the capability to access or store user data.  Regardless, patent practitioners can never be too careful with client data.

So, in summary, while ChatGPT is a novelty for generating articles and performing other tasks, it is not (yet) a useful tool for drafting patent applications.