The Machine That Asked to Be Called an Inventor

In 2018 a computer system named DABUS—its creator, the physicist and AI researcher Dr. Stephen Thaler, called it a "Device for the Autonomous Bootstrapping of Unified Sentience"—produced two things its maker said it had invented on its own: a food container with a fractal surface geometry meant to nest and grip more efficiently, and a flashing emergency light, dubbed a "neural flame," designed with a pulse pattern that the human eye finds hard to ignore. Dr. Thaler then did something no one had done before. He filed patent applications in a string of countries, and on the line where the inventor's name goes, he wrote the name of the machine.

It was, in part, a deliberate provocation. Dr. Thaler wanted to force patent systems to answer a question their statutes had never imagined anyone would ask: can a machine be an inventor? And underneath that question lay a harder one that will outlast DABUS by decades. As AI systems get better at generating genuinely new solutions to technical problems, how much must a human still do for a patent to issue at all?

The answers came back, jurisdiction by jurisdiction, and they were strikingly consistent on the headline and strikingly inconsistent on the details. The United States, the European Patent Office, and the United Kingdom all said the same thing: an inventor must be a natural person, so DABUS cannot be named. Australia briefly said yes before its Full Court reversed; South Africa registered a DABUS patent only because its system grants patents without examining them. But behind that near-unanimous "no machine inventors" headline sits the question that actually governs day-to-day practice—what counts as enough human contribution when an AI does much of the work?—and on that question the jurisdictions diverge, hedge, and keep changing their minds. The United States, as we will see, changed its formal answer twice in under two years.

That divergence is not academic. AI now sits at the center of drug discovery, materials science, protein design, semiconductor floorplanning, and countless other fields where a patent is the difference between a defensible investment and a free gift to competitors. A company that builds its research around AI tools needs to know whether the inventions that come out the other end can be patented, who must be named, and how to structure its work so that the answer stays "yes." This article maps the global picture and translates it into practical guidance for filing across borders.

The companion question—who owns what an AI helps create, once you know a human invented it—gets its own full treatment in AI-generated inventions: who owns what the machine creates. For the wider terrain of AI law beyond patents, see our overview of artificial intelligence key legal issues. And because every patent question that crosses a border eventually becomes an enforcement question, our guide to global patent litigation strategies is the natural sequel to this one.

What "Inventor" Has Always Meant

For most of patent history, no one bothered to define "inventor" carefully, because the meaning seemed obvious. An inventor was a person who had an idea. The U.S. Patent Act, when it does define the term, calls the inventor "the individual or, if a joint invention, the individuals collectively who invented or discovered the subject matter of the invention" (35 U.S.C. § 100(f)). "Individual," courts had long assumed, means a flesh-and-blood human—not a corporation, not a partnership, and certainly not a machine. The statute's neighboring provisions reinforce that reading. Section 115 requires each inventor to execute an oath or declaration, and to swear to it; an oath is the sort of thing only a person can give. The Patent Act, in short, was built around the silent premise that inventors are people.

The legal heart of inventorship is a mental event called conception. The classic formulation, nearly a century old, describes conception as "the formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention, as it is thereafter to be applied in practice" (Townsend v. Smith, 36 F.2d 292, 295 (C.C.P.A. 1930)). Conception is the moment the finished idea exists in someone's head, clearly enough that an ordinarily skilled person could build it from the description. The word "mind" was doing quiet but decisive work in that sentence. If conception requires a mind, and a machine has no mind in the legal sense, then a machine cannot conceive—and cannot invent.

European law made the same assumption from a different direction. The European Patent Convention gives the inventor "the right to be mentioned" before the patent office (Article 62 EPC) and treats the right to a patent as something that passes from the inventor to an applicant (Article 60 EPC)—legal relationships that only a person can hold. A machine cannot hold a right, cannot transfer one, and cannot be mentioned as the bearer of one. British law took the same path through section 7 of the Patents Act 1977, which speaks of the inventor as the "actual deviser" and routes the right to grant through people and the property rights they can hold.

None of this was controversial as long as machines were tools. A chemist who ran a molecular-modeling program still conceived the molecule; the program was a fancier slide rule. An engineer who used CAD software still designed the part; the software was a fancier drafting table. The machine was always the means and the human was always the mind. AI is what cracked that comfortable arrangement—because for the first time, the definite and permanent idea of the solution can appear somewhere other than a human head.

A closer look at conception—the concept that does all the work

Because so much turns on conception, it is worth understanding what the doctrine actually requires, in terms a non-lawyer can follow. Conception is not the same as having a hunch or stating a goal. The Federal Circuit has long held that an idea is "definite and permanent" enough to count only when the inventor has a specific, operative solution in mind—not merely a problem to be solved or a research plan to pursue. In Burroughs Wellcome Co. v. Barr Laboratories, Inc., 40 F.3d 1223, 1228 (Fed. Cir. 1994), the court explained that conception is complete when the idea is "so clearly defined in the inventor's mind that only ordinary skill would be necessary to reduce the invention to practice, without extensive research or experimentation." The inventor must have the finished invention in view; what remains can only be the routine work of building it.

