In this article
Here is the situation, and I want to describe it accurately, because the accuracy is the whole point.
It is late. You have been at this for either nine hours or eleven hours — the distinction has stopped mattering — and somewhere in the middle of it the thing stopped being a vague irritation about how nobody has fixed the obvious problem with [the mount, the valve, the scheduling logic, the way the sensor fails in cold weather] and became, instead, a thing. A real one. It has parts. The parts have relationships to one another. You can see it.
You got here with help. You typed the irritation into a chat window and something typed back, and then you typed again, and around the fourth or fifth exchange the model said something you hadn’t thought of, and you thought oh, and then you thought wait, and then the two of you — and I am going to keep using “the two of you” for a moment because that is honestly how it felt, whatever the metaphysics — spent several hours refining it. Maybe you built it. Maybe there is a repo. Maybe the repo is public.1
And now it is Monday, or nearly, and you are experiencing the specific vertigo of having made something and not knowing what the next verb is.
Listen: the next verb is not iterate.
Yes, You Can Patent an Invention You Developed With AI
Short answer: using AI to develop an invention does not disqualify it from patent protection. That is a narrower statement than it sounds, and the narrowness matters.
I want to be unambiguous about the part that is unambiguous, because there is a genre of anxious internet content suggesting that AI involvement poisons a patent application, and that genre is mostly wrong. Then I want to be equally unambiguous about the four things this does not mean, because there is a competing genre — mostly written by people selling things, myself included — that overshoots in the other direction.
Under U.S. law an inventor must be a natural person. This was settled in Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), when the Federal Circuit considered whether an AI system called DABUS could be named as an inventor and concluded, essentially, that “individual” means a person — the kind of holding that seems obvious until you notice that every patent office on earth had to say it out loud. The Supreme Court declined to review it. The UK Supreme Court reached the same conclusion in December 2023, as did the EPO and Australia’s Full Federal Court.2
The follow-on question — fine, but does using AI wreck my patent? — was answered on November 28, 2025, when the USPTO rescinded its earlier guidance and replaced it with something considerably more relaxed. The February 2024 framework had required a natural person to make a “significant contribution” to each claim, analyzed under the Pannu factors. That is gone. The current guidance treats an AI system as an instrument — analogous, in the Office’s own comparison, to software or laboratory equipment — and applies the traditional conception test uniformly. If you are the only human involved, the Pannu analysis drops out entirely.
Did you conceive it? Probably. You probably conceived it at hour six, when you thought wait.
Four things that does not mean
1. It does not mean your patent will issue. Inventorship is one requirement among several, and it is not the hard one. Your invention still has to be new under §102, non-obvious to a person skilled in the field under §103, and directed to eligible subject matter under §101. That last one deserves flagging for this audience specifically: if what you built is software, subject-matter eligibility under the Alice framework is a live and genuinely difficult hurdle, and it sinks a meaningful number of software applications that are otherwise novel and well-drafted. Nobody can tell you how yours will fare without looking at it. Anyone who tells you otherwise before looking at it is not being straight with you.
2. It does not mean conception is automatic. The November 2025 guidance removed the Pannu-factor screen for solo inventors. It did not remove the requirement that a human conceive the invention — it just stopped applying a special test to find out. If your actual contribution was a two-line prompt and wholesale acceptance of what came back, the conception question is real rather than rhetorical, and it is a question about the specific facts of what you did. The honest version is: most people reading this conceived their invention and used a model to develop it. Some didn’t. You know which you are, and if you genuinely don’t, that uncertainty is itself worth raising with an attorney before you file.
3. It does not mean this is settled law. USPTO guidance is examination policy, not statute. The Office says so itself. Courts are not bound by it, Congress can legislate over it, and the Federal Circuit has not yet decided a case about a thin human contribution to a heavily AI-generated invention. Thaler answered whether a machine can be named as an inventor. It did not answer how much a human must do. That question is open, and a guidance document that swung once in November 2025 can swing again.
4. It does not mean you own it. This is the caveat most likely to actually affect you, and it has nothing to do with AI. Inventorship and ownership are different things. If you built this on a work laptop, on work time, in your employer’s field of business, or while subject to an invention-assignment agreement you signed on your first day and have not read since — your employer may own it regardless of who conceived it. The same goes for university IP policies and for contractor agreements with an IP clause. You can be the undisputed inventor of something you do not own. Go read your employment agreement before you file anything. Also relevant: if a co-founder or collaborator was in those conversations with you, the Pannu joint-inventorship analysis still applies as between the two of you, and getting the human inventorship wrong on an application is its own category of problem.
