I use artificial intelligence to help me draft patent applications. I also use it when responding to office actions, refining claim language, and summarizing prior art. I want to be upfront about that — AI is part of my practice, and it genuinely saves time.
But here is what I have learned after working with these tools: AI is useful precisely to the extent that I tell it what to say. Left to its own, it produces something that looks like a patent application but isn’t. And that distinction matters enormously — especially if you’re an inventor thinking about writing your own with AI.
AI Is a Tool. That’s It.
The best analogy is a word processor. Microsoft Word is an extraordinarily powerful tool. It catches typos, formats documents, checks grammar, and can even suggest rewrites. But Word doesn’t make you a grammarian. It just makes your brief look clean on the page.
AI is the same. It can reformulate your ideas into polished, professional-sounding language. It can generate boilerplate descriptions of technology, pull together explanations of how systems work, and produce a document that has the shape and feel of a patent application. For someone who has never written a patent before, that output is convincing.
That is exactly the problem.
AI Doesn’t Know What You Don’t Know
There’s a phrase I keep coming back to when clients ask me about this: AI doesn’t know what you don’t know.
What I mean is this. When I draft a patent application, I’m not just describing an invention. I’m making strategic decisions that affect the scope of protection for years, sometimes decades. I’m thinking about what the prior art looks like, how a USPTO examiner is likely to approach the claims, where the narrowest defensible claim sits, and what claim language might be exploited by a competitor down the road. I’m thinking about continuation strategy and when to file a full application, how this application fits into a broader portfolio, and whether the specification needs to be drafted broadly enough to support future claims that the inventor hasn’t even thought of yet.
None of that thinking is written down anywhere in a way that an AI can access. It lives in the heads of patent attorneys who have spent years before the USPTO — prosecuting applications, getting rejections, arguing with examiners, watching patents get invalidated in court, and learning from each of those experiences. That knowledge isn’t on the internet. Most of it never will be.
It doesn’t know what you don’t know and as such, it might send you down the wrong route.
AI is trained on what’s publicly available: issued patents, published applications, legal treatises, practitioner articles. It’s very good at sounding like a patent attorney. It is not a person that can draw the right information out of you.
What AI Can’t Help Doing: Hallucinations
There’s a property of every AI tool currently on the market — including the ones I use — that I want you to understand before you trust one with a patent application. It’s called hallucination, and it’s not a bug that will be patched in the next release. It’s a feature of how these systems work.
A large language model doesn’t know things. It predicts the next plausible word given its training data. When the model encounters the edge of what it was trained on, it doesn’t say “I don’t know.” It produces confident, plausible, and sometimes completely wrong output.
In a casual email, a hallucination is an annoyance. In a patent application, it can be catastrophic.
AI fabricates prior-art references that don’t exist. If you rely on an invented citation in your background section, you’ve introduced a misrepresentation into a document you will sign under penalty of perjury on a USPTO declaration. AI fabricates case citations. Mayo v. Prometheus is a real decision — but AI will happily generate a citation to a nonexistent Federal Circuit case and paraphrase a fake holding that sounds reasonable. AI invents technical details. If the specification says the mechanism operates at 150°C and the actual invention fails above 90°C, enablement under 35 U.S.C. §112 is broken the moment the issue is raised. AI uses claim terms not defined in the specification. An examiner catches this immediately as a §112(b) indefiniteness problem.
Patent attorneys can usually spot these errors because of the experience of drafting hundreds of applications and reading thousands of office actions. The patterns are familiar. They will not be familiar to you. That’s what makes AI’s output so dangerous in DIY hands — it looks right, right up until it doesn’t.
The Strategy Gap Is Bigger Than It Looks
I write articles for this blog not about what inventors need to know, but about what they want to know. People don’t read articles about things they don’t already find interesting. That means most of the technical craft of patent prosecution — the kind of insight that comes from handling hundreds of office actions and learning how to read an examiner’s pattern of rejection — never gets written up anywhere. I don’t write about things you need to know. It’s not interesting enough to publish. I write about things you want to know.
Claims strategy is a perfect example. A well-drafted independent claim is neither too broad nor too narrow. Too broad, and it gets rejected or later invalidated. Too narrow, and a competitor designs around it easily. Finding that middle ground requires understanding the specific technology, the specific prior art landscape, and how similar claims have fared at the USPTO. AI will produce claims. It will not find that middle ground on its own, because finding it requires judgment that isn’t derivable from any published text.
