We’re experimenting with something that many patent firms are reluctant to touch, reviewing AI-written invention disclosures from startups, SMEs and other technical companies, with a view to turning suitable disclosures into provisional patent applications. And the cost? …less than fixing it after filing!
If your engineering or product team has already used ChatGPT, Claude or another AI tool to document your company’s invention, we are willing to look at what you’ve produced and see whether it would be useful for a Provisional patent application. We’ll fix minor issues that would normally get a rejection from the patent offices.
The reality is that AI is already deeply embedded in patent drafting and technical documentation. Almost every patent application these days is likely to involve some level of AI assistance, whether that’s a simple AI review of human-written text, engineers using AI to organise their technical notes, or a full-on dump from an all-night Claude Co-work session.
The results vary greatly – so does the amount of AI-interference.
The legal position is reasonably clear. AI cannot simply be the sole inventor, but AI tools can be used to assist human inventors and practitioners when used responsibly. There are still plenty of unresolved questions about AI, inventorship, disclosure and patent validity, and many of those questions have not yet been fully tested by the courts. I’m not going to get into the jurisprudence of AI inventions here. Instead, I want to deal with a much more practical question:
For more on that issue, see DABUS Cries for Attention: Understand Why AI Inventorship was Refused.
We have also written about how companies should approach human contributions to AI-assisted inventions.
Some of the better AI-generated patent descriptions I’ve seen from companies follow roughly the same structure. They start with the Field of the Invention and the Background to the technical problem. They then give a reasonably straightforward, lay summary of the invention before getting into the technical detail. From there, describe your company’s invention in enough detail that someone can understand the components, how they relate to each other and, importantly, how the invention actually works.
Focus primarily on the technical aspects rather than the size of the market, sales projections, anticipated customer demand or other commercial advantages.
This is also why we distinguish between patenting your product and patenting the underlying invention.
You don’t need to produce a perfect patent specification. This is a Provisional Patent Application.
For more on where company drafting can make sense, see Patents: When Do You Do It Yourself?
The claims are the most important, and most difficult part of a patent application.
They’re also where AI-generated specifications tend to go badly wrong.
AI can produce text that looks remarkably like a patent claim. It has the semicolons. It has the “comprising”. It may even have beautifully numbered dependent claims.
That doesn’t mean it’s a good claim.
A skilled patent practitioner is thinking about considerably more than whether the claim sounds patent-like. We’re thinking about what the actual inventive concept is, what needs to be in the claim, what should be left out, what prior art might exist, what competitors might do differently and whether the terminology used in the claims is properly supported by the specification.
Fortunately, a bad AI-generated claim isn’t necessarily fatal.
Even when the claims are completely wrong, the AI-generated material will often tell us enough about your company’s technology that we can work out what the claims should say.
AI loves descriptive titles.
Give it a new widget and you may receive something along the lines of:
“Intelligent Sensor-Based Adaptive Widget Management System Using Real-Time Machine Learning Optimisation.”
That might describe the invention beautifully.
It’s probably not the title we would choose.
Patent titles are generally better when they’re concise and reasonably broad. Don’t try to squeeze the entire inventive concept into the title.
This is one of the most dangerous weaknesses because it’s an absence error.
An AI-generated specification may confidently tell us that the company’s invention includes:
Wonderful.
But how are they connected? What does the sensor actually measure? What signal does it generate? What happens to that signal? What calculation does the processor perform? How does that calculation cause the device to do whatever makes the invention useful?
Patent specifications have disclosure and enablement requirements. Merely naming the ingredients isn’t necessarily enough.
A useful test for your engineering team is to ask:
“Could another technically skilled person understand from this description how to make and use the invention?”
If an important part of the answer is “well, obviously you’d just make that bit work”, that’s usually the part that needs more explanation.
Then there are the dozens of comparatively mundane drafting issues.
Paragraph numbering. Antecedent basis. Inconsistent terminology. Reference numerals. Formatting. The order of sections. Text appearing unnecessarily in drawings. Calling the same component a “controller”, “processor” and “control module” in three consecutive paragraphs.
None of these necessarily means your disclosure is useless. BUT you will get a Patent Office rejection that costs you dearly to fix.
In fact, don’t have one of your engineers spend three hours trying to make an AI document look exactly like a professionally prepared patent specification before sending it to us.
We can fix formatting.
Have them spend that time making sure the technical substance is there.
AI is extremely happy to tell you why your company’s invention is wonderful. In fact, every month we receive perpetual motion machines to patent (something to do with our name, I think). AI happily encourages the delusion.
It’s faster. Lighter. Smarter. More efficient. More intuitive. Customers will love it. Manufacturing costs will fall. Productivity will soar.
That’s great for a pitch deck.
It’s much less useful to a patent practitioner unless those advantages help explain the underlying technical invention.
If the device is lighter, what technical feature makes it lighter?
If it’s faster, what architecture or process makes it faster?
There are two things that AI-Generated patent submissions omit, that you won’t realize is missing, and that will eventually kill your patent. Way too much to get into in this blog but we will spot it.
If your engineers, founders or product team have been using ChatGPT, Claude or another AI tool to develop or document an invention, don’t assume that work needs to be thrown away and started again.
Give us substance.
We want to understand:
Diagrams, CAD drawings, sketches, photographs, test results, pseudocode, architecture diagrams and engineering notes can also be extremely useful.
And don’t worry too much about making the disclosure sound like a patent.
If your company is also preparing to pitch, demonstrate or announce the technology, read our guide to public disclosures and patent strategy.
We expect the claims may be wrong. We expect some terminology to be inconsistent. We expect AI enthusiasm about how revolutionary the invention is. And we expect there may be important technical details or alternative embodiments missing.
Those are things we can work on.
What we’re interested in is whether, somewhere underneath all that AI-generated prose, your company has an invention worth protecting, initially as a Provisional patent, then 12-months later as a fully explored and formatted patent.
We’re offering qualifying startups, SMEs and other technical companies the opportunity to submit their AI-written invention disclosures for our review, with a view to preparing provisional patent applications from suitable submissions.