None of that is a reason not to proceed. It is a reason to proceed with an accurate picture, which is the only kind worth having.
What to do: keep your records. Chat logs, notes, the napkin, commit history, dated anything. Not because anyone is coming for you, but because the difference between a person who can demonstrate their own contribution and a person who cannot is a difference that occasionally matters enormously, is free to maintain in advance, and is impossible to manufacture afterward.
That is the good news, hedged about as much as good news can be hedged while remaining good. The rest of this is not good news.
Claude Cannot Run a Prior Art Search — and the Way It Fails Is Worse Than You Think
Short answer: web search is not patent search, and the errors are designed to survive your spot-check.
You have probably already asked it. Everyone does. You typed something like has anyone patented this and got back a fluent, organized, plausible answer, possibly with patent numbers in it, and felt a small relief that you may still be carrying around.
Put it down.
The obvious objection is the wrong one
“The model just makes things up” is true but insufficiently frightening, because a fabrication is at least findable. You click the link, there’s nothing there, you learn your lesson, you move on with a healthy new skepticism. Fabrication is a failure mode with a smoke detector attached.
The actual failure mode is subtler, and the definitive example is almost too on-the-nose to be real.
The Latham & Watkins incident
In May 2025, in a federal copyright case in the Northern District of California, lawyers at Latham & Watkins — a firm whose partners bill north of two thousand dollars an hour — filed an expert declaration on behalf of their client. The client was Anthropic. The product at issue in the litigation was Claude. An associate had asked Claude to format a legal citation for the declaration, and Claude had done so, and the citation was wrong.
But notice how it was wrong. Per the associate’s own declaration explaining the error, the article was real. The link worked. The volume number was right. The page numbers were right. The publication year was right. What Claude got wrong was the title and the authors — the two fields nobody re-reads once the link resolves. The Latham team clicked through. They reviewed the article. They corrected the volume and page numbers. They did not notice the title.
Opposing counsel noticed. As Fortune reported, drawing on Reuters’ coverage of the hearing, and as the Volokh Conspiracy documented, Magistrate Judge Susan van Keulen called it “a very serious and grave issue” and observed that there is “a world of a difference between a missed citation and hallucination generated by AI.” The associate called it “an embarrassing and unintentional mistake,” which it plainly was.3
So: the model, defended in court by the best lawyers money can buy, tripped those lawyers on the way into the courtroom. And it did it not by inventing something out of nothing but by getting a real thing almost entirely right.
What this means for your novelty check
Transpose that onto your prior art question. The model hands you a patent number. The number is real. The assignee is real. The date is real. The characterization of what that patent actually claims — the only part that matters to you, since claims are the fence and everything else is landscaping — is subtly, confidently, unverifiably off.
You spot-check it. It survives the spot-check. Everything survives the spot-check. That is what the spot-check is bad at.
And I should be precise here, because you are technical and will catch me if I’m not: the model can search the web. That is a real capability and I am not going to pretend otherwise. But searching the web is not searching patents. A prior art search means full-text search across USPTO and EPO databases, classification searching through the CPC hierarchy, non-patent literature, and then the slow unglamorous work of comparing claim language to your thing. A summary of web results is a categorically different object that has been dressed to look like the same object.
The real tools are free and unpleasant to use: USPTO Patent Public Search, Google Patents, Espacenet, PatentsView. Go be bored in them for an afternoon. Boredom is diagnostic here.
Check Your Claude Privacy Setting: The Five-Year Retention Problem
Short answer: on consumer plans, your invention disclosure may be retained for five years. It takes eleven seconds to check.
In August 2025 Anthropic updated its consumer terms. If you are on Claude Free, Pro, or Max — including when you use Claude Code from one of those accounts — there is a setting governing whether your chats and coding sessions are used to train new models.
| Setting | Data retention |
|---|---|
| Training on | 5 years |
| Training off | 30 days |
Commercial arrangements — Team, Enterprise, the API, Bedrock, Vertex, Gov, Education — sit outside this entirely and are not used for training under Anthropic’s commercial terms.