That standard has two consequences that matter enormously for AI-assisted work. First, conception is about the solution, not the problem. A person who brilliantly identifies an unmet need—"someone should find a non-flammable, fast-charging electrolyte"—has not conceived anything in the patent sense, however valuable that insight is commercially. Conception belongs to whoever first holds the definite idea of the answer. The Federal Circuit drove the point home in Hitzeman v. Rutter, 243 F.3d 1345 (Fed. Cir. 2001): an inventor must have a definite and permanent idea of the operative invention, including an appreciation that it will work for its intended purpose, not merely a hope or a goal.

Second, the law sharply distinguishes conception (the mental act) from reduction to practice—actually building the invention, or describing it in a patent application well enough to show that it works. Inventorship attaches to conception, not to reduction to practice. This is why a person who merely carries out someone else's fully formed idea is a skilled technician, not a co-inventor (Sewall v. Walters, 21 F.3d 411, 416 (Fed. Cir. 1994)), and why a person who runs the confirmatory experiment that verifies a colleague's conception does not thereby become an inventor. The Federal Circuit said it crisply in Burroughs Wellcome: "a person who simply provides the inventor with well-known principles or explains the state of the art without ever having a firm and definite idea of the claimed combination as a whole does not qualify as a joint inventor."

Map that onto the AI problem and the difficulty comes into focus. If a human holds the definite idea of the solution and uses AI to build, simulate, or verify it, conception lived in the human mind and the human is the inventor—the AI merely helped reduce the idea to practice. But if the AI is the first place the definite, operative solution ever appears, the conception did not happen in any human mind, and the traditional doctrine has no inventor to point to. The whole global debate is, at bottom, an argument about which of these two stories a given AI-assisted invention tells.

Where AI Breaks the Old Picture

Machine-learning systems can now surface solutions that no human directed, anticipated, or—sometimes—fully understands. A neural network trained on chemical data can flag a molecule with useful properties no chemist set out to find. A generative-design system can evolve a mechanical bracket no engineer would have drawn, lighter and stronger than anything in the textbooks. Reinforcement-learning agents have produced antenna geometries and chip layouts that look, to human eyes, almost deranged—and outperform the human designs anyway. When that happens, the conception test stops being a formality and becomes a genuine puzzle. If the inventive idea never formed in a human mind, who is the inventor? The AI? The people who built it? The people who trained it, chose its data, or framed the problem? Or is there simply no inventor, and therefore no patent?

To make this less abstract, hold two examples in mind throughout this article. The first is DABUS itself: the extreme case, where (on Dr. Thaler's own telling) no human conceived the inventions at all. Every major patent office has rejected that scenario. The second is far more common and far more important. Picture Dr. Elena Marchetti, a researcher at a hypothetical company called Helix Materials, who is trying to find a safer electrolyte for fast-charging batteries. She frames the problem, chooses and tunes an AI discovery system the company calls FORGE, curates the training data, sets the performance constraints, runs the search, and then—from thousands of candidate molecules FORGE proposes—recognizes that one obscure compound is the breakthrough, refines it, and validates it in the lab. (This scenario, and the rival company introduced later, are hypothetical, used to illustrate doctrine.) Dr. Marchetti is nowhere near the DABUS extreme. But how much of what she did counts as conception? That is the question every patent system is now struggling to answer, and the question on which Helix's patent will live or die.

United States: Humans Only, and a Sharp Return to First Principles

DABUS loses, cleanly

The U.S. answer to the headline question is settled. The USPTO rejected the DABUS applications in 2020, and the Federal Circuit affirmed in Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022). Writing for the panel, Judge Stark held that the Patent Act's use of "individual" unambiguously means a natural person. The court leaned on the statutory text and on the Supreme Court's instruction in Mohamad v. Palestinian Authority, 566 U.S. 449 (2012), that "individual" in a statute ordinarily means a human being unless Congress says otherwise. It refused to let policy arguments about encouraging innovation override the statute's text, observing that nothing in the Patent Act suggested Congress meant to break from the ordinary meaning—and that if AI inventorship is to be recognized, Congress, not the courts, must do it. The Supreme Court declined to take the case (143 S. Ct. 1783 (2023)). So in the United States, an AI system cannot be named as an inventor. Full stop.

The question that actually matters—and a major reversal

That left the practical question untouched: when a person like Dr. Marchetti invents with AI, what must she have done to qualify as the inventor? Here the U.S. position has whipsawed, and any reader relying on older guidance needs to know that the ground moved.

In February 2024, the USPTO issued "Inventorship Guidance for AI-Assisted Inventions" (89 Fed. Reg. 10043 (Feb. 13, 2024)), which answered the question by borrowing the test for joint inventorship among multiple humans—the so-called Pannu factors from Pannu v. Iolab Corp., 155 F.3d 1344, 1351 (Fed. Cir. 1998). Under Pannu, a person is a joint inventor only if they (1) contribute in some significant manner to conception, (2) make a contribution to the claimed invention that is not insignificant in quality when measured against the full invention, and (3) do more than merely explain well-known concepts or the state of the art. The 2024 guidance treated the human-plus-AI relationship as something like a partnership, to be evaluated factor by factor: had the human contributed "significantly" to conception, given everything the AI did? The guidance was widely discussed and widely criticized for importing a multi-person joint-inventorship analysis into situations where only one human was involved—arguably making it harder to patent AI-assisted work by inviting examiners to weigh the human's contribution against the machine's, as if the machine were a rival inventor whose share had to be subtracted out.