The same is true of responding to office actions. When I respond to a rejection, I’m not just arguing the legal standard — I’m reading the examiner’s specific objections, thinking about what amendment will open the least number of new vulnerabilities, and deciding whether to argue or amend, and how hard to push. That calibration is experiential. An AI can draft the response. It cannot make those calls.
Inventors Understand How; They Don’t Understand What the Inventive Concept Is
There’s a distinction that underlies everything I’ve just described, and I want to state it plainly:
Inventors understand how their product works. They don’t understand what the inventive concept is.
Those are different things. Knowing that your product works — how the parts fit together, what the mechanism does, what the prototype does in testing — is the engineering understanding. Knowing the inventive concept is something else. It’s understanding which specific relationships between features make the product novel; which features, expressed at which level of abstraction, give you claim scope that a competitor can’t design around; which aspects are central and which are incidental; and how the inventive step survives the art a USPTO examiner is going to cite against you.
That understanding doesn’t usually exist before drafting begins. It emerges through the drafting process itself, with a patent attorney who has been through that process many times. The inventor has the raw material. The attorney finds the shape. AI can polish text. It cannot find the shape of a claimed invention from a technical description, because the shape isn’t in the description. It’s in the relationship among features, interpreted against prior art and claim construction.
What Does Your Invention Feel Like?
Here is a question I ask inventors early in a drafting engagement: what is it like to actually use the thing you invented? Not the mechanism. The experience. What does the user notice first? What do they notice ten minutes in? What makes them keep using it tomorrow instead of going back to whatever they used before?
Those answers, more often than you might expect, reveal the real claim scope. A new kitchen tool’s novelty might turn out to live not in the parts list but in the angle between the handle and the blade that produces a wrist-neutral grip. A new fastener’s novelty might live not in its geometry but in the acoustic feedback it gives the user when it’s fully seated. A new medical device’s novelty might live not in its mechanism but in the tactile cue that tells the clinician it has reached the right depth.
AI cannot ask that question meaningfully. AI has never held anything. The inventor can answer the question, but rarely knows they need to — they’re too close to their own design to see it. This is the human dimension of patent drafting, and it is not going away.
The Quality Illusion
Here is the real danger, and I want to be direct about it: AI-generated patent applications look good. The language looks professional. The format is correct. The descriptions are coherent. An inventor who has never seen the inside of a patent prosecution file has no way of knowing that the claims are too narrow, that the specification doesn’t adequately support the broader embodiments, that the disclosure enables a competitor to work around the key inventive concept, or that a subtle choice of words will haunt them in litigation years from now.
The output looks like a patent application. But appearances in patent law are particularly dangerous. A patent that gets granted on an AI-drafted application without the proper guidance isn’t necessarily a valuable patent. It might have a scope so narrow it protects nothing, or enablement issues that make it vulnerable to challenge the moment it matters.
I sometimes think of it as a beater car with a new paint job. The paint job looks good from far away. Come close and its still a beater car.
What AI Is Genuinely Useful For, In My Opinion
I don’t want to leave the wrong impression. Within a professional practice, AI is a real tool with real value. I use it to get a first draft of a detailed description on the page faster. I use it to help reorganize how an inventor explained their invention to me. I use it to check whether my claim language is internally consistent and to quickly survey how similar claims have been drafted in related applications.
But in every case, I’m directing it. I know what the finished product needs to say before I ask AI to help me say it. I’m not asking AI what the patent should cover — I already know that. I’m asking it to help me express what I’ve already decided.
That’s a completely different use case from an inventor who opens an AI tool with no prosecution experience and asks it to write a patent application from scratch. In the first case, AI is accelerating the work of an expert. In the second, it’s producing the appearance of expert work with none of the substance behind it.
If You’re Going to Use AI, Use It Here: Invention Disclosure
Despite everything I’ve said, AI has a legitimate and genuinely useful role for the solo inventor or small-business owner thinking about patent protection. That role is not patent drafting. It is invention-disclosure preparation.
A great invention disclosure is the single most useful artifact you can produce about your own invention. It documents your conception. It forces you to articulate your technical advantages. It surfaces embodiments you hadn’t thought of. It has independent legal value as evidence of conception and reduction to practice — whether or not you ever file. And if you do come to me or another patent attorney, a thorough disclosure changes what we can do in the drafting process, because 80% of the quality of a patent application traces to the quality of the disclosure it’s built from. For the full afternoon-long method I recommend, see How to Write a World-Class Invention Disclosure With AI in One Afternoon.