I don’t know what your setting says. A great many people clicked through the notification in a hurry in the autumn of 2025 and could not tell you today what they chose. It is changeable at any time. Go look.
What this does not mean
Now the necessary caveat, and I want to make it honestly rather than in the weaselly manner of someone with a product to sell: typing your invention into a chatbot is almost certainly not a public disclosure. It does not, by itself, destroy your novelty. A private conversation is not “available to the public” in the sense 35 U.S.C. §102 means. Anyone telling you otherwise is either confused or selling you something, and I am, admittedly, selling you something, which is exactly why I’m being careful about this one.
The risk is not disclosure. The risk is duration — the ordinary hygiene question of how long a complete technical description of your unpatented invention should sit on someone else’s infrastructure, and whether five years is your preferred answer.
Samsung’s Device Solutions division permitted ChatGPT internally on March 11, 2023. Within about twenty days there had been three leaks, including semiconductor source code. Generative AI was banned across company devices on May 1. So it goes. Amazon, JPMorgan, Goldman Sachs, and Citigroup all reached similar conclusions around the same time, and those are not organizations with a reputation for hasty caution.
If You Have Shipped Anything, a Twelve-Month Clock May Already Be Running
Short answer: public use and on-sale bars start a fuse that ends in permanent forfeiture — and the grace periods abroad are shorter, narrower, and easier to forfeit than the American one.
This section applies to you specifically, and did not apply to the version of this essay written for people who only talked about their inventions.
You built something.
Under 35 U.S.C. §102, certain events start a twelve-month clock that ends in permanent, unrecoverable forfeiture of your U.S. patent rights. The two relevant ones are the public use bar and the on-sale bar. The Supreme Court confirmed in Helsinn Healthcare v. Teva (2019) that even a confidential sale can trigger the on-sale bar, which is the sort of holding that makes patent attorneys go quiet at parties.
An incomplete list of things that can start that clock:
- Pushing the repo public
- Shipping a beta to users outside your organization
- Demoing at a hackathon, a meetup, a demo day
- The launch thread
- Taking pre-orders
- Pitching an investor without an NDA
Read that list as a list of things worth asking about, not a list of verdicts. Whether any given act actually triggers a bar is intensely fact-specific — it turns on what exactly was shown, to whom, under what obligations of confidence, and whether the invention was ready for patenting. A private pitch to one investor may well not be a public disclosure. A conference demo of a working unit very likely is. The point is not that you have already lost; it is that these questions have real answers that depend on details you can still remember today and won’t in a year, which is an argument for writing them down and asking someone now.
The part that is genuinely unfair
The United States gives you a twelve-month grace period after your own disclosure. The rest of the world is stingier, and in the most important market, effectively merciless.
Europe is the strict one. EPC Article 54 defines novelty against everything made available to the public before your filing date, and Article 55 carves out only two narrow exceptions — disclosure resulting from an evident abuse against you, and display at certain officially recognized international exhibitions. Your launch thread is neither. For practical purposes, if you disclosed before filing, Europe is gone.
Elsewhere there are grace periods, but do not mistake them for the American one. Japan provides twelve months under Article 30 of its Patent Act — but you have to claim it, with a statement filed at the time of application and supporting documentation within thirty days. Miss the procedure and the substantive protection evaporates. China’s Article 24 period is six months and applies only in specified situations, such as certain government-recognized exhibitions, prescribed academic meetings, and disclosures made without your consent. A launch is not on that list either.
So the accurate version is not “you lose everything abroad.” It is: the safety net is shorter, narrower, conditioned on paperwork most inventors don’t know exists, and in Europe essentially absent. That is a worse situation than most people assume and a better one than the internet usually claims.
Which produces a genuinely routine outcome: an inventor discloses publicly, remains perfectly fine in the U.S., and has quietly forfeited Europe and complicated everywhere else — without a single person telling them, at any point, that this was happening.
Also, and separately: under the America Invents Act the U.S. is first-inventor-to-file. Not first to invent. If someone else files before you, the fact that you thought of it first is a matter of purely private significance.
Why your idea is likelier to have a twin than you think
There is something you should know about how you got here.
In July 2024, researchers Anil Doshi and Oliver Hauser published a study in Science Advances finding that generative AI makes individual people’s ideas more creative — genuinely, measurably, especially for less-creative people — while making the population of ideas more similar to one another. Everyone gets better. Everyone gets better in the same direction. Subsequent work has confirmed the clustering.
