Here is the thing about artificial intelligence and patents that nobody — not the attorneys, not the founders, not the VCs in their quarter-zips — is willing to say plainly:
Nobody knows what they're doing.
Not completely.
Not in the way that a tax attorney knows what she's doing when she files a return, or a plumber knows what he's doing when he sweats a copper joint. The entire field of AI patent law is, as of this writing in early 2026, a kind of magnificent improvisation — a jazz performance where the musicians are also building the instruments while the audience throws tomatoes and money in roughly equal proportions.[^1]
You've built something. Or you are building something. Or you are the lawyer for someone who is building something. The something involves artificial intelligence — a term we will use here with the full knowledge that, in common parlance, includes everything from neural networks to machine learning, to deep learning and more (even if such usage drives technologist up a literal wall).
You want to know whether you can patent it, whether you should patent it, how to patent it without accidentally handing your competitors a detailed instruction manual, and what happens when someone with a patent you've never heard of sends you a letter suggesting that you owe them money.
These are reasonable questions. This is a guide that attempts to answer them. It will not be short. The subject does not permit brevity. What it will be is honest, which is more than can be said for most legal marketing content, and useful, which is the only thing that justifies its length.
Part I: Whether You Can Patent the Thing at All
The first question every AI founder asks — usually at about 2 a.m., in a moment of either inspiration or panic — is the foundational one: Can you actually patent an AI algorithm?
The answer, which manages to be both "yes" and "it depends" simultaneously, is a kind of legal quantum superposition that collapses only upon observation by a patent examiner.[^2]
The United States Patent and Trademark Office will not allow you to patent a mathematical formula, an abstract idea, or a law of nature. These are the rules. They have been the rules since 1853, when the Supreme Court decided Le Roy v. Tatham and declared that a principle of nature cannot be monopolized. Nobody disagrees with this in theory. In practice, the question of what constitutes a mere "abstract idea" versus a patentable "application" of that idea has generated more litigation, more scholarly articles, and more forehead-creasing confusion than perhaps any other question in American intellectual property law.
The landmark case is Alice Corp. v. CLS Bank, decided in 2014. It established a two-step test for patent eligibility under Section 101 that has since been applied to thousands of AI patent applications with the kind of inconsistency that would get you fired from any job that required measuring things accurately. Getting your machine learning patent past the Alice hurdle requires understanding not just the law but the specific examiner, the specific art unit, and the specific way you describe what your invention actually does — which, in the AI context, is complicated by the fact that many inventors cannot fully explain what their invention actually does, because the invention is, by design, a system that teaches itself to do things the inventors did not explicitly program it to do.
The situation changed meaningfully — or at least the direction of the prevailing winds shifted — with two developments that every person reading this document should understand.
First: The Federal Circuit's decision in Recentive Analytics v. Fox Corp., handed down in April 2025. The court held that simply applying a conventional machine learning technique to a new data domain does not make the resulting invention patent-eligible.
If your claim amounts to "we used ML to solve Problem X" and the only novel thing is the X, you are in trouble. The claim must recite something technically specific about how the AI system works — a particular architecture, a novel training methodology, a concrete technical improvement over prior approaches — or it falls on the wrong side of the eligibility line. This was, to use the technical legal term, a big deal.
If your claim amounts to "we used ML to solve Problem X" and the only novel thing is the X, you are in trouble. The claim must recite something technically specific about how the AI system works — a particular architecture, a novel training methodology, a concrete technical improvement over prior approaches — or it falls on the wrong side of the eligibility line. This was, to use the technical legal term, a big deal.
Second: Ex parte Desjardins, a PTAB decision designated precedential in November 2025 and incorporated into the MPEP the following month.
This one went the other way.
The Board held that a machine learning system designed to address a specific technical problem — in this case, catastrophic forgetting — was patent-eligible because the claims were directed to a concrete improvement in computer technology, not to the abstract concept of "using AI." If Recentive was a door closing, Desjardins was a window opening, and the resulting cross-breeze is what patent prosecutors are currently navigating.[^3]
The Board held that a machine learning system designed to address a specific technical problem — in this case, catastrophic forgetting — was patent-eligible because the claims were directed to a concrete improvement in computer technology, not to the abstract concept of "using AI."
The practical upshot, which is what you actually want to know: You can patent AI inventions. But your claims need to do more than recite a function. They need to describe a technical how. What the USPTO will and won't allow for machine learning models is a moving target, but the trajectory since late 2025 has been cautiously favorable for well-drafted applications that get specific about architecture, training methodology, or measurable technical improvements.