AI is good at helping you produce a thorough disclosure. Here are a few prompts you can use to get started:
Prompt 1 — Problem statement.
I’m an inventor preparing an invention disclosure for my patent attorney. Here’s my invention: [describe]. Ask me 15 questions, one at a time, that will help me articulate (a) the specific problem my invention solves, (b) what people currently do instead, and (c) why those current approaches fail. After each of my answers, ask a sharper follow-up before moving to the next question.
Prompt 2 — Technical advantages.
Explain — in precise technical terms — why my invention is likely to work better than [the closest conventional approach]. Identify the specific physical, chemical, mechanical, or computational principles that give rise to the advantage. Be conservative: if an advantage is speculative, say so.
Prompt 3 — Prior art recall.
I’m going to describe my field and invention. Based on what I say, ask me about any prior art, prior products, prior patents, prior academic papers, or prior public disclosures I might be aware of but haven’t mentioned. Do not tell me about prior art you “think” exists — I need my memory of what I know, not your guesses.
Prompt 4 — Disclosure-gap review.
Review the disclosure I’ve written so far: [paste it in]. Identify 10 gaps, ambiguities, or areas where a patent attorney would likely ask follow-up questions. Don’t try to fill them — just identify them so I can address them before my attorney meeting.
About the Craft I’m Not Sharing
I gave you a handful of starter prompts above. They’re worth using. But I want to be direct about what I actually do when I draft a patent application with AI — because it’s the honest answer to the question, “why couldn’t I just do what you do?”
I don’t work from a library of refined prompts. What I do is closer to a conversation. I draft a passage. I ask AI to try a particular framing. I read what it produces, see what’s off, tell it what’s wrong, ask it to rework a specific phrase, push back when the rewrite misses the point, and keep going until the passage says what the invention requires. The next passage is a different conversation, guided by different considerations. I’m still drafting the application. I’m just not typing every word letter by letter like we used to.
What isn’t captured in a prompt — and what separates my work from AI’s — is the real-time judgment behind that conversation. Knowing when a claim limitation has crept in too narrow. Knowing when a specification passage won’t support a later claim amendment. Knowing when a rewrite reads well but weakens the enablement case. Knowing what to push back on, and how hard. That judgment comes from years of prosecuting applications and reading office actions — not from a prompt list. It isn’t publishable because it isn’t a document. It lives in the directing.
The practical implication for you: the starter prompts above will help you produce a stronger invention disclosure. They won’t produce what I produce when I draft. If you want drafting-grade output, you need drafting-grade judgment steering it — and that only happens inside an attorney engagement.
What I’m Not Saying
I want to be careful here. I am not claiming that a patent attorney produces perfection. No one does. A patent drafted by a trained attorney can still face rejection, require amendment, or end up narrower than hoped. We don’t guarantee outcomes.
What I am saying is that the work product of a trained attorney is materially better than what an inventor produces from an AI tool on their own. Not perfect — better.
The honest choice in front of you isn’t “perfect patent or no patent.” It’s this:
- Option 1: Invest in attorney-drafted patent protection.
- Option 2: Accept that your AI-assisted application may have flaws you will need to accept.
Both paths are legitimate. For many solo inventors the right answer is Option 2 for now — the invention may not be commercially developed enough, the market may not justify the spend, the money may not be there. What doesn’t exist is Option 3 — “I’ll DIY with AI and pay an attorney a few hundred dollars to clean it up.” The structural problems in an AI-drafted application aren’t surface-level. Fixing them isn’t editing — it’s re-drafting, which costs the same as drafting from scratch.
The Bottom Line
AI will not guide you through the patent process. It doesn’t know which claims to fight for and which ones to let go. It doesn’t know your competitive landscape well enough to make filing strategy decisions. It doesn’t know what it doesn’t know — and neither will you, if you’re relying on it without the background to evaluate what it gives you.
The technology will keep improving. Where AI is in ten years is genuinely hard to predict. But right now, in April of 2026, the thing that makes a patent application worth having is not how it looks on the page. It’s the judgment behind it.
If you’re an inventor wondering whether AI tools can replace working with a patent attorney, my honest answer is: not yet, and not in the ways that matter most. A good patent is an investment. The value is in the protection it actually provides — not the document you file.
Questions about protecting your invention? Contact our office today to discuss your patent strategy.