In brief. Every major patent system to consider the question has held that an AI system cannot be an inventor—but that settled answer conceals the live one: when a human invents with AI, how much human contribution is enough? The U.S. answer changed dramatically in late 2025, when the USPTO scrapped its 2024 framework and recast AI as an ordinary tool, with inventorship judged by the century-old law of conception. This article traces the global DABUS litigation, explains the conception and joint-inventorship doctrines that decide these cases, walks through the new U.S. framework and its consequences, and lays out the documentation, disclosure, ownership, candor, trade-secret, and foreign-filing strategies that R&D-driven companies should adopt now. It is not legal advice; this area moves quickly, and specific situations call for qualified patent counsel.

It is two in the morning in a pharmaceutical lab that looks more like a server farm than a bench. No one is pipetting. A generative model has spent the night proposing kinase inhibitors against a protein target, scoring each against a fitness function, and discarding the failures. By dawn it has surfaced a molecule that no chemist on the team—and, as best anyone can tell, no chemist anywhere—has ever drawn. The structure is novel, non-obvious, and, the assays will soon confirm, startlingly effective. It is, in every commercial sense, an invention worth a fortune.

Now answer a deceptively simple question: who invented it?

The patent system has run for more than two centuries on an assumption so basic it was rarely stated—that inventions come from human minds, and that the inventor is the person who had the idea. The Constitution frames the whole enterprise around "Inventors" and their "Discoveries," authorizing Congress to "promote the Progress of Science and useful Arts" by securing exclusive rights for limited times. U.S. Const. art. I, § 8, cl. 8. For most of that history, the assumption never had to be examined. A microscope reveals what the eye cannot, but the insight stays human. A spreadsheet crunches numbers no human could, but the hypotheses come from people. Modern generative AI is different in a way that strains the old vocabulary: it does not merely accelerate human cognition; in some workflows it appears to originate the solution.

To keep the discussion concrete, we will follow one company throughout. Heliotrope Therapeutics is a venture-backed biotech whose discovery platform—an in-house generative model called HELIX—proposes candidate molecules against protein targets. The kinase inhibitor in the opening scene is HELIX's output. It becomes Heliotrope's lead candidate, potentially worth hundreds of millions of dollars. Dr. Amara Okafor, the medicinal chemist who defined the target profile, curated the training data, wrote the constraint parameters, and selected and refined HELIX's output, believes she invented it. Heliotrope's investors want a patent. Whether they can get one—and on what theory—is the subject of this article. Every doctrinal point below ultimately comes back to Dr. Okafor and what she can truthfully say she did.

This question has moved from seminar-room speculation to urgent controversy as AI capabilities have accelerated. Courts, patent offices, and legislatures across the globe are now deciding cases that test the foundations of intellectual property law, and the answers will shape innovation incentives for decades. This article addresses the patent side of the story. For the broader landscape of AI legal risk—liability, privacy, employment, antitrust, and contracting—see our companion overview of artificial intelligence's key legal issues; for a comparative deep dive into how different countries treat machine contributions, see our analysis of AI and inventorship from a global perspective.

Why "Who Is the Inventor?" Is a Legal Question, Not a Philosophical One

Before the doctrine, a word on why this matters so much—because the stakes are not abstract.

Inventorship is not a courtesy or a line on a press release. In U.S. patent law it is a legal status with hard consequences attached. The named inventors are, in the first instance, the owners of the patent right; everything downstream—assignments, licenses, the company's ability to sue infringers—runs through them. Get inventorship wrong, and the consequences cascade. A patent that names the wrong people, or omits a real inventor, can be held invalid or unenforceable. Inventorship defects are a staple of patent litigation precisely because they can take down an otherwise strong patent: an accused infringer who can show that a true inventor was left off, or a non-inventor named, has a path to defeating the patent entirely. (Our overview of patent infringement claims and defenses surveys the broader menu of invalidity attacks.)

So "who invented it?" is not a metaphysical riddle about machine consciousness. It is a question with a statutory answer, a body of case law behind it, and real money riding on getting it right. When the inventor is a person plus an extraordinarily capable software tool, the law has to decide how to characterize the human's role—and that characterization decides whether Heliotrope gets a defensible twenty-year exclusivity or a molecule any competitor can copy the moment its structure is published.