One more thing, because everyone is asking it and pretending they aren't: Can you patent a prompt? The answer, right now, is probably not as a standalone invention, though prompt-dependent systems with novel architectures might qualify. This area is so new that there is essentially no precedent, which means it is either an opportunity or a trap, depending on your tolerance for ambiguity.
Part II: Whether You Should Patent the Thing at All
Here is something that a patent law firm's blog almost never tells you, because it is not in their financial interest to tell you: Sometimes you should not patent your AI. There are at least five situations where other forms of protection — trade secrets, defensive publication, copyright, contractual restrictions, or simply moving faster than your competitors — make more strategic and economic sense than filing a patent application.
The central tension is this: A patent requires disclosure. You must describe your invention in sufficient detail that a person skilled in the art could reproduce it. For an AI company, this means potentially revealing the architecture, the training methodology, the data pipeline innovations, or the loss function modifications that constitute your competitive advantage.
You are, in a very real sense, publishing a paper about the thing that makes your company valuable, and you are doing it in exchange for a twenty-year right to exclude others from practicing that specific thing — a right that can be challenged, designed around, or simply ignored by companies in jurisdictions that do not respect your patent.[^5]
The patent-versus-trade-secret decision is therefore not a legal question so much as a strategic one, and it requires you to think honestly about several things that founders would rather not think about:
How easy is your innovation to reverse engineer?
How quickly is the technology evolving?
How much of your competitive advantage is in the model itself versus the data you trained it on?
Do you plan to license the technology?
Do you plan to sell the company?
The last question matters more than most founders realize. AI patents affect startup valuations in ways that are both quantifiable and psychological.
Venture capitalists evaluate patent portfolios during due diligence not because they are patent experts — most are not — but because patents serve as a proxy for defensibility, for the existence of a moat, for the likelihood that some larger company cannot simply replicate what you've built and bury you with distribution advantages. A weak or nonexistent IP portfolio can signal to investors that your innovation is undifferentiated, which is another way of saying that you are just a team, and teams can be hired away.
But how do you build an AI patent portfolio when you are bootstrapping, which is to say when you have more conviction than cash?
The answer involves provisional applications (which cost substantially less than non-provisionals and buy you twelve months), micro-entity fee reductions (seventy-five percent off, if you qualify), strategic prioritization (patent the thing that makes you money first, patent the thing that makes you interesting second), and a clear-eyed understanding that not every invention in your pipeline needs to be patented.
Some should be trade secrets.
Some should be published defensively to prevent competitors from patenting them.
Some should simply be built and shipped as fast as possible, because speed is its own form of protection in markets that move this quickly.
Before you file anything, though — before you spend a dollar on prosecution — you should understand whether you are free to operate at all. A freedom-to-operate analysis is the process of determining whether your product might infringe someone else's patents, and if you are an AI company in 2026, the answer is: maybe. The landscape is vast. AI patent applications appear in roughly forty-two percent of all USPTO technology categories. Finding whitespace in a crowded landscape is both a defensive necessity and a strategic opportunity, but you cannot find whitespace if you do not look. Here is how we do it.
Part III: How to Patent the Thing
Here is where it gets technical. I apologize in advance, but only slightly, because the people who need this section most — in-house patent counsel, patent prosecutors, and the particularly ambitious founder who reads prosecution office actions for recreation — want specificity, not apology.
Drafting AI patent claims that actually survive examination is an art, and like most arts it is easier to demonstrate than to describe. The post linked in this sentence contains before-and-after examples — actual claim language showing what fails and what succeeds at the USPTO — because it turns out that the single most useful thing a patent blog can do is show you the work, rather than talking about the work in abstract terms, which is ironic given that the whole problem with AI patents is abstraction.
The key principle, post-Recentive: Your claims must recite technical specificity. Not "a machine learning model trained on data to produce predictions" — that is a wish, not an invention. Instead: the specific model architecture, the specific training protocol, the specific way the system handles a known technical problem. The more your claims read like an engineering specification and less like a marketing brochure, the better your chances.
The disclosure dilemma is the corresponding challenge on the specification side. Disclose too little, and you face enablement and written description rejections. Disclose too much, and you have published a cookbook for your competitors — one that remains publicly available even if your claims are ultimately narrowed or invalidated. The calibration of disclosure is, I would argue, the single most consequential decision in AI patent prosecution, and it requires a different approach depending on whether your AI system is static (trained once, deployed as-is) or dynamic (continuously learning, adapting to new data in production).
