How to use AI to help draft a patent application

Artificial intelligence tools are now being used very widely, including for legal and technical writing. In many contexts, AI can help create and review documents at a level which would otherwise require significant experience, training and time. Patent drafting is no exception.

That said, AI is still only a tool. Like any tool, it can be used well or badly. A familiar principle from computing remains highly relevant: garbage in, garbage out. If the instructions are vague, incomplete or misleading, the result is unlikely to be reliable. If the instructions are clear, precise and well structured, the result can be surprisingly useful.

This article considers how to get the best out of AI when using it to help draft a patent application.

It should be said at the outset that this guidance is based on our own experience. AI platforms are developing rapidly and any observations of this kind should be treated as a snapshot in time and from one particular professional perspective. Even so, our experience to date is that the main AI platforms, including systems such as ChatGPT, can be extremely helpful when used carefully.

There are also many specialist products which purport to draft patents, review inventions or carry out legal analysis automatically. At present, many such services appear in substance to operate as an additional front end over more general AI platforms, while charging substantial fees. In our view, much of the benefit that these tools claim to offer can often be obtained by careful and intelligent use of a high-quality mainstream AI platform.

So what does careful use actually look like?

The first point is that the AI platform should be told clearly what the objective is. Are you trying to prepare a United Kingdom patent application, a European patent application, a United States patent application, or a PCT international patent application under the Patent Cooperation Treaty? That matters, because the language, structure and legal assumptions may differ depending on the jurisdiction and filing route. If you do not specify the intended route, the AI system may draw on mixed drafting styles and produce text which is formally inconsistent or strategically unhelpful.

A useful starting instruction may therefore be along these lines: please assist me in preparing a draft patent application intended for filing first in the United Kingdom, using terminology and structure appropriate for UK patent practice. A similar instruction can of course be adapted for Europe, the United States or the PCT route.

The second point is even more important. When describing the invention, focus first on how it works. Many inventors begin by explaining what the product is for, what its benefits are, why it is commercially attractive, or what problem it seeks to solve. All of that may be relevant, but it is not enough on its own. For patent drafting, the core question is: how does the invention actually work?

The AI system should therefore be given a careful explanation of the technical operation of the invention. What are the physical or functional components? How do they interact? What steps occur, and in what order? What features are optional, and what features are essential? What alternatives or variants are possible? What materials, dimensions, control features, sensors, software elements, or mechanical relationships are involved? The more clearly those matters are explained, the more useful the resulting draft is likely to be.

This point is not merely stylistic. If the inventor cannot explain how the invention works, the subject matter may not yet be ready for patent drafting. It may still be only an early concept or commercial aspiration. That does not necessarily mean a patent is impossible, but it often means that a human discussion is needed to identify what the invention actually is. That is one of the important ways in which a patent attorney adds value.

A third useful step is to ask the AI system to help test patentability at a preliminary level. This must be done carefully. It is usually not enough simply to ask whether an invention is patentable. The better course is to ask narrower and more precise questions.

For example, after giving a full explanation of the invention, you might ask: given the explanation of my invention provided earlier in this chat, please review whether the invention appears likely to be novel in light of publicly available patent literature and technical material. Novelty is one of the core requirements for patentability. Upon that foundation, questions of inventive step can then be considered.

A further prompt might ask: given the explanation of my invention provided earlier in this chat, please assess whether the invention appears likely to satisfy the requirement of industrial applicability. That may sound straightforward, but it can be very important. Some proposals are not refused because they are uninteresting, but because they fall into categories which patent law treats differently, for example abstract ideas, business methods, certain software-related subject matter, or other excluded material depending on the jurisdiction.

Again, this is where precision matters. The standards applied in the United Kingdom, Europe and the United States are not identical. A prompt which does not identify the intended jurisdiction may produce an answer which is legally mixed and therefore of limited value.

A fourth point is that AI should be used iteratively, not just once. It is often better to build the draft in stages. One may begin by asking the AI to summarise the invention. Then ask it to identify the broad inventive concept. Then ask it to propose possible independent claims. Then ask it to identify fallback features for dependent claims. Then ask it to prepare a fuller specification description based on those elements. Finally, ask it to check the draft for internal consistency, unclear terminology, unsupported claim language, and unnecessary limitations.

This staged approach is usually more effective than simply asking the AI to draft a complete patent application in one step. A one-step prompt often produces text that looks polished but is structurally shallow. A staged approach gives the user more control and allows important errors to be corrected earlier.

It is also often helpful to challenge the AI directly. Once a draft has been created, one may ask questions such as: which terms in this draft are vague or indefinite? Which claims may be too narrow? Which features appear to have no clear support in the description? Which parts of this application rely on assumptions not expressly stated? Where might added matter issues arise if amendments are made later? These are the sorts of questions that move the process from mere text generation towards more disciplined drafting.

Another practical point is that inventors should not rely unquestioningly on the confidence of AI-generated output. AI systems often produce fluent and persuasive language even when the underlying legal or technical analysis is weak. In patent work, that is dangerous. A draft may look professional while containing omissions, inconsistencies, over-broad statements or jurisdictionally inappropriate language. It is therefore wise to treat AI output as a working draft, not as a finished legal instrument.

Confidentiality should also be considered. Before uploading sensitive technical information into any AI platform, the user should consider the platform’s terms, privacy settings, training policies and data handling arrangements. That is a commercial and legal consideration in its own right.

The general lesson is therefore simple but important. To get the best results from AI patent drafting tools, be precise. Define exactly what you want the AI to do. Tell it which jurisdiction or filing route is intended. Give it a clear technical explanation of how the invention works. Break the task into stages. Ask focused questions about novelty, inventive step, industrial applicability and claim structure. Then review the result critically rather than assuming that polished wording is necessarily good drafting.

In our experience, this approach can improve the usefulness of AI substantially. The final result is still unlikely to be perfect, but it may become a much better starting point for further work.

A further and often valuable step is then to refer the resulting draft to a patent professional for review. At Patent Outsourcing Limited, we can review self-drafted and AI-assisted patent applications and provide comments on potential deficiencies and suggestions for improvement. There is no true substitute for a patent application drafted from first principles following direct discussion with the inventor. However, we recognise that some clients do not wish to take that full route or incur the associated cost at the outset. In those circumstances, a focused review service can be a useful way to avoid serious pitfalls and obtain professional input before filing.

For related information, please also see our article “AI and self-drafted patent applications”.

Can AI draft a patent application?

AI can help generate a patent application draft, but the quality of the result depends on the instructions given, the technical detail provided, and the jurisdiction for which the application is intended.

How should I prompt AI to write a patent application?

The best approach is to define the jurisdiction, explain clearly how the invention works, and ask the AI to build the application in stages rather than in a single prompt.

Can ChatGPT help draft a UK patent application?

Yes, it can assist with drafting and review, but the user should specify that the intended application is for UK practice and should not assume that the resulting draft is ready to file without careful review.

What information should I give AI for patent drafting?

You should explain how the invention works, what components or steps are involved, which features are essential, which are optional, and what variants may also be relevant.

Can AI assess whether an invention is patentable?

AI can help with a preliminary review of novelty, inventive step and industrial applicability, but such analysis should be treated cautiously and not as a substitute for professional legal advice.

Why does an AI drafted patent application still need review?

Because patent drafting is not only about producing fluent text. It also involves legal structure, technical disclosure, claim strategy and jurisdiction-specific conventions which AI may not handle reliably without careful guidance and checking.