The Doctrine That Decides Everything: Conception

If you remember one word from this article, make it conception. In U.S. law, conception is the touchstone of invention. The Federal Circuit's classic formulation, drawn from Hybritech Inc. v. Monoclonal Antibodies, Inc., 802 F.2d 1367 (Fed. Cir. 1986), and elaborated in Burroughs Wellcome Co. v. Barr Laboratories, Inc., 40 F.3d 1223 (Fed. Cir. 1994), defines 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 hereafter to be applied in practice." An idea is "definite and permanent" when the inventor has a specific, settled solution in mind—not merely a hope, a goal, or a research problem, but a particular way of achieving the result that requires no more than ordinary skill to reduce to practice.

Two features of the conception doctrine do enormous work in the AI context.

First, conception is the mental part of invention—the eureka, the formation of the idea—as distinct from reduction to practice, the building or testing that confirms the idea works. Reduction to practice can be "actual" (you make the thing and it functions) or "constructive" (you file a patent application that fully describes how to make and use it). Crucially, you do not have to perform the reduction to practice yourself to be the inventor; the law has always allowed inventors to delegate construction and testing to technicians and others without those helpers becoming co-inventors. A scientist who conceives a molecule and hands it to a lab to synthesize is still the inventor. That principle, old as the patent system, turns out to be exactly the lever that lets a human "invent" with an AI: if the human conceived the definite-and-permanent idea, the machine that helped realize it is, in legal terms, doing the work of a skilled assistant or a piece of equipment.

Second, conception must be of the complete and operative invention. A mere wish—"I want a molecule that inhibits this kinase"—is not conception. Burroughs Wellcome makes the point vividly: an inventor has not conceived an invention until 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." When the answer still requires genuine invention to find—when you have a target but no solution—you have a research plan, not a conceived invention. This is why, as we will see, simply asking an AI to "solve problem X" and accepting whatever it returns is doctrinally dangerous: posing the problem is not conceiving the solution.

Hold those two ideas. Nearly everything in the modern AI-inventorship framework is an application of them.

The DABUS Saga: Testing Patent Law's Limits

No case has done more to clarify—and complicate—the landscape than the global litigation over DABUS, an AI system built by Dr. Stephen Thaler. DABUS, the "Device for the Autonomous Bootstrapping of Unified Sentience," is, in Thaler's telling, not a tool that implements human goals but a system that genuinely conceives inventions on its own. Whether that description is technically accurate is contested; what matters legally is that Thaler litigated as though it were true.

Thaler filed patent applications in more than a dozen jurisdictions, naming DABUS as the sole inventor of two innovations: a food container with a fractal surface said to improve grip and heat transfer, and an emergency light beacon that flashes in an attention-grabbing pattern. In each filing Thaler explicitly stated that he was not the inventor—that DABUS had conceived the inventions independently—and that his ownership claim derived solely from owning the machine. The applications were engineered as test cases. By refusing to name any human inventor, Thaler created a pure specimen of autonomous machine invention that could not be dodged through clever claim drafting or strategic silence about AI's role. Had he simply listed himself, as several offices pointedly observed, the cases might have ended very differently—or never begun.

United States: The Federal Circuit Speaks

The USPTO rejected the applications on a short ground: the Patent Act requires inventors to be "individuals," and an individual is a human being. Thaler took the fight through the federal courts to the Court of Appeals for the Federal Circuit, the specialized appellate court with nationwide jurisdiction over patents.

In Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), the court unanimously affirmed. Its analysis is a clinic in statutory textualism. The key provision is 35 U.S.C. § 100(f), which defines "inventor" as "the individual or, if a joint invention, the individuals collectively who invented or discovered the subject matter of the invention." Drawing on the Supreme Court's reasoning in Mohamad v. Palestinian Authority, 566 U.S. 449 (2012)—which held that "individual" in the Torture Victim Protection Act means a natural person—the Federal Circuit concluded that "individual," absent some clear contrary indication, presumptively means a human being. The Patent Act supplied no such indication; if anything, neighboring provisions reinforced it, referring to inventors by personal pronouns and requiring inventors to execute an oath stating that they believe themselves to be "the original inventor." Machines do not take oaths. Congress, the court reminded, is presumed to "say what it means and mean what it says."