On November 28, 2025, the USPTO rescinded that 2024 guidance in its entirety and replaced it with revised guidance, issued to implement Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence" (Jan. 23, 2025), which directs federal agencies to clear away policies that needlessly impede AI development. The new framework is a deliberate return to first principles, and it changes the analysis in four practical ways.

First, it reaffirms—without wavering—that only natural persons can be inventors. That part did not change, and could not, because it is statutory. Second, and this is the heart of it, the revised guidance treats an AI system as a tool, analogous to laboratory equipment, computer software, research databases, or any other instrument that assists in the inventive process. Using a powerful tool does not diminish your inventorship any more than using a powerful microscope, a mass spectrometer, or a supercomputer does. Third, the guidance discards the central move of the 2024 approach: the Pannu factors apply only when deciding whether multiple human beings are joint inventors, and are simply inapplicable when a single human invents with AI assistance—because an AI cannot be a "joint inventor" at all, and so there is no second contributor against whom to measure the human's share. When one person is involved, examiners apply the ordinary conception test and nothing more: did this person conceive the invention? Fourth, the guidance confirms that these principles run across utility, design, and plant patents alike, and that a priority or benefit claim to an earlier application must share at least one human inventor in common—so a chain that rests on AI "inventorship" somewhere up the line will not be honored.

The net effect is a more permissive, more conventional posture. For Dr. Marchetti, the question is no longer "did she clear a multi-factor significant-contribution hurdle against the machine?" It is the familiar one every inventor has always faced: did she conceive the invention, using FORGE as a tool? On the facts above—she framed the problem, shaped and constrained the AI, recognized the significance of an obscure result, and refined and reduced it to practice—the answer is comfortably yes. The conception lives in her mind; FORGE is the instrument.

What still trips people up

The simpler standard does not make the issue disappear; it relocates it. Conception is intensely fact-bound, and the hard cases are real. If a person did nothing but type a one-line prompt—"find me a better electrolyte"—and accept whatever came out, a U.S. examiner, and far more dangerously a future litigation adversary attacking the patent's validity, could fairly ask whether that person conceived anything at all. The traditional doctrine has always policed the line between an inventor and a mere "supervisor" or "person who poses a problem." In Garrett Corp. v. United States, 422 F.2d 874 (Ct. Cl. 1970), the court reminded us that one who merely suggests an idea of a result to be accomplished, rather than the means of accomplishing it, is not a joint inventor. AI sharpens that line rather than erasing it. The practical defense is the same one good inventors have always relied on: contemporaneous records showing the specific technical judgments the human made. We return to documentation below, because it is where cross-border filing strategy actually lives.

European Patent Office: A Firm No, and a Tool-Based Yes for the Rest

The EPO rejected the DABUS applications at first instance under Rule 19(1) EPC for failing to designate a proper inventor, and its Legal Board of Appeal affirmed in decisions J 8/20 and J 9/20 (December 2021). The Board reasoned from the structure of the Convention: the inventor has a right to be mentioned (Article 62 EPC) and a right that passes to the applicant (Article 60 EPC), and only a person with legal capacity can hold and transfer such rights. A machine has no legal personality, so it can be neither the bearer of the right to be mentioned nor the source of a derived right to apply. Dr. Thaler's fallback—that he owned DABUS and therefore succeeded to whatever it produced, "as employer" or by accession—failed too, because there was no inventor's right for him to succeed to. Policy arguments about fostering innovation, the Board said, were for the legislators who write and amend the Convention, not for the Office reinterpreting it.

On the practical question, the EPO has not issued a single dedicated pronouncement of the kind the USPTO has now issued twice. Instead it works through its established framework for computer-implemented inventions, reflected in the Guidelines for Examination (notably G-II, 3.3 and 3.3.1 on artificial intelligence and machine learning), and the upshot is broadly tool-based and compatible with the new U.S. posture. AI and machine-learning methods are patentable in Europe if they meet the ordinary requirements—novelty, inventive step, industrial applicability, sufficiency of disclosure—and if they have technical character. A bare mathematical method or algorithm has no technical character and is excluded "as such" under Article 52(2) and (3) EPC; the EPO's Guidelines note that terms like "support vector machine," "reasoning engine," or "neural network" usually refer to abstract mathematical models devoid, by themselves, of technical character. But AI applied to a technical problem—controlling a physical system, classifying technical measurements, improving how a computer itself works, or processing medical, audio, or image data for a technical purpose—can acquire technical character and cross the threshold.

For inventive step, the EPO uses its problem-and-solution approach and assumes the skilled person already commands standard machine-learning techniques. Merely applying a known AI method to a known problem is therefore usually obvious; the inventive contribution has to lie somewhere more specific—in a novel architecture, in a technically meaningful choice of training data, in the technical application, or in a real technical insight about how to make the system work. And the disclosure must let a skilled person reproduce the system. For machine-learning inventions, that sufficiency requirement (Article 83 EPC) can demand disclosure of the training-data characteristics, the model architecture, and the training methodology—because a specification that recites only "we trained a model and it produced this output" may teach the public nothing it could actually repeat.