For patent counsel managing larger portfolios, two strategic considerations deserve attention. First, continuation strategies allow you to maintain "living" patent families that evolve with your technology — an approach that is particularly valuable in AI, where the product you ship in eighteen months may bear only a family resemblance to the product you described in your initial filing. Second, prosecution by the numbers — allowance rates, average timelines, and examiner-level statistics across AI art units — can inform both filing strategy and outside counsel selection in ways that gut instinct alone cannot.
And because we are living in the year 2026, a word about using AI tools in patent prosecution itself. Yes, AI-powered drafting and prior art search tools exist. Yes, they can reduce costs and accelerate timelines. No, they do not eliminate the need for human judgment. And yes, there are real ethical and confidentiality questions about feeding client inventions into AI systems that may retain or learn from the input. This is a topic on which reasonable people disagree, and the disagreement is evolving faster than anyone's blog post can track, including this one.
Part IV: The Inventorship Question, or, Can a Robot Be Thomas Edison?
No. A robot cannot be Thomas Edison. Not legally, anyway.[^6]
The question of whether an AI system can be named as an inventor on a patent application has been litigated more or less to death, thanks to Stephen Thaler and his AI system called DABUS, which he submitted as the sole inventor on patent applications in virtually every major jurisdiction on earth. The global status of AI inventorship after Thaler is: no. Not in the US, not in the UK, not in the EU, not in Australia (which briefly said yes, then reversed itself on appeal). An inventor must be a natural person. This is the law.
What the law does not resolve — and what matters much more for your day-to-day operations — is the question of inventorship when AI is used as a tool in the inventive process. If an engineer prompts an AI system and the AI produces an output that becomes the basis for a patentable invention, who is the inventor? The engineer? The engineer who designed the AI? The person who curated the training data? All of them?
The 2025 USPTO inventorship guidance simplified the prior framework by eliminating the Pannu factors and treating AI as a tool no different from a microscope or a calculator. The traditional conception test applies: the inventor is the person who contributed to the conception of the claimed invention. But documenting human contributions to AI-assisted inventions is harder than it sounds, because the line between "I told the AI what to do" and "the AI did the creative work" is blurry, contested, and often reconstructed after the fact by people who have financial incentives to reconstruct it in particular ways.
Part V: When Someone Comes for You
There are two kinds of AI companies: those that have received a patent demand letter, and those that will receive a patent demand letter.
This is not pessimism. It is arithmetic. Non-practicing entity suits targeting AI companies increased by fifteen to twenty percent in 2025. AI patent trolls are specifically targeting startups, and they are sophisticated about it — timing demand letters to coincide with funding announcements, when the startup has cash and the investors are pushing for clean resolutions. More than half of AI patent lawsuits are filed against companies with annual revenue under twenty-five million dollars. The targets are not Google and Microsoft.[^7] The targets are you.
If you receive a demand letter, do not respond immediately. Do not panic. Do not ignore it. Do engage counsel with specific experience in AI patent assertions. Do evaluate whether your company has insurance that covers patent claims. Do assess the validity of the underlying patent, because a remarkable number of demand letters are built on patents that would not survive an inter partes review.
Proving AI patent infringement is genuinely difficult, and this difficulty works in the defendant's favor. AI systems are opaque by nature — the source code is proprietary, the training data is confidential, the internal workings of a neural network are not inspectable in the way that a mechanical device or a chemical process is inspectable. Courts have adapted by accepting circumstantial evidence — marketing materials, published papers, job postings describing the accused technology — but discovery in AI patent cases remains complex and expensive.
On the proactive side, defensive patent strategies for AI companies — including membership in organizations like the LOT Network or Unified Patents' AI Zone, defensive publication programs, and cross-licensing arrangements — can reduce your exposure before a demand letter arrives. The time to build a defensive strategy is before you need it. This is the kind of advice that sounds obvious until you realize that ninety percent of AI startups have done nothing about it.
Part VI: Generative AI, Open Source, and the Questions That Didn't Exist Three Years Ago
We are now in territory that the law has not caught up to, which is uncomfortable for lawyers and exciting for everyone else.
Generative AI has blurred the lines between patent, copyright, and trade secret protection in ways that make traditional IP categories feel like maps drawn before the continents shifted. When a company fine-tunes a foundation model, what exactly is the protectable IP? The fine-tuning method might be patentable. The training data might be copyrightable (or might infringe someone else's copyright — a question currently being litigated in approximately one thousand jurisdictions). The model weights might be best protected as trade secrets, assuming you can keep them secret, which is increasingly difficult in an ecosystem that values open publication and research collaboration.