The most important sentence in the opinion is the one the court did not decide. The Federal Circuit expressly declined to address "whether inventions made by human beings with the assistance of AI are eligible for patent protection." That careful limitation walled off the questions that matter most commercially—the AI-assisted questions—and left them to be worked out elsewhere. The Supreme Court denied certiorari in April 2023. A denial is not a merits ruling, but it signaled that the human-inventor requirement will stand absent an act of Congress.

United Kingdom: The Supreme Court's Definitive Ruling

The United Kingdom became the first major jurisdiction whose highest court ruled on AI inventorship, and its reasoning is the most thorough of the lot. After the UK Intellectual Property Office rejected Thaler's 2018 applications, he climbed through the High Court and Court of Appeal to the Supreme Court. In Thaler v. Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49, the court unanimously ruled against AI inventorship—and addressed two distinct issues that map neatly onto the deeper problem.

First, inventorship. The court held that an "inventor" under the Patents Act 1977 must be a natural person. Reading the statute as a coherent whole, it found that the legislative scheme presupposes human inventors with legal capacity. Parliament did not imagine machine invention in 1977, but the structure and language of the Act compelled the human reading anyway.

Second—and more revealing—ownership. Could Thaler claim the inventions simply because he owns DABUS, by analogy to an employer owning an employee's inventions? The court said no, on several grounds. Section 7 of the Act exhaustively lists who may obtain a patent: the inventor, and those who derive rights from the inventor through a rule of law (such as employment) or an enforceable agreement. Owning the machine that produced the idea is nowhere on that list. The court also refused to extend the old common-law doctrine of accession—the rule that the owner of a thing owns what it naturally produces, as the farmer owns the calf her cow bears—to intangible inventions, noting that accession has always been about new tangible property. Lord Kitchin stressed that the court was interpreting existing law as applied to Thaler's particular claims; whether AI-generated inventions should be patentable, and how the law ought to evolve, are questions for Parliament. The ownership holding matters because it forecloses the intuitive "I built it, so I own its output" theory that many AI developers instinctively reach for.

European Patent Office: The Legal Board of Appeal

At the European Patent Office, the Receiving Section rejected the applications, and the Legal Board of Appeal affirmed in decisions J 8/20 and J 9/20 (December 2021, with full reasoning published in 2022). The Board grounded itself in Article 81 EPC, which requires applications to designate the inventor, and Rule 19(1), which requires the designation to give the inventor's family name, given names, and address—formal requirements that presuppose a natural person with legal capacity, which an AI lacks.

But the Board also pointed, almost helpfully, at a door Thaler had bolted shut from the inside. It observed that it was "not aware of any case law which would prevent the user or the owner of a device involved in an inventive activity to designate himself as inventor under European patent law." Translation: had Thaler simply named himself—even for an invention he attributed to DABUS—the EPO might well have accepted the designation at the formalities stage. The reason is structural: the EPO's review of inventor designations is essentially formal, and the Office does not investigate whether the named inventor truly invented. Disputes over who actually invented are left to national courts. That formality creates a theoretical pathway to a European patent on an AI-heavy invention through the designation of a human whose creative role was thin—a pathway that carries its own risks under national law, and, for any U.S. counterpart application, under the duty of candor discussed below.

Australia, South Africa, Germany, Switzerland: A Fragmented Map

The DABUS filings produced a genuinely fragmented global picture, which is itself a strategic fact for any company filing internationally.

South Africa granted a patent naming DABUS as inventor in 2021—the first country to do so. But South Africa runs a registration system without substantive examination, so the grant reflects the absence of a gatekeeper rather than a considered ruling that machines can invent.

Australia briefly broke from the common-law pack when a single judge of the Federal Court ruled for Thaler in 2021, finding nothing in the Patents Act 1990 that required a human inventor. The Full Court of the Federal Court reversed in Commissioner of Patents v. Thaler [2022] FCAFC 62, holding that "inventor," properly understood, denotes a natural person. The High Court of Australia declined special leave, leaving the Full Court's view in place.

Germany's Federal Patent Court took a middle path that is worth noting because it points toward where harmonization might eventually land: the named inventor must be a natural person, but the AI system may be mentioned alongside the human—acknowledging the machine's role while preserving the human requirement.

Switzerland's Federal Administrative Court, ruling in 2024, accepted a subsidiary argument that the EPO had hinted at: a human who recognizes the significance of an AI-generated output and acts on it may be treated as the inventor, allowing the application to proceed with the human named.