Run Dr. Marchetti's electrolyte through this and the answer is again favorable, but for European reasons: the invention solves a concrete technical problem (a safer, faster-charging battery chemistry), the inventive technical contribution is identifiable (the specific compound and the insight that led to recognizing it), and FORGE is used as a tool by an identifiable natural-person inventor. The question Europe asks—did a natural person make the inventive technical contribution?—maps neatly onto Dr. Marchetti.

United Kingdom: The Highest Court Speaks

The United Kingdom produced the most authoritative judicial word on the headline question anywhere in the world. After the UKIPO and the lower courts rejected the DABUS applications, the Supreme Court took the case and, in December 2023, ruled unanimously against Dr. Thaler in Thaler v. Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49. Lord Kitchin's judgment held three things in plain terms. First, an "inventor" under section 7 of the Patents Act 1977 must be a natural person; the statute consistently speaks of the inventor as "the actual deviser of the invention," and a "deviser" is a person who devises—a human. Second, Dr. Thaler could not derive the right to apply from DABUS, because a machine cannot own property—including an invention—and so has nothing to transfer, and there is no rule of law by which the owner of a machine automatically owns whatever the machine generates. Third, the policy case for recognizing AI inventors, however serious, is a matter for Parliament, not the courts. The Court was careful to note what it was not deciding: it expressed no view on whether inventions devised with the assistance of AI are patentable, or on who might be entitled to them—only that DABUS, a machine, could not be named as the inventor.

On the practical side, the UKIPO's approach centers on identifying the "actual deviser"—the person whose intellectual effort produced the inventive concept. The emphasis is on conception over implementation. A person who conceives a solution and uses AI to implement or optimize it can be the inventor; a person who merely hands a problem to an AI and accepts the output may not have contributed the inventive concept at all. Recognizing the significance of an AI-generated result can count, the UKIPO indicates, where doing so required genuine technical judgment rather than mere observation. This is, in substance, the same conception inquiry the United States has now returned to and that the EPO conducts through its own vocabulary—convergent destinations reached by different routes. Notably, after the Supreme Court ruling the UK government consulted on AI and patents and declined, for the time being, to change the law, concluding that the human-inventor rule remains workable. The door to AI inventorship in Britain is closed, and Parliament has so far chosen not to reopen it.

Asia: Agreement on the Rule, Variation on the Margins

The major Asian patent offices reach the same headline conclusion—humans only—while differing in tone and in how much human contribution they expect.

China. The China National Intellectual Property Administration (CNIPA) treats inventors as natural persons, consistent with the requirement that an inventor designation name the "person" who made the inventive contribution to the substantive features of the invention. In keeping with the scale of China's AI industry, CNIPA's practical posture is comparatively flexible: human involvement in defining the problem, choosing the AI tools, directing the process, and evaluating outputs can support inventorship even where the AI generated large portions of the technical solution. China's 2024 Guidelines for Patent Applications Involving Artificial Intelligence and related guidance signal an accommodating stance toward AI-assisted invention while holding firm on the natural-person rule. At the same time, China has pressed hard on disclosure, insisting that AI and machine-learning inventions be described well enough to reproduce—algorithms, training data, model architecture—so that a specification reciting only the AI's output, with no path to generating it, can be rejected for insufficiency.

Japan. The Japan Patent Office likewise requires human inventors, reading Japan's Patent Act—which defines an "invention" as the "highly advanced creation of technical ideas utilizing the laws of nature" (Article 2(1))—to require a human creative act, since a machine cannot "create" within the statute's meaning. Using even a very sophisticated AI tool does not defeat human inventorship if a human directed the creative process; but if the AI autonomously generated the inventive concept with no meaningful human intellectual contribution, no valid inventor exists, and there is no one to name.

South Korea. The Korean Intellectual Property Office aligns with the international consensus that inventors must be natural persons. Like the others, it emphasizes documenting the human contributions: who contributed, what each did, how the AI was used as a tool, and how human conception related to the AI's outputs. KIPO has been an active voice in international discussions on AI and IP, but it has not departed from the human-inventor rule.

South Africa: the asterisk. South Africa registered a DABUS patent in July 2021, making headlines as the first jurisdiction to list an AI as inventor. The explanation is procedural rather than philosophical: South Africa uses a deposit-style registration system that grants patents meeting formal requirements without substantive examination of patentability or inventorship. The grant reflects the absence of examination, not a considered decision that machines can invent, and it would be vulnerable to revocation if ever challenged in court.

Australia: a reversal worth remembering. Australia briefly broke ranks. A single judge of the Federal Court held in Thaler v. Commissioner of Patents [2021] FCA 879 that DABUS could be named, reasoning that the Patents Act 1990 did not define "inventor" or expressly require a human, that the word's ordinary meaning had evolved, and that recognizing AI inventors better served the system's purpose of encouraging innovation. The Full Federal Court reversed unanimously in Commissioner of Patents v. Thaler [2022] FCAFC 62, holding that "inventor" necessarily refers to a natural person and that entitlement under the Act must flow from a human inventor. The High Court of Australia denied special leave to appeal. The Australian episode is the exception that proves the rule: even the one court that said yes was overturned, leaving a clean global consensus that change, if it comes, must come from legislatures rather than judges.

