So you think you have an AI invention. Do you really?
The rules for patenting AI inventions in the U.S. shifted meaningfully in late 2025 and early 2026 in a direction that favors inventors. But the rules are specific, and they reward people who think about patentability while they’re building, not after.
If you forget everything else in this article, let this be your one takeaway. The USPTO will now grant patents on AI inventions that improve how the AI technology itself works. It will still refuse patents on AI inventions that just apply existing AI to a new problem.
Here is the same takeaway, in a real-life scenario: imagine explaining your invention to a senior ML engineer at another company. If their reaction would be “okay, so you trained a standard model on your industry’s data”, you’re in the danger zone. If their reaction would be “huh, that’s a clever way to make the model itself behave differently”, you’re in the patentable zone.
If you’re still finding hard to wrap your mind around the “how to develop an AI invention that improves how the AI technology itself works”, don’t worry. What follows is a step-by-step process on how to think about patentability while you are developing your AI concept.
STEP 1: REFRAME YOUR PROJECT AROUND A TECHNICAL PROBLEM
Most engineers and founders naturally describe their work in terms of business problems: “We help logistics companies predict delays.” “We help radiologists catch cancers earlier.”
Patents don’t work that way. The USPTO wants to see a technical problem with current AI or computing technology, AND a technical mechanism that solves it.
So before you start any patent conversation, translate your project into technical-problem language. Two examples:
Example 1.
Business framing
“We predict equipment failures from sensor data.”
Technical framing
“Standard recurrent models lose accuracy on irregularly-sampled multi-sensor data because of [specific mechanism]. We modified [specific component] to handle this by [specific approach].”
Example 2.
Business framing
“We summarize legal documents.”
Technical framing
“Standard transformer attention has quadratic memory cost on long documents. We developed a sparse attention pattern that exploits [specific structural property of legal text] to achieve linear scaling while preserving accuracy on [specific tasks].”
Notice that the technical framings aren’t just a collection of fancy words. They’re more specific about what technical thing was broken and what technical thing fixes it. That specificity is what patents protect.
If you can’t write a technical framing of your invention, that’s a signal. Either dig deeper into what’s actually novel about your technical approach, or accept that this particular work may not be patentable (which is fine: trade secrets, copyright on code, and first-mover advantage may serve you better).
STEP 2: FIND OUT WHERE YOUR TECHNICAL NOVELTY ACTUALLY LIVES
Sit down with your technical team and answer this question concretely: if a competitor stripped away everything generic and used only what’s genuinely new about our approach, what would be left?
The answer is your invention. Be ruthless. The following things are almost never the answer:
“Our dataset” (datasets aren’t patentable inventions)
“Our prompts” (prompts to a general-purpose LLM are very hard to patent)
“Our UI”
“Our integrations”
“How well our model performs on benchmark X”
The answer should be something like:
“A novel loss function that prevents [specific problem]”
“A specific way of structuring the model’s memory across inference steps”
“An architecture in which [component A] feeds [component B] in a way that solves [problem]”
“A training procedure that achieves [result] by doing [specific thing]”
If you find more than one such answer, you potentially have more than one patentable invention. Track them separately.
STEP 3: BUT DID YOU DOCUMENT IT?
The November 28, 2025 USPTO inventorship guidance is clear: AI is a tool, and a human has to be the inventor. This affects how you should document your work day-to-day, especially if your team uses LLMs or AI coding assistants heavily.
What to capture, in a lab notebook, a structured doc, emails or anything really that keeps verifiable track of dates:
The technical problem the human team identified, and when
The hypotheses the humans formulated about how to solve it
The approaches the humans chose to try (and why)
The prompts, configurations, or guidance humans gave to any AI tools they used
The outputs the AI produced and how the humans evaluated, selected, or modified them
The follow-on refinements the humans made
You don’t need court-grade records. You need contemporaneous evidence that humans were doing the conceiving, with AI as a tool. A timestamped Word doc with periodic updates is far better than nothing.
A trap to avoid: if you also file patents internationally, never list AI as an inventor on a foreign filing, because the U.S. application that claims priority from it will be rejected.
STEP 4: SO WHAT IS THE TECHNICAL IMPROVEMENT. REALLY, WHAT IS IT?
This is the step most people skip, and it matters enormously when the patent application is drafted.
For each piece of technical novelty you identified in Step 2, your or your team should be able to answer four questions clearly:
What was wrong with the prior state of the art? (Be specific. Not “models were slow.” Consider “memory cost scaled quadratically with sequence length, which made [class of applications] impractical.”)
What’s the mechanism of your improvement? (How does it actually work, not just what it achieves.)
What’s the measurable result? (Memory reduced by X%. Speed increased by so many Tokens Per Second. Error rate down Z points on benchmark W.)