The implication is not comfortable. You did not have a private conversation. You had a conversation that thousands of people with similar irritations and similar prompts are having with similar models, and the models are pulling all of you toward the same neighborhood of solution space. Your idea is more likely to have a twin than it would have been five years ago, not less.
So it goes. But it goes faster now.
What a Provisional Patent Application Actually Does — and What It Costs
Short answer: it buys a filing date and twelve months for $65 to $325. It is the correct next move. It is also easy to do uselessly.
A provisional patent application is a strange instrument, and most descriptions of it are either too flattering or too dismissive.
What it is: a filing that establishes a date. Per the USPTO’s own description, it is never examined. It is not published on its own — though be careful with that fact, because if you later claim priority to it, the provisional becomes part of the file history and generally becomes publicly available when that later application publishes. It is not a vault. Don’t put anything in it you intend to keep as a trade secret forever.
It confers exactly zero enforceable rights on its own. It expires after twelve months, automatically, and cannot be extended. During those twelve months you may say “patent pending,” which is legally meaningless and commercially useful, a combination that describes a surprising amount of the world.
What it costs, under 37 CFR 1.16(d):
| Entity status | Provisional filing fee |
|---|---|
| Micro entity | $65 |
| Small entity | $130 |
| Large entity | $325 |
Most independent inventors qualify as micro entities. I checked these against the USPTO fee schedule on the day I wrote this — it was last revised July 1, 2026, and the provisional fee did not move.
Sixty-five dollars. This is a real number and it is why everybody tells you to file a provisional.
Now the part nobody tells you.
Why DIY Provisional Patents Fail: The §112 Problem
Short answer: a provisional protects only what it adequately describes. A thin one is worse than none at all.
Under 35 U.S.C. §112, your filing must contain a written description of the invention sufficient to enable someone skilled in the field to make and use it. When you later file the real application and claim priority back to your provisional date, that date holds only for claims your provisional adequately supported.
The Federal Circuit has been unsentimental about this for decades. In New Railhead Mfg. v. Vermeer Mfg., 298 F.3d 1290 (Fed. Cir. 2002), the court held the provisional’s specification had to describe the claimed invention in full, clear, concise, and exact terms — and the priority claim failed. In Dynamic Drinkware v. National Graphics, 800 F.3d 1375 (Fed. Cir. 2015), the court reaffirmed that a provisional confers a priority date only where it provides §112 support. The pattern in those cases is the same and it is grim: an inventor believed they had a priority date, built a year of decisions on that belief, and discovered in litigation that they did not.
Which produces the actual hazard, and it is not the one you’d guess:
A thin provisional is worse than no provisional at all.
Not because it costs more. Because it produces confidence. The inventor who files nothing knows they are exposed and behaves accordingly — cautiously, with NDAs, quietly. The inventor who files three pages of vague description believes they are covered, and proceeds to demo, pitch, post, and publish on the strength of that belief, and finds out at the exact moment when nothing can be done about it.
The numbers, and what they do and don’t show
An analysis of USPTO filing data found that applicants representing themselves abandon at a rate of about 76%, against roughly 35% for applicants with professional representation. Between 40% and 48% of provisionals are never converted to anything at all.4
I want to be careful about what that proves, because it is less than it appears and I would rather tell you than let you assume.
Abandonment is not the same thing as a §112 failure. People abandon applications because the company folded, the money ran out, the market moved, or the idea stopped seeming good in daylight — none of which is a drafting problem. Nobody publishes a statistic on how many provisionals were technically inadequate, because that determination generally only gets made in litigation, and the overwhelming majority of provisionals never come near a courtroom. So the honest characterization of that 76/35 gap is: it is consistent with a drafting-quality story, and also consistent with a selection story where people who hire attorneys are the ones with more at stake. It is suggestive. It is not proof.
What is established, and doesn’t depend on any statistic, is the legal mechanism. §112 support is required. Priority holds only where support exists. The Federal Circuit has enforced this repeatedly. That is the part you should act on.
And the mechanism behind the gap is not mysterious either. Writing a patent application is a specific, learnable, deeply unnatural skill — the entire discipline consists of describing your invention broadly enough to cover what a competitor would build to get around you, while describing it specifically enough to satisfy §112, and doing both simultaneously in a document you will not be permitted to add anything to later. Ever. Not one sentence.