The table below summarizes where the major offices land.

Jurisdiction Can AI be named inventor? Source of the rule Practical posture on human contribution
United States No 35 U.S.C. § 100(f); Thaler v. Vidal (Fed. Cir. 2022) Traditional conception test; AI is a tool; Pannu only for multi-human joint inventorship (2025 guidance)
EPO No Articles 60, 62 EPC; J 8/20, J 9/20 Tool-based; technical character and identifiable inventive contribution required
United Kingdom No Patents Act 1977 s. 7; Thaler [2023] UKSC 49 "Actual deviser"; conception over implementation
China (CNIPA) No Patent Law (natural-person inventor) Relatively flexible on human contribution; strict on disclosure
Japan (JPO) No Patent Act art. 2(1) Human creative act required; AI as tool permitted
South Korea (KIPO) No Korean patent law Documentation of human contribution emphasized
Australia No Patents Act 1990; Thaler [2022] FCAFC 62 Natural-person requirement after reversal
South Africa Granted in practice Non-examining deposit system No substantive inventorship review

The pattern is unmistakable: a globally settled "no" on machine inventorship, reached through at least four different statutory routes, with the live disagreement pushed down into the human-contribution analysis—where the offices range from China's relative generosity to the more demanding, fact-intensive scrutiny familiar in the United States and Europe.

The Spectrum That Decides Real Cases

Strip away the jurisdictional vocabulary and a single picture emerges. Human involvement in AI-assisted invention runs along a spectrum, and where a given project sits on that spectrum predicts the answer almost everywhere.

At one end is full human conception: the person conceives the complete invention and uses AI only to implement, simulate, or optimize it. Inventorship here is unproblematic; the AI is a calculator with ambitions. Next comes human direction with AI generation: the person frames the problem, selects and tunes the AI, sets constraints, and interprets results, while the AI generates the specific solution. This is Dr. Marchetti's zone, and it is where most real R&D lives. A step further is minimal human involvement: the person poses a problem, the AI autonomously finds the solution, and the person merely notices it is valuable. At the far end is autonomous AI invention—the DABUS scenario—where the machine identifies both problem and solution with no meaningful human direction. Everyone rejects that end; nearly everyone accepts the first; the fight is in the middle.

What pushes a middle-of-the-spectrum project toward valid human inventorship? The recurring factors, drawn from the various offices' guidance and from ordinary conception doctrine, are easy to state and worth internalizing. Merely identifying a problem, without more, tends not to be enough—inventorship turns on conceiving the solution, not naming the gap. Designing or meaningfully adapting an AI system to elicit a particular kind of result can contribute, where simply running an off-the-shelf tool unmodified generally cannot. Curating training data with real technical judgment can contribute, because the data shapes what the system can find. Specifying constraints that embody technical insight about the desired solution can contribute. Recognizing and selecting a valuable output from many can contribute—but only if the recognition required genuine technical expertise rather than the equivalent of picking the prettiest picture, since the law has always refused inventorship to those who merely appreciate the value of someone else's idea. And refining or improving an AI output through additional human insight clearly contributes, at least as to the refined elements. Dr. Marchetti's project checks most of these boxes; a bare-prompt project checks none. The difference is not luck. It is process—and process can be designed.

Two projects, one law: a worked comparison

To see how the same legal framework sorts real cases, compare Helix Materials with a hypothetical rival, ChemBot Inc., both chasing the same battery breakthrough.

At Helix, Dr. Marchetti spends months narrowing the technical problem—not "a better battery," but a specific class of electrolyte that resists thermal runaway above a defined temperature while preserving fast-charge ion mobility. She and her team adapt FORGE for the task, choosing and cleaning the training data (rejecting datasets that would bias the search toward known-flammable families), and they encode their hard-won constraints into the search objective. FORGE returns roughly four thousand candidates. Dr. Marchetti's team recognizes that one unusual fluorinated compound—buried far down the ranked list because a naive scoring metric undervalued it—is the answer, a judgment that required deep domain expertise no off-the-shelf user would have. They then modify the molecule to improve its stability and validate the result in the lab. Across every major jurisdiction, this is a clean human invention. In the United States, Dr. Marchetti conceived the solution using FORGE as a tool; at the EPO, the inventive technical contribution is identifiable and the problem is technical; in the UK, she is the "actual deviser"; in China and Japan, the human creative act is evident. The patent names Dr. Marchetti (and any colleagues who contributed to conception), and it stands.

At ChemBot, an executive types a single instruction into a general-purpose model—"design a non-flammable, fast-charging electrolyte and give me the molecule"—and forwards whatever it returns to the patent department, naming himself as inventor. Nothing in this story is conception. He did not hold a definite, operative idea of the solution; he stated a goal and accepted an output. He did not design or adapt the tool, curate data, encode technical constraints, or exercise technical judgment in recognizing or refining the result. Under the U.S. conception test, there is no human inventor to name. At the EPO, no natural person made the inventive technical contribution. In the UK, there is no "actual deviser." A patent procured on these facts is not merely weak; it is vulnerable to being held invalid or unenforceable for failing to name a proper inventor—a point we turn to next. The striking thing is that both companies used AI to reach a similar chemical result. The law does not punish ChemBot for using AI; it asks what the humans did, and at ChemBot the humans did almost nothing inventive.