The throughline is convergence on the result—no machine inventors—reached by strikingly different routes, with meaningful disagreement at the edges about how a human may be substituted in. We map these national differences in detail in our companion article on global perspectives on machine contributions to innovation, and we situate them within broader global patent litigation strategy.

USPTO Guidance: A Framework That Changed Its Mind

While the courts disposed of the narrow question—no AI inventors—the broad question festered: what about the ordinary, commercially dominant case in which a human and an AI collaborate? How much human contribution is enough? The USPTO answered through administrative guidance, and its answer flipped within two years as administrations changed. Understanding both versions matters, because patents prosecuted under the first framework are out there, and the doctrinal debate between the two is the live one.

The February 2024 Guidance: AI Through the Lens of Joint Inventorship

In February 2024, responding to directives in the Biden administration's Executive Order on Safe, Secure, and Trustworthy AI, the USPTO issued its first comprehensive inventorship guidance for AI-assisted inventions, published at 89 Fed. Reg. 10,043 (Feb. 13, 2024). Its central, sensible move was to confirm that AI-assisted inventions are not categorically unpatentable—AI's mere involvement does not poison the well—while cautioning that "an AI system—like other tools—may perform acts that, if performed by a human, could constitute inventorship under our laws."

To decide when a human qualifies as inventor of an AI-assisted invention, the 2024 guidance borrowed a tool from a neighboring doctrine: the factors of Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998). To understand the move—and the later reversal—you have to understand Pannu. It is a joint-inventorship case. When two or more people contribute to a single invention, 35 U.S.C. § 116 makes them joint inventors even if they did not work together physically, did not contribute equally, and did not each contribute to every claim. Pannu gave courts a test for separating real co-inventors from mere helpers. To be a joint inventor, a person must (1) contribute in some significant manner to the conception or reduction to practice of the invention, (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 to the real inventors well-known concepts or the current state of the art.

The 2024 guidance treated the human-plus-AI pairing as if it were a joint-inventorship problem, asking whether the human's contribution cleared the Pannu bar against the backdrop of what the AI did. Five guiding principles fleshed it out: (1) using an AI system does not negate inventorship if the human contributed significantly; (2) merely recognizing a problem or holding a general goal is not enough; (3) merely presenting a problem to an AI is not enough—the human must do more than prompt; (4) developing an essential building block (designing the AI, curating the training data, constructing the prompts and parameters) can be a significant contribution even without involvement in every later step; and (5) "intellectual domination"—mere ownership or control of the AI—confers nothing.

Applied to Heliotrope, the 2024 framework made Dr. Okafor's inventorship a genuinely fact-intensive question. Her target profile, curated training set, and constraint parameters looked like "essential building blocks" under the fourth principle. But an examiner could still probe whether her contribution to the claimed molecule—the specific structure HELIX surfaced—was "significant" and "not insignificant in quality" against the whole. The framework invited a claim-by-claim audit weighing human contribution against machine contribution, which created prosecution risk and, perversely, an incentive to downplay HELIX's role in the very documents where candor is required.

The November 2025 Revised Guidance: AI as an Ordinary Tool

The ground shifted with the USPTO's Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54,636 (Nov. 28, 2025), issued under Director John Squires and implementing Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence." The revised guidance rescinded the 2024 guidance in full and reframed the entire inquiry.

Its first and central move is to abandon Pannu in the AI context. The reasoning is clean: Pannu and its progeny address joint inventorship among multiple natural persons. But an AI system cannot be an inventor at all—that is the settled holding of Thaler v. Vidal. So when a single human invents with AI assistance, there is no second inventor in the picture and therefore no joint-inventorship question to analyze. Importing a joint-inventorship test was, on this view, a category error: it treated the machine as a quasi-co-inventor whose share had to be measured, when the machine is not an inventor in any degree.

Its second move is to return the analysis to conception—the doctrine we opened with. The touchstone is simply whether a natural person formed "a definite and permanent idea of the complete and operative invention"—a specific, settled solution, not a goal or a research plan. If the human conceived the claimed invention, she is the inventor. Full stop.

