The rise of the AI-assisted provisional
IP Australia's Australian IP Report 2026 records provisional filings rising from 4,343 in 2024 to 6,867 in 2025, with more than half of 2025 provisionals filed without a patent attorney. The report attributes part of that growth to digital tools, including AI, that make drafting faster and cheaper. It also notes the risks: more speculative filings, and questions about disclosure quality and inventorship.
The real question is not whether a tool was used to produce the words, but who made the decisions that determine whether the specification will hold up.
A provisional is tested on substance, not polish
Under section 40(1) of the Patents Act 1990, a provisional must disclose the invention clearly enough and completely enough for a skilled person to perform it. Each claim in your later complete application gets the provisional's date only if the provisional disclosed that claimed invention to that standard. We explain the consequences of falling short in the pitfalls of a poorly prepared provisional patent application.
None of this is checked when the provisional is filed. A well-formatted document that misses the point is accepted just as readily as one that gets it right. The problems only surface at examination or in litigation, when it is too late to fix them.
Where AI-only drafts go wrong
The first problem is that the output depends entirely on the input. An AI tool describes what it is told about. Most inventors, quite reasonably, explain their invention in terms of what it does and why customers will want it. The features that will matter in examination, the specific technical means that distinguish the invention from what already exists, are often the ones they don't think to mention. A tool that is never told about them can't describe them.
The second problem is plausible detail that isn't true. AI tools are good at filling gaps with technical content that sounds right: dimensions, materials, operating ranges, test results and alternative embodiments the inventor never considered or tried. A specification should describe what the inventor has actually devised and how it can genuinely be put into practice. Detail that turns out to be wrong, or that describes versions that don't work, can undermine the patent rather than strengthen it.
The third problem is length without support. An AI draft can run to many pages and still say little that is specific. Background material, generic statements of advantage and repeated summaries add length but provide no support for any particular claim. What matters is whether each feature you will eventually claim is described clearly, with its function, its implementation and its reasonable alternatives.
A specification that looks professional and a specification that will survive examination are not the same thing. Only the second one protects your invention.
The fourth problem is that the claims and description don't fit together. Effective drafting starts from the claims: deciding what the invention is, defining it broadly, and then setting up narrower fallback positions. The description is then written to support each of those claims. AI drafts are usually written the other way around, as a general description with claims either absent or bolted on at the end. We explain why the two must be drafted together in can you prepare a patent application yourself?.
The fifth problem is that nobody has looked at the prior art. Deciding what to emphasise in a specification depends on knowing, or anticipating, what already exists. An AI tool working from the inventor's description has no idea which features are routine in the field and which are genuinely new. It may give pages to features that are well known and a single sentence to the one that matters.
The sixth problem is inventorship. Australian law requires a human inventor. In Thaler v Commissioner of Patents, the Full Federal Court held in 2022 that an AI system can't be named as an inventor, and the High Court refused special leave to appeal. If an AI tool suggests features or embodiments that end up in your claims, questions can arise about who actually devised them. The inventive contribution needs to come from the people named as inventors.
How to use AI sensibly
AI can be genuinely useful to an inventor preparing for a patent application. It can help you organise your notes, describe how your prototype works, list the variations you have considered and prepare clear explanations of your drawings. All of that makes a better starting point for your attorney and can make the process faster.
What it can't do is take responsibility for the judgments that determine whether your priority date will hold: what to claim, how broadly, what detail is needed to support each claim, and how the specification will stand up against prior art an examiner hasn't found yet. Those judgments come from training and experience in patent law and practice.
How we prepare a specification
At Patentec we define the invention in a set of claims first, typically 20 or more from broad to narrow, based on a detailed discussion with the inventor. We then build a detailed description, usually 20 to 50 pages, around those claims, checking that every claimed feature has explicit support and that the technical content is accurate and genuinely reflects the invention. The result is a provisional prepared to the standard of a complete specification, ready for an international-type search and for the stages that follow, as set out in our approach to patent applications.
If you have drafted a provisional with AI, or are thinking about it, contact Patentec for a free initial consultation. We can review what you have, tell you frankly whether it is likely to hold up, and explain what a properly prepared specification would add.