Getting inventorship wrong is not a paperwork error

It is tempting to treat inventorship as a formality—a name to be filled in—but mis-stating it carries real consequences that surface at the worst possible moments. Historically, naming the wrong inventors or omitting a real one could render a patent invalid outright, and while modern law is more forgiving of honest mistakes, the danger has not disappeared. U.S. law allows correction of inventorship under 35 U.S.C. § 256, and courts treat correctable, good-faith errors leniently; Pannu itself arose because the question of who invented mattered to validity. But correction is available only where the error can be cured without deceptive intent, and an inventorship problem entangled with bad faith can sink the patent entirely through the doctrine of inequitable conduct—the "atomic bomb" of patent law, which can render an entire patent unenforceable. (We unpack that doctrine in inequitable conduct in patent prosecution and the evidentiary hunt for it in finding evidence of inequitable conduct.) Even an innocent error must be caught and corrected, which means it must first be noticed—often during litigation, when an adversary scrutinizing the patent finds that the named inventor could not have conceived what the claims describe.

For AI-assisted inventions this is a live risk rather than a theoretical one. A defendant accused of infringing Helix's electrolyte patent will look hard for any sign that no human actually conceived the claimed compound, because if it can show that, it may defeat the patent without ever reaching the merits of infringement—a strategy explored in our guides to what constitutes patent infringement: claims and defenses and comprehensive patent infringement litigation. The contemporaneous record of Dr. Marchetti's technical judgments is precisely what answers that attack. The ChemBot patent, by contrast, has nothing to offer in response. The same exposure appears in transactions: when an acquirer's lawyers conduct diligence before buying an AI-driven R&D company, an inventorship gap on a key patent can stall or shrink a deal while everyone scrambles to reconstruct who contributed what—if it can be reconstructed at all. Inventorship done carelessly is a latent defect that compounds with time; done carefully, it is a quiet source of strength.

The Twin Problem: Disclosing a Black Box

AI-assisted inventions raise a second difficulty that sits right alongside inventorship, and the two are joined at the hip. A patent is a bargain—exclusive rights in exchange for teaching the public how to make and use the invention. That teaching obligation lives in 35 U.S.C. § 112(a), which demands both a written description that shows the inventor possessed the claimed invention and an enablement disclosure sufficient to let a person of ordinary skill make and use it without undue experimentation. The trouble is that machine-learning systems can be opaque even to their creators, and a specification that says, in effect, "we fed data to a model and this molecule came out" may satisfy neither requirement.

Start with where AI sits in patent law's oldest fault line: the divide between the predictable and unpredictable arts. The Federal Circuit has long held that the detail required to satisfy written description and enablement varies with the complexity and predictability of the field (Ariad Pharm., Inc. v. Eli Lilly & Co., 598 F.3d 1336, 1351 (Fed. Cir. 2010) (en banc); Capon v. Eshhar, 418 F.3d 1349, 1357–59 (Fed. Cir. 2005)). Mechanical, electrical, and computer inventions have traditionally been treated as predictable, requiring less detail to support broad claims; chemical and life-science inventions are treated as unpredictable, requiring more. AI scrambles this tidy division. Is a generative model that designs molecules a "computer invention" (predictable) or a "chemical invention" (unpredictable)? Stakeholders and the USPTO's own AI policy reports have split on the question, and the answer may turn on the particular invention. The safe assumption for a drafter is the demanding one: write as if the art is unpredictable and a skilled reader will need real detail.

Enablement supplies the operative test. Under In re Wands, 858 F.2d 731, 736–37 (Fed. Cir. 1988), whether a disclosure requires "undue experimentation" turns on a constellation of factors—the breadth of the claims, the nature of the invention, the state of the art, the level of ordinary skill, the predictability of the field, the direction the inventor provides, the presence of working examples, and the sheer quantity of experimentation the disclosure leaves to the reader. A specification that names a "convolutional neural network" to solve a prediction problem, but leaves the reader to guess at the architecture, the hyperparameters, and the training regime, may force exactly the kind of trial-and-error the Wands factors condemn. And the Supreme Court's enablement decision in Amgen Inc. v. Sanofi, 598 U.S. 594 (2023), tightened the screws on broad functional claims: a patentee who claims an entire genus by what it does must enable the full scope of that genus, not just a few examples. For an AI invention claimed in functional terms—"a model trained to predict stable, non-flammable electrolytes"—Amgen is a warning that the claim's reach cannot outrun the specification's teaching.

The written-description requirement adds its own demand: the specification must show that the inventor was in possession of the claimed invention, not merely that the inventor wished for it. Where the inventive work was done by an opaque model, possession can be genuinely hard to demonstrate. Several offices have pressed on exactly this point in practice. China's CNIPA has rejected applications that describe an AI's output without enabling its generation; the EPO's sufficiency requirement under Article 83 EPC can demand disclosure of training-data characteristics, architecture, and methodology for machine-learning inventions. The convergent lesson is that AI-assisted applications should be drafted to disclose enough of the technical pathway—the data, the constraints, the architecture and training methodology where relevant—to satisfy the most demanding office.

