Using AI to Draft a Patent Application: Legal & Commercial Considerations for Inventors
Relying on AI to draft a patent application, even at the provisional stage, introduces risks that are effectively ticking timebombs for your patent rights. While AI tools are often promoted as a way to democratise expertise, they are not a substitute for the legal, technical, and strategic rigour required in patent preparation. Missteps at this stage can irreversibly compromise intellectual property rights. This article focuses on the legal deficiencies that arise from AI-assisted drafting, rather than the more commonly discussed issues of hallucinations or confidentiality.
Completeness vs Substance
We are now seeing an increasing number of disclosures, supposedly “ready to file” applications and applicants that have prepared and filed their own provisional applications and are now seeking our assistance with material that is clearly generated largely through AI platforms. These documents often look polished and complete, but you cannot judge a book (or a patent application, as may be the case) by its cover.
At the core of the issue is disclosure. The patent system rewards applicants based on what they have actually invented, not what they hope to achieve. This requires a clear and complete description of how the invention works in practice. The specification must enable a skilled reader to perform the invention without undue experimentation. High-level descriptions of results, benefits, or intended functionality are not enough.
AI-assisted drafting tends to fall short in a consistent way. It produces language that is fluent and convincing, but often abstract. The focus is placed on what the system does rather than how it does it. The result is a specification built around outcomes, buzzwords, and generalised functionality, with insufficient technical detail to support those claims. Key steps are omitted, alternatives are underdeveloped, and the disclosure assumes knowledge that has not been properly articulated.
This problem is exacerbated by a common misunderstanding about provisional applications. There is a tendency to treat provisionals as informal placeholders, something that can be filed quickly and refined later. In practice, this assumption creates problems because a provisional application only establishes a priority date for what it properly discloses at the time of filing. If critical features, variations, or implementation details are missing or inadequately described, you cannot rely on that earlier date for those aspects at a later stage.
This becomes particularly important during examination. If something was not described with sufficient detail at filing, it cannot later be added to strengthen the claims. Patent offices are strict, and becoming stricter all the time, when it comes to the introduction of new matter. If it was not included in the application at the time of filing, it cannot be added later. In those situations, either the claim is invalid for claiming something that the application does not support, or it must be narrowed to what was actually disclosed. In many cases, this results in protection that does not align with commercial objectives. The fact that the original application was prepared quickly or with the assistance of AI does not attract any leniency.
Even where AI-generated application progresses through initial stages, it often lacks strategic depth. A well-prepared specification is drafted with future claims in mind. It deliberately includes variations, fallback positions, and different levels of generality to preserve flexibility during prosecution. AI does not engage in that type of forward planning. It does not identify which aspects of the invention are likely to carry commercial value, nor does it structure the disclosure to support a claim strategy with adequate breadth. The result is often an application that locks the applicant into a narrow or poorly supported position at the earliest stage.
It is also important to understand that grant does not immediately equate to strength or commercial value. Patent offices assess applications against legal requirements, not commercial usefulness. As a result, even a poorly prepared application may proceed to grant, particularly if the claims are narrow or the deficiencies are not fully tested during examination. However, the real test comes later, during enforcement or commercial due diligence. At that point, omissions in the original disclosure, lack of technical detail, or poorly supported claim scope can significantly limit the ability to enforce the patent or to assert it against competitors. In some cases, the granted claims may be so narrow, or so disconnected from the actual product, that they offer little practical protection or value despite the patent being in force, and this point leads us to commercial considerations.
Decisions, Risks, and Commercial Consequences
Many of these risks extend beyond the patent office. Patent rights are often relied on in discussions with investors, partners, or licensees. In those situations, the strength of the underlying application will be tested through due diligence, usually by a patent attorney acting for the other party. An application that has been self-filed or appears to be heavily AI-generated will almost always attract closer scrutiny.
From a due diligence perspective, this is ultimately an issue of risk. If corners appear to have been cut at the drafting stage, the reliability and scope of the resulting rights are uncertain and can become a point of concern at critical stages of enforcement or commercial negotiations. An experienced patent attorney will look for indicators such as lack of technical detail, inconsistent terminology, and absence of fallback positions. These issues are often not difficult to identify, and they can materially affect whether a transaction proceeds at all.
Ultimately, the issue is not that AI produces poor writing. In many cases, it produces very good writing. The problem is that patent drafting is not primarily a writing exercise. It is an exercise that combines legal, commercial and strategic factors which requires careful consideration and balance of disclosure, scope, and future positioning. AI, in its current form, does not perform that function.
Based on our observations to date, treating AI to draft a patent as a substitute for a patent attorney, particularly at the provisional stage, is a high-risk approach. The consequences of getting it wrong are not easily fixed and, in many cases, cannot be fixed at all.
If you have queries about the drafting and prosecution of patents, please contact us at mail@mbip.com.au or via our online form.