Your code is not a disclosure
I say this because I suspect the thought has already occurred — I have a working implementation and a thorough README, surely that’s the description.
It isn’t. Working code demonstrates one implementation. §112 asks for a description of the invention, which includes the alternatives you didn’t build, the variations you rejected, the range of the thing. A repo dump is a photograph of a single point in a space you are trying to claim.
What Patent Pending Made Simple Costs, and What You’re Actually Buying
This is an advertisement.
You knew that. I knew you knew. There is a particular insult in pretending otherwise for two thousand words and then producing a button at the bottom as though it were a natural consequence of the reasoning, so let me just tell you what I’m selling and you can decide.
Patent Pending Made Simple is software that walks you through writing a provisional. It asks you structured questions — the ones a patent attorney would ask, in the order they’d ask them, including the deeply annoying ones about how a competitor would design around you, which is the question most inventors have never asked and cannot answer on the first attempt. Then it drafts the document. You file it, or we file it.
For a concrete example of how a general-purpose AI conversation differs from that patent-specific process, see the PatentPending.ai vs. ChatGPT workflow comparison.
| Tier | Price | What’s included |
|---|---|---|
| Starter | $599 | Software only; you review and file |
| Expert | $1,399 | Adds a strategy call and attorney review |
| White Glove | $1,599 | Adds paralegal preparation and filing |
USPTO fees ($65–$325) are separate. There is a money-back guarantee.
The honest comparison: attorney-drafted provisionals are commonly quoted starting around five thousand dollars, though published figures vary by firm, technology, and complexity, and the AIPLA’s biennial Economic Survey is the closest thing to an industry benchmark. So this is cheaper than the typical attorney-drafted route. It is not free, and the free option — you, a text editor, a weekend — remains available and is the one that carries the risk described two sections up.
What you’re actually buying is the sequence of questions. That’s it. That’s the product. The drafting is the easy part; knowing what you failed to say is the hard part, and the hard part is what a decade of drafting these things teaches you and what a blank document conspicuously does not.
The disclaimers, which are real
This is not legal advice, and reading it does not make anyone your lawyer. No one can promise you a patent will issue. Anyone who does promise you that is running a different kind of business — and you should know that the FTC has documented invention-promotion schemes taking inventors for five to twenty thousand dollars and up, and that the USPTO maintains a public complaint forum specifically because this industry needed one.
The Order of Operations
Build. Then search — properly, in the boring databases. Then file something that would actually hold. Then ship, pitch, post, and tell people.
That’s the entire essay. Everything above is elaboration on the ordering.
The reason the ordering matters is that three of those four steps are reversible and one is not. You can rewrite the code. You can redo the search. You can amend, refile, continue, and convert. What you cannot do is un-disclose. Once the thing is public it is public in every jurisdiction at once and permanently, and the safety nets that exist are shorter than the American one, conditioned on procedures you have to know about in advance, and in Europe barely there at all.
You made something at hour six on a Saturday. It’s yours. It has been yours the whole time.
Go find out what your privacy setting says, and then let’s write it down properly, while writing it down still does something.
Footnotes
-
If the repo is public, skip to the section on statutory clocks, which is about you, and then come back. I’d rather you read this out of order than not get to it. ↩
-
The DABUS applications, for what it’s worth, covered a food container based on fractal geometry and a flashing emergency light. A patent office in South Africa granted one of them, which sounds like a dramatic outlier until you learn that South Africa does not substantively examine patent applications, so the grant means approximately what a receipt means. ↩
-
I am aware of the recursive problem here — that this essay is in part about the unreliability of a system, and that you have no particular reason to trust my account of the incident rather than checking it. Please check it. The declaration is on the public docket in Concord Music Group, Inc. v. Anthropic PBC (N.D. Cal.), and the coverage is linked above. This is, I think, the correct relationship to have with any confident summary of a thing you did not witness, including this one. ↩
-
These figures come from third-party analyses of USPTO filing data rather than a single official USPTO statistic, which is the sort of thing I would want disclosed to me, so I’m disclosing it to you. ↩
- Claude invention
- AI-assisted invention
- prior art search
- disclosure risk
- provisional patent application