Then there is the open-source question. Most AI startups build on open-source frameworks — PyTorch, Hugging Face models, various pretrained architectures available under a menagerie of licenses with different restrictions and obligations.
What can you patent when your foundation is open-source?
How do copyleft licenses interact with patent claims? What happens during due diligence when an investor discovers that your core model is fine-tuned from a foundation model with license terms you haven't fully read?[^8]
Who owns AI-generated output is the question that ties all of this together, and the answer — which varies by platform, by jurisdiction, by the specific terms of service governing the AI system that produced the output — is the kind of lawyerly mess that keeps general counsel awake at night and keeps patent attorneys employed during the day.
Part VII: The World Is Not Flat, but the Patent Landscape Might Be
The final cluster of things you need to know involves geography, which is to say: other countries exist, they have their own patent systems, and those systems do not agree with each other about much.[^9]
Filing AI patents globally requires understanding three fundamentally different approaches to the same question.
The United States uses the Alice/Mayo framework and asks whether your claims are directed to an abstract idea.
The European Patent Office asks whether your invention has "technical character" under its computer-implemented inventions guidelines.
China requires a "technical solution using technical means yielding a technical effect," which sounds circular because it is circular, but which in practice has resulted in a more permissive approach to AI patentability than either the US or Europe.
The numbers tell a story. China accounts for sixty to seventy percent of global AI patent filings. The US-China AI patent race has produced approximately 300,000 Chinese AI patent applications in 2024 alone — roughly four times the US volume. The quality and international enforceability of these patents varies enormously, but the sheer volume creates freedom-to-operate challenges that no globally operating AI company can afford to ignore.
And for companies operating in or selling into Europe, the EU AI Act's implications for patent strategy deserve urgent attention. The Act requires detailed technical documentation for high-risk AI systems, including algorithms, datasets, and testing protocols. In Europe's absolute novelty regime, these regulatory disclosures could destroy patent novelty if they precede patent filing. The full applicability date for most operators is August 2026, which is — checks calendar — very soon.
The AI patent landscape is complicated, evolving, and occasionally absurd. It rewards companies that think strategically and punishes companies that think reactively. It requires different things of founders, general counsel, and patent counsel, but it requires something of all of them, and that something is: attention. Sustained, informed, occasionally tedious attention to the details of how intellectual property law intersects with a technology that is itself only partly understood by the people who build it.
You do not need to understand all of this at once. You do not need to read all thirty of the deeper analyses linked throughout this guide in a single sitting, though you are welcome to, and there is coffee. What you need is a framework for knowing which questions to ask, when to ask them, and who to ask — and the self-awareness to recognize when you are in territory where improvisation, however brilliant, is no substitute for expertise.
The musicians are still building the instruments. The music, somehow, is getting better.
[^1]: The tomatoes, in this metaphor, are Section 101 rejections. The money is venture capital. Both arrive without warning and in quantities that seem almost personally targeted.
[^2]: This is not a metaphor that patent examiners find amusing. Patent examiners, as a class, find very few things amusing, though this is perhaps unfair, and certainly one should not generalize about a group of professionals whose individual examiner allowance rates vary from 12% to 94% within the same technology center, which tells you something important about the predictability of this process that no amount of legal analysis can convey quite as efficiently.
[^3]: I am aware that this paragraph contains two architectural metaphors in sequence. I am not sorry. Patent law is a house with many rooms and most of them are locked and nobody can agree on who has the keys.
[^5]: Which is to say: many jurisdictions. The international patent enforcement landscape is, to put it charitably, uneven.
[^6]: Whether a robot could be Thomas Edison in the sense of being a tireless, obsessive workaholic who engaged in extensive and sometimes ethically questionable competitive practices is a question for a different blog.
[^7]: Google and Microsoft also get sued, obviously. But they have legal departments the size of small cities and patent portfolios that function as nuclear deterrents. You, presumably, do not.
[^8]: What happens is that the deal gets complicated. "Complicated" is a polite word for what actually happens to deal timelines when license issues surface during diligence.
[^9]: They agree that perpetual motion machines are not patentable. This is approximately the extent of the consensus.
This guide is updated quarterly to reflect new case law, USPTO guidance, and legislative developments. It is intended as general information and does not constitute legal advice. If you have specific questions about your AI patent strategy, contact us — we actually like talking about this stuff.