Can AI Replace Your Patent Attorney?

I am seeing an alarming increase in the uninformed use of Large Language Models (LLMs), like ChatGPT, in the formal invention process, and even in the drafting of utility patents. Many inventors are choosing to forego the expensive traditional process of hiring a patent attorney for what could be tens of thousands of dollars, and instead learning the process themselves and filing a utility patent with the USPTO for as little as $70 for a micro entity. However, division of labor evolved for a reason, and the typical inventor cannot feasibly learn all the related legal knowledge necessary for such an endeavor, nor is this efficient for the progress of technology as a whole. Nevertheless, it’s understandably enticing, and even more so with the fruition of mind-bendingly powerful AI tools, like GPT-4, in the last few years.

Now the patent drafting process is more accessible than ever, allowing inventors to simply plug their ideas into a chat prompt and receive something akin to an entire patent specification in seconds. But what dangers does this pose for the daring inventor, and is a modern LLM really a viable replacement for a bar-hardened patent attorney? Also, can you trust a public model like ChatGPT trustworthy to handle your unprotected intellectual property?

The Competence of LLMs

Someone who knows how an LLM works understands that the model only returns a string of words, with each next word informed by the large textual datasets the model was trained on. When you ask ChatGPT to write a set of patent claims based on data you provide it about your invention, there is no promise that the claims it returns aren’t legalese gibberish. Of course there will likely be some coherence to what the model returns, but you need to understand that the USPTO is picky. If one claim isn’t worded exactly how they want, they could potentially reject issuance of the entire patent. The USPTO is also slow, meaning you just wasted years waiting for that patent to be issued, only to have it rejected, and even if you revise and refile, you may be waiting years more. Getting your claims written exactly how the USPTO likes them is why you’re paying the patent attorney their thousands of dollars. Could a selectively generated string of words convince a patent examiner? Yes. Will you trust ChatGPT to do so? That is up to you.

Due to the recency of “competent” AI tools, and the sluggishness of patent offices, there is not much data on the success rates of LLM-drafted patent specifications yet. The time is coming up where we may start seeing statistics on this which will hopefully lead to a more informed decision on the risk for the frugal inventor. It’s not currently required to disclose the use of AI tools in the drafting of a patent specification submitted to the USPTO, so it will be difficult to know which issued patents or rejected applications are LLM-assisted without the disclosure of the inventor. All this is not to say that an inventor or attorney should not use LLMs to help in the drafting process. I wholeheartedly support the use of AI tools for the tasks they excel at, like volume-based tasks like web crawling, brainstorming, and creating templates. Many patent attorneys are even creating customized models to speed up their more menial tasks, like background searching and drafting a patent skeleton that they will later fill in using their human expertise. But the consensus still remains that there is no competent AI replacement for the hardened expertise of most professionals, including patent attorneys, and development engineers like myself.

Do Your Homework, Save a Fortune

As a side note, in September the USPTO released a guidance update for the assessment of “AI-assisted inventions”, which applies more to an LLM’s intellectual contribution to the patent’s claims than to its use in the drafting of those claims. However, these two matters are nonetheless entangled. When an inventor uses an LLM in the drafting of their patent claims, the LLM’s results can very easily contain additional or skewed information which might be misinterpreted or misunderstood by the inventor if they are not well-versed in legalese (which they rarely are). While the USPTO assures us that the use of an LLM in the process of invention “does not negate that person’s contributions as an inventor”, they also warn that “Maintaining ‘intellectual domination’ over an AI system does not, on its own, make a person an inventor of any inventions created through the use of the AI system”. This means that if a person entrusts an LLM with the full creative invention process, that person is not automatically a “contributor” to the (supposedly novel) inventive claims made by the LLM (this is one of the three so-called “Pannu factors” that a person need to meet to qualify as a significant contributor to an invention).

Here’s my advice for inventors: do a little diligence to inform yourself on the patent application process, what patent attorneys do (and why we pay them so much to do it), and how LLMs work. Then decide for yourself how to spend your time and money. A smart option for you may be to use an LLM to create a rough draft that you then send to an attorney for revision and finalization. This may save you significant money over paying an attorney to do everything in-house, and would definitely be a significant time-saver over learning to competently draft a utility patent all on your own.

Protect Your IP From Public LLMs

Lastly, I want to make this important note. Due to its virality and accessibility, OpenAI’s ChatGPT (and other lesser-used models like Antropic’s Claude and Google’s Gemini) has been the go-to for anything we feel enticed to use an LLM for, including patent drafting for some. However, there are some serious concerns with trusting public models such as these with your IP. The conversations between a user and a public LLM like ChatGPT are often stored in third-party servers, and such LLMs often have terms of use that allow certain individuals to access this information. Both of these facts are a liability for anyone submitting any sensitive information into these models. It’s possible that by putting sensitive information into ChatGPT to draft your claims that you may be unknowingly invalidating your IP through public disclosure (laws on this vary by country). The risk may be low, but not zero. Luckily, there are ways around this, one of which is to host your own private LLM. Models are commonly available now that are small enough to run even on a modest computer, though you can expect a much slower computation time from these than one hosted by a megacorp. Nevertheless, it seems to me a small price to pay for IP security. You can find out more about how to host your own private model here, as well as information about what specific model and settings are right for your needs.

I hope this helps you inventors make informed decisions about managing your IP and getting your idea onto the shelf! Don’t hesitate to email me with any questions or feedback at benjamin@ppd-usa.com.

Share:

_
_
_
_