Why does the mechanism produce the result? (Why does it work? The causal story tying the technical choice to the measurable outcome.)
If you can answer these four questions, your patent attorney can write a specification that satisfies the current USPTO guidance. If you can’t, you have technical work to do before you have a patent application worth filing.
A practical tip: run an internal “patent disclosure” review for each candidate invention before engaging counsel. Have the inventors write up the four answers above in a 2–3 page document. This dramatically improves the quality and reduces the cost of the patent prosecution process.
STEP 5: TIMING IS EVERYTHING
Patents protect what you’ve invented at the time of filing. Not an AI model that might be patentable once you are done tinkering with it. The new USPTO posture rewards inventions that are concrete and demonstrated, not aspirational.
Some practical timing rules:
File when the technical mechanism is real, not when the idea exists. A working prototype with measurable improvements over a baseline is enormously stronger than a concept.
File before public disclosure. Showing the invention at a conference, in a blog post, in a paper, or in a sales demo can start clocks and create prior art. Once you’ve publicly disclosed in the U.S., you have a one-year grace period to file, but most other countries give you no grace period at all. Public disclosure can permanently destroy your foreign patent rights.
File before launching publicly if international protection matters to your business.
Consider a provisional application if you have a real invention but need more time to refine it. A provisional gives you 12 months to file the full application while locking in your priority date.
STEP 6: DO YOU KNOW ANY EXPERTS IN YOUR FIELD? KEEP THEM CLOSE.
Even with the favorable new posture at the USPTO, some applications will still receive rejections claiming the invention is an “abstract idea.” There’s now a procedural tool for fighting back: the Subject Matter Eligibility Declaration (SMED).
In plain terms, a SMED lets you submit expert evidence. That usually looks like a written declaration from a qualified technical expert, explaining why someone skilled in your field would understand your invention as a real technical improvement, not just abstract math.
You don’t need to file a SMED proactively. But you should think early about who could serve as a credible declarant: a respected researcher, an engineer with deep expertise in your area, an academic. If you build relationships with such people during your inventive process (advisory boards, collaborations, co-authored papers), you create options for later.
A SMED with a credible expert behind it can turn around a § 101 rejection that would otherwise have killed the application. It is not a sure way to overcome a § 101 rejection but it can make a difference for many patent applications in the AI field.
COMMON MISTAKES. DON’T DO THAT. PLEASE DON’T.
Pitching the business problem instead of the technical mechanism. Your patent attorney can only work with what you give them. If your invention disclosure reads like a sales pitch, the resulting application will too.
“We use AI to...” claims. Almost any claim that starts this way “we use AI to identify fraud,” “we use AI to schedule deliveries” is in trouble. Generic AI applied to a domain isn’t an invention. The invention has to be in the how.
Confusing performance with invention. Beating a benchmark is not the same as inventing something patentable. Many teams have state-of-the-art performance (2,900 - 3,100+ TPS for example) with no patentable invention behind it.
Treating prompts as inventions. Prompts to general-purpose LLMs are extraordinarily hard to patent and even harder to enforce. If your “invention” is fundamentally a prompt or prompt-engineering technique applied to someone else’s model, look at trade secret protection or copyright instead.
Listing AI as an inventor. Don’t. Please don’t. Anywhere. Ever. It’s a fast way to invalidate your application.
You did not document your concept trail. If a competitor later challenges your patent in court or at PTAB and argues a human didn’t really do the inventing, your contemporaneous records are your defense. Don’t wait until that moment to create them.
Mistaking the USPTO’s friendlier posture for a free pass. The USPTO is more willing to grant AI patents now, but courts apply the same eligibility framework they’ve applied for years. A patent that survives examination but can’t survive litigation is not worth much.
WORKING CHECKLIST. BECAUSE IT WORKS. WELL, YOU DO.
Before you start working on your patent application with a patent attorney for an AI invention, you should be able to answer “yes” to all of these:
We can describe the invention as a solution to a technical problem, not just a business problem
The technical novelty lives in how the AI itself works, not in the dataset, domain, or UI
We can articulate the mechanism of the improvement, not just the result
We have measurable evidence that the mechanism produces a technical improvement
We have contemporaneous records showing human conception of the invention
We have not yet publicly disclosed the invention, or we’re within the one-year U.S. grace period
We can identify a credible technical expert who could vouch for the improvement if needed
If you can check all seven boxes, you’re in good shape to engage patent counsel and move forward. If you can’t, the gaps tell you what to work on first.
THIS GUIDE IS GENERAL INFORMATION, NOT LEGAL ADVICE. PATENT DECISIONS FOR SPECIFIC INVENTIONS SHOULD BE MADE WITH QUALIFIED COUNSEL WHO CAN EVALUATE YOUR PARTICULAR TECHNOLOGY AND BUSINESS SITUATION.