Here is the uncomfortable symmetry, and the reason inventorship and disclosure are two sides of one coin: the more autonomous the AI's role, the harder both problems become. The more the machine did, the weaker the human's claim to conception—and the harder it is to teach the public how to reproduce a result the inventors themselves cannot fully explain. There is even a perverse incentive lurking here, one the USPTO's patentability commentators have flagged: faced with hard disclosure and eligibility hurdles, companies may keep their best AI systems as trade secrets rather than patent them, starving the public of exactly the disclosure the patent system exists to extract. (On the choice between patenting and secrecy, see legal protection of software: copyrights, patents, trade secrets, and contracts and building a trade secret protection program from scratch.) The practical drafting answer is to err toward disclosure of the technical pathway and to claim around the human-contributed elements—a strategy we develop below.

And eligibility, lurking underneath

Even before written description and enablement come into play, an AI invention has to survive patent eligibility under 35 U.S.C. § 101. Many AI claims read as data analysis or mathematical computation, and courts and examiners may find them directed to an "abstract idea" under the two-step framework of Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014), and Mayo Collaborative Services v. Prometheus Laboratories, 566 U.S. 66 (2012). A claim caught at step one can still survive at step two if it recites "significantly more" than the abstract idea—an inventive concept tied to a concrete technical improvement, in the spirit of Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016), where a self-referential database table that improved how the computer worked cleared the bar. The USPTO's 2019 Revised Patent Eligibility Guidance and its later AI-specific examples push examiners to ask whether the claim integrates any abstract idea into a practical application. Dr. Marchetti's electrolyte, claimed as a physical composition of matter rather than as an algorithm, sidesteps most of this; but a claim to FORGE itself, or to the method of training it, lives squarely in Alice territory and must be drafted with a technical improvement front and center. Our deep dives on patent eligibility after Alice and on overcoming obviousness rejections under Section 103 carry these threads further; the latter matters because the EPO and USPTO alike will treat the routine application of a known AI method as obvious unless the inventive contribution is something more.

Turning Law Into Practice: Build the Record Before You Need It

Because the formal rule is settled and the practical rule turns on human conception, the single most valuable thing a company can do is generate, in the ordinary course of research, a clear record of what its people actually contributed. This is not paperwork for its own sake; it is the evidence that determines whether a patent issues, survives a validity challenge, or holds up when an acquirer's lawyers go looking for problems during diligence.

In practice, that means capturing—contemporaneously, not reconstructed years later—the technical judgments a human made at each stage. Who identified the specific technical problem, and on what insight? Who chose the AI system, and why that one? Did anyone design or modify the system for this problem? Who selected and curated the training data, and what judgment went into it? Who set the constraints, and what technical reasoning drove them? Who evaluated the outputs and, if there were many, who selected among them and on what basis? Who refined the result, and how was it integrated into the larger invention? A research process that can answer those questions in writing is a research process that produces patentable inventions; one that cannot is gambling that no examiner or adversary ever asks. The old engineer's habit of keeping a witnessed, dated lab notebook turns out to be the perfect prophylactic for the AI age—not because the law requires a particular form, but because the record of human judgment is the thing that proves conception. Our guidance on preparing an invention disclosure for your patent attorney translates these questions into a disclosure workflow.

Ownership is the next link in the chain, and it is easy to overlook. Establishing that a human invented does no good if that human's rights never reach the company. AI-driven research often involves outside contractors, academic collaborators, and employees in multiple countries, each with different default rules about who owns workplace inventions. A well-drafted assignment regime—present-tense "hereby assigns" language, compliant with each relevant jurisdiction's invention-assignment statutes—closes the gap. We cover the drafting in employee invention assignment agreements and the secrecy side in drafting enforceable non-disclosure agreements for technology transactions. For the deeper question of how AI-assisted inventions are owned once invented, see again AI-generated inventions: who owns what the machine creates.

Two further practices pay for themselves on cross-border filings. Draft for the most demanding jurisdiction, not the most permissive. Because China and the EPO can demand robust disclosure of the AI system itself—data characteristics, architecture, training methodology—an application written to satisfy them will usually satisfy everyone, while one written only for the lightest-touch office may fail abroad. And draft claims to foreground the human-contributed, technically inventive elements. Where the humans supplied specific technical insights and the AI supplied others, claims built around the human contributions are more defensible than claims that read as a description of what the machine did on its own. When an examiner pushes back—on inventorship, on eligibility, or on enablement—our guide to responding to patent office actions walks through the moves. And when the resulting portfolio has to be enforced or defended across borders, global patent litigation strategies and the lessons in global patent wars become the playbook.

Why Not Just Change the Law?

If the natural-person requirement is the obstacle, why not amend the statutes to let machines be inventors? Several proposals circulate. One would recognize AI inventorship outright and assign the resulting rights to the AI's owner or operator. Another would treat AI-assisted inventions as a special category with their own rules and a calibrated human-contribution threshold. A third would create an entirely new, sui generis form of protection tailored to machine-generated innovation, with its own term, requirements, and perhaps a shorter monopoly to reflect the lower human cost of creation.

None has gained real traction, and there are principled reasons for caution. The patent bargain trades disclosure and a limited monopoly for the incentive to invent and to publish. If no human ingenuity is being incentivized—if the machine would have produced the invention regardless of any patent reward—then a core justification for the monopoly evaporates. There are also line-drawing nightmares: who would owe the duty of candor to the patent office, who would sign the inventor's oath, and how would obviousness be judged against a hypothetical skilled person if the "inventor" is a system that can search chemical space faster than any human team? The USPTO, the UKIPO, and the EPO have all, after study, left the human-inventor rule in place, and major economies have consulted and declined to legislate. For now, the realistic near future is continued application of the human-inventorship rule, with guidance evolving at the margins. The U.S. swing from the 2024 Pannu-based approach to the 2025 tool-based approach is a vivid reminder that the margins can move quickly—and in either direction—even when the core rule does not.

Frequently Asked Questions

Can an AI system be named as an inventor on a patent anywhere? Not in any jurisdiction that examines patents. The United States, EPO, United Kingdom, Australia, China, Japan, and South Korea all require a human inventor. South Africa registered a DABUS patent, but only because its registration system grants patents without substantive examination; that grant is an outlier, not a recognition that machines can invent.

If I use ChatGPT or another AI tool to help develop an invention, can I still be the inventor? Almost certainly yes—provided you, a human, conceived the invention, with the AI serving as a tool. Under the USPTO's 2025 guidance, using AI is treated like using any other powerful instrument, no different in principle from a microscope or a simulation package. What matters is that a person formed the definite and permanent idea of the operative solution and exercised meaningful technical judgment, not that an AI was somewhere in the workflow.

What if the AI generated the actual solution and I just recognized it was valuable? This is the genuinely hard case. Merely appreciating the value of an output the machine produced—without contributing the inventive concept—may not be enough to qualify as the inventor, because conception has always belonged to whoever first holds the definite idea of the solution, not to whoever notices it is useful. If your recognition required real technical expertise (selecting a non-obvious candidate from thousands, understanding why it works, and refining it), you have a much stronger claim than someone who simply accepted the top result.

Are the USPTO's Pannu factors still relevant to AI inventions? Only obliquely. After the November 2025 guidance, the Pannu joint-inventorship factors apply when determining whether multiple humans are co-inventors. They no longer govern the single-human-plus-AI situation, because an AI cannot be a joint inventor at all. For a lone human inventing with AI, the test is ordinary conception.

What is the biggest practical risk for a company doing AI-driven R&D? Two risks, intertwined. First, naming an inventor who cannot show conception, leaving the patent vulnerable to invalidity or unenforceability—and to attack in litigation or diligence. Second, failing to disclose enough of the technical pathway (data, architecture, training methodology) to satisfy written-description and enablement requirements, especially at the EPO and CNIPA. Good contemporaneous documentation and demanding disclosure drafting address both.

Does this affect copyright too? This article is about patents and inventorship. AI authorship in copyright follows a related but distinct human-authorship rule, with its own line of Copyright Office decisions and litigation; see our coverage of copyright infringement claims against generative AI for that adjacent battlefield.

Conclusion: A Settled Rule, a Moving Target, and a Process Answer

On the question that makes headlines, the world has reached consensus: an inventor must be a human being. No major jurisdiction recognizes AI inventors, and appellate courts in the United States, the United Kingdom, and Australia have closed the door to getting there by judicial interpretation. Only legislatures can change that, and none has.

But the consensus on the headline masks genuine movement on the question that governs everyday practice—how much a human must contribute when AI does much of the work. The United States has now twice changed its formal answer, landing in late 2025 on a tool-based, conception-centered framework more permissive and more conventional than the 2024 approach it replaced. The EPO, the UK, and the major Asian offices reach broadly compatible results through their own doctrines, with China the most flexible on human contribution and among the strictest on disclosure. And running underneath all of it is the twin problem of teaching a black box—written description, enablement, and eligibility hurdles that grow harder precisely as the machine's role grows larger.

For a company like Helix Materials—and for the thousands of real companies it stands in for—the lesson is reassuringly practical. Inventions that emerge from genuine human-AI collaboration, where people frame the problems, shape the tools, exercise technical judgment, and recognize and refine the results, are patentable across the major jurisdictions today. What converts that abstract eligibility into an issued, enforceable patent is process and documentation: building, in the normal course of research, a clear contemporaneous record of what the humans contributed, and drafting disclosures robust enough for the most demanding office in the filing plan. The companies that do this will preserve their patent options as the law continues to shift around them. The ones that wait for perfect clarity will be waiting a long time—and inventing, in the meantime, without protecting what they invent.

Related Articles

This article is for general informational purposes only and does not constitute legal advice, nor does it create an attorney-client relationship. Patent law and AI guidance in particular are evolving rapidly and vary by jurisdiction; the discussion here may not reflect the most recent developments. Consult qualified patent counsel about your specific circumstances before acting.