The Ups and Downs of AI-Assisted Patent Drafting
Dear LLM, please write a patent application
The task of drafting a new patent application can to some extent be automated with the help of an LLM. Based on almost any input material, it can produce an entire application text or shorter passages that we can add to our own text. We can try to improve the output gradually with continued prompting. However, we need to be careful because a badly formulated claim or mistaken phrase can render the application worthless. I will here explore some pitfalls of LLM-assisted drafting that patent attorneys should be aware of.
The data security perspective is paramount. If our data is not secure in the LLM, the invention we write about may leak to the public domain before our application is finished and filed. The information we use when working on client A’s application may also inadvertently infiltrate the work we do for client B, which amounts to a breach of confidentiality. However, for the sake of this article, I will assume that these problems can be avoided by selecting the right LLM and learning how to use it.
My focus will be on the practical aspects of patent drafting: how do we produce the material we need, and how long does it take? I will review two different workflows. The first is inspired by J.C.Schütte’s article Beyond the blank page: Rethinking workflows when drafting patent applications with AI-assistance (epi Information 3/2025). I will call it the LLM-first workflow. Here the input to the LLM can be an invention report, some prior art documents and other material related to the invention. After some preparatory prompting the LLM is requested to generate a claim set and a description, i.e. a more or less complete first draft of the application. In this workflow the LLM lays the groundwork, the attorney reviews the draft and revises it into a complete application (manually or with additional LLM assistance).
A second, attorney-first workflow is a less radical departure from traditional patent drafting. Here the attorney handles the invention report, studies the prior art and writes at least the most important claims and key parts of the description manually. This material is input to the LLM. The attorney then conducts a dialogue with the LLM, for example by asking it to evaluate the clarity of the claims and to expand different sections of the description (perhaps one at a time). The attorney evaluates each LLM suggestion and considers if it is good and relevant enough to be added to the application text (possibly after some additional modifications). The preparation and improvement of figures can be included in the same iterative process. In this workflow the attorney lays the groundwork manually and then expands the draft into a complete application with LLM assistance.
The expected benefits of both workflows in comparison to fully manual drafting are that (1) at least some steps in the drafting process are automated, and therefore quicker, and (2) the LLM can generate material which the attorney might not otherwise have thought about. Let’s examine how certain and strong these benefits are in the two workflows.
The quality of LLM-generated material
A first draft is generated much more quickly in the LLM-first workflow than in the attorney-first workflow. A few model applications written by the prompting patent attorney might be given as examples to the LLM. Any number of earlier patent publications or case law can also be given as input. Combining this background information with the invention report, the LLM might be able to generate an application text in the attorney’s preferred style and format so that it contains options which would not have occurred to the patent attorney.
However, the LLM is likely to struggle in extracting the necessary features of the invention from the invention report. It may be adept at identifying features of the invention which differ from the known prior art documents, but this is only a preliminary step in writing a good independent patent claim. It is much more difficult to decide which of those features need to be excluded from the claim because the invention can be implemented without them.
In most fields of technology, an invention is a practical solution to a practical problem. The considerations which make some features essential and others non-essential are therefore also going to be practical. Solution A may work well but require costly modifications to the device. Solution B may not work quite so well, but it might still be preferable because less modification is needed to accommodate it.
Most inventors assume that the background knowledge which led to the selection of solution B over A isn’t of interest and shouldn’t be included in an invention report. They
may put a lot of weight on experimental evidence which supports the best embodiment but say little or nothing about other alternatives. Consequently, many invention reports contain too little information to form a useful, stand-alone starting point for claim drafting. This is the most important pitfall in the LLM-first workflow.
In most fields of technology, patent attorneys discern the necessary features of the invention (which will form the independent claim) by asking if the same effect can be achieved by other means than the ones described in the invention report. The easiest way to find answers is to access the inventor’s practical knowledge in an interview. This is far more likely to generate a high-quality, legally secure patent claim than an LLM-draft which works only on the basis of an invention report and additional material which lacks a direct connection to the invention at hand.
In the attorney-first workflow the independent claims are produced manually, so their quality depends only on the attorney. The LLM is tasked with expanding the material which the attorney has produced, and it is up to the attorney to decide where the manually written sections must be retained and where LLM-assistance is requested.
Intermediate workflows could also be conceived. In the LLM-first workflow, the attorney can of course review and revise LLM-generated claim proposals before prompting the LLM to generate the entire application text. Conversely, in the attorney-first workflow the attorney can have the LLM review the first draft of the claims and request suggestions for improving their clarity.
The quality of LLM-generated material also depends on how skilfully the attorney prompts the LLM, just as the quality of a manual work product depends on its maker. It is therefore not possible to declare definitively that either workflow would always produce a claim or an application of higher quality. But in the end, the attorney has final responsibility and cannot rely blindly on the LLM.
The amount of attorney work
Timesaving is a tricky question in LLM-assisted drafting. On one hand, the LLM generates lots of text in a matter of seconds. On the other, the more text the LLM produces, the longer it will take for the attorney to review it.
Reviewing and revising a full application draft might be a reasonably quick task in an ideal scenario where the LLM generates a high-quality draft. But if the text produced by the LLM has fundamental flaws, the prompting attorney is back in square one. The attorney’s reviewing work can expand dramatically if the LLM made an error in the independent claim, for example by including features which are not necessary for the invention. Major updates may then be needed throughout the description, because all parts of an application text reflect the content of the independent claim.
After performing a thorough review of an entire application draft in the first round of the LLM-first workflow, the attorney may be reluctant to prompt the LLM to rework a new iteration because it would be too much work to start the review from the beginning. The attorney could try to rework and improve the first draft in dialogue with the LLM, but this might require skilful prompting and a lot of double-checking to see if the end result really corresponds to the attorney’s intentions. Revision can be hard work.
In the attorney-first workflow the expected workload is easier to predict. The attorney can put a regular effort into writing the key parts of the application manually and then spend a shorter time on the gradual expansion of the application text with LLM-based suggestions. It is up to the attorney to decide when the manual work comes to an end and LLM-assisted work begins. As the attorney gains more confidence in working with the LLM, the time spent on manual work can perhaps be gradually shortened.
In a way, any LLM output is a lottery where the outcome is influenced by the attorney’s prompts and by the input material given to the LLM. If all the output is legally sound and relevant in the LLM-first workflow, the process can be extremely efficient. But if it is inadequate, revising the draft into a high-quality application is going to be hard work. The stakes and the rewards of the lottery are therefore high in the LLM-workflow. They are lower in the attorney-first workflow which relies on the attorney’s own skill in the beginning and then splits the work of the LLM into a series of smaller tasks where it is easier to revise or discard the LLM output after review.
Experience versus search power
When experienced patent attorneys review new inventions, they may recognize similarities to earlier cases which they either read about or handled personally. Based on this experience, they may (in the attorney-first workflow) decide to formulate the claim in a certain way or to include a particular passage in the description to forestall potential problems in the upcoming examination or a possible litigation.
An LLM which in the LLM-first workflow models its claim on earlier patent publications, or on earlier applications written by the attorney but chosen more or less at random, is unlikely to optimize the claim or the description in the same way. It lacks the knowledge of subsequent prosecution history which influenced the attorney’s judgment.
On the other hand, an LLM could be given access to a far broader range of case law and guidelines than the attorney could ever hope to memorize. This could include case law across jurisdictions where the attorney has never worked, any one of which could be selected as the primary focus based on the client’s filing strategy. This fountain of information may not have immediate relevance to specific terminological choices, but having the LLM review the application from the vantage point of a selected body of case law and guidelines may help in optimizing the application. Skilful prompting and attorney review will be needed to separate relevant information from irrelevant, but the benefits could be significant.
Consequently, the psychological experience of the attorney does not have to compete against the database-experience of the LLM. They can, in both the LLM-first and the attorney-first workflows, complement each other through dialogue.
Cautionary notes on AI-assisted drafting
In both workflows discussed above, there are two pitfalls which attorneys should be aware of. First, if a client wants to file just a single patent application (with no other patentable inventions forthcoming) the attorney can freely generate as much material as possible to the application text. But innovative clients usually make many inventions in the same field of technology. An application which is filed today might some day become damaging prior art for future applications.
The LLM can be prompted to list an almost unlimited number of options relating to the invention at hand. Some of the suggested options may today be speculative, not yet ripe for commercial use. But let’s say we decide to include them all in the description of application 1 and file it. What if the same client a few years later comes up with a new idea relating to one of the options we mentioned in application 1? It is now commercially viable. We file a new patent application 2 for this invention, but alas, application 1 is public and now constitutes prior art. It may not be novelty-destroying, but troublesome enough to force us to narrow the claim of application 2. The protection we obtain for the new invention in application 2 will then not be optimal.
Of course, the same risk exists if the attorney writes the application 1 manually. But most attorneys will not engage in a lot of speculative freestyling when they write an application. A patent application needs to explain and claim the invention while providing material for good claim amendments. Attorneys should keep this in mind also when reviewing LLM-assisted application drafts. Including everything is not necessarily a good idea. Skilful prompting is needed to entice added value from the LLM without sliding into excessive speculation.
The attorney should of course also consult the client to decide whether or not borderline LLM-generated options, which seem insightful but highly speculative, should be included or excluded.
The second pitfall is that we don’t want to make binding statements which could be used against the client in future court proceedings. Again, patent attorneys are trained to avoid them when writing manually, but LLM’s produce them regularly. There is a risk that an LLM-generated passage contains a statement we don’t want to make, and the attorney might miss it when reviewing the LLM output. This risk can of course be reduced by prompting the LLM to avoid statements of this kind and having it perform an extra check for potentially binding phrases. But the responsibility lies with the attorney.
Finally, currently practicing patent attorneys learned to draft applications without any help from LLMs. Should the next generation learn the trade through manual drafting or LLM-assisted drafting? I would advocate the former alternative, at least in the very beginning and for now. You only write your first application once and doing it yourself probably sets you on a steeper learning curve than prompting and reviewing would do. There will no doubt come a day when LLM-assisted patent drafting has become so routinized that a new trainee can jump straight into that world. But first, our generation must lay the foundations.
Conclusion
Anyone can nowadays generate a text which looks like a real patent application, using almost any input material. More and more, patent attorneys are likely to hear from clients who have produced their own application text with a large language model. But they may not recognize the limitations of the LLM and the low quality of its output. A 100% LLM-generated patent application would be a high-risk investment, more likely to fail than to succeed.
To improve the quality of their patent application drafts with LLMs for the benefit of their clients, attorneys can start by experimenting with the attorney-first workflow. It gives a clear and controlled understanding of the added value provided by the LLM. The LLM-first workflow promises greater efficiency gains but is harder to evaluate and more dependent on how skilfully the attorney uses the LLM. In these workflows and in any mixed workflow which could be imagined the attorney must take full responsibility for the final product.
I have explained why human expertise is still needed for high quality patent applications, without discounting the benefits that LLMs can bring to the table. To keep up with the times, patent attorneys should learn how to use artificial intelligence tools and keep their clients informed of the value that their human contribution still brings to the patent drafting process.