Its third move follows naturally: recharacterize AI as an instrument. Generative models and all other AI systems are, the guidance says, "analogous to laboratory equipment, computer software, research databases, or any other tool that assists in the inventive process." Just as inventors have always drawn on the services, ideas, and aid of others—and on instruments and software—without those sources becoming co-inventors, they may use AI without the AI becoming one. The same inventorship standard now governs all inventions, AI-assisted or not. There is no special "AI track."

The practical consequence is significant. By treating AI as no different in principle from a mass spectrometer or a literature database, the guidance removes the special scrutiny the 2024 framework invited. For Heliotrope, the question is no longer how Dr. Okafor's contribution stacks up against HELIX's. It is simply whether she conceived the claimed molecule. If, through her target definition, parameter design, candidate selection, and structural refinement, she formed a definite and permanent idea of the complete and operative compound as claimed, she may be named inventor—without a Pannu audit of the machine's share.

Two cautions keep this from being a free pass. First, conception must be genuine, not recited. An applicant cannot launder thin human involvement by reciting magic words; the file has to support a real story of a human forming the idea. Second, the duty of candor survives intact (more on that below). The reframing makes the standard friendlier to AI-assisted work; it does not relax the truthfulness the system demands.

Critics argue the new approach may prove too permissive—blessing patents where the human contribution was thin and the AI did the creative heavy lifting. Supporters answer that the century-old conception standard is robust enough to police abuse, that it is the same standard applied to every other tool, and that imposing special burdens on AI-assisted inventions would handicap American innovators against competitors abroad. Both sides are right about something, which is why this is the debate to watch. How examiners actually probe conception in AI-heavy files will be worked out file by file; our guide to responding to patent office actions covers the general craft of overcoming rejections.

A final framing point: inventorship is only one gate. Even a properly attributed AI-assisted invention must still clear patent-eligible subject matter under 35 U.S.C. § 101—a separate and historically treacherous hurdle for software-adjacent inventions. Many AI inventions get tangled in the abstract-idea exception, and the USPTO's own AI patentability guidance treats § 101 as the central battleground for machine-learning claims. We treat that problem at length in patent eligibility after Alice and survey the underlying analysis in overcoming obviousness rejections under Section 103. Inventorship and eligibility are independent locks on the same door; you have to pick both.

When More Than One Human Is Involved: Joint Inventorship in the AI Lab

Real AI-assisted projects rarely have one human and one machine. They have teams—a chemist who set the target, a data scientist who curated the training corpus, an ML engineer who tuned the model, a junior researcher who ran the screen. The 2025 guidance scraps Pannu for the human-versus-machine question, but Pannu and § 116 are alive and well for the human-versus-human question. Among the people, ordinary joint-inventorship law still decides who belongs on the patent.

That distinction is easy to lose and expensive to get wrong. Suppose at Heliotrope the lead molecule's claimed structure traces to Dr. Okafor's hypothesis and refinement, but a colleague, Dr. Reyes, independently conceived a key structural modification that appears in a dependent claim. Under § 116 and Pannu, Reyes is likely a co-inventor of at least that claim—and inventorship in the United States is determined claim by claim, so a person who contributed to even one claim is an inventor of the patent. Leave Reyes off, and the patent carries an inventorship defect. Add a manager who merely supervised and approved budgets, and you have named a non-inventor. Either error is a live invalidity or unenforceability theory in litigation.

Two practical rules follow. First, decide inventorship by reference to the claims, not the org chart or the credit politics. Who conceived the subject matter of these claims? Contributing money, lab space, encouragement, or general supervision does not make someone an inventor; conceiving claimed subject matter does. Second, the AI does not change the human-side analysis. You run the same § 116/Pannu inquiry among the people that you would in any team invention; you simply do not add the machine to the roster. Our discussion of employee invention assignment agreements explains why getting the human roster right also matters for clean chain of title—because each true inventor must assign, and a missing inventor is a missing signature.

Ownership, Assignment, and the Chain of Title

Naming the right inventors is necessary but not sufficient. In the United States, inventorship and ownership start in the same place—the inventors own the invention by default—but they diverge the moment the inventors assign. For a company, the entire value of the patent depends on a clean chain of title from each human inventor to the entity.

This is where the UK Supreme Court's ownership holding in Thaler echoes loudly in U.S. practice. You cannot own an AI's output by owning the AI, because the AI is not an inventor and has nothing to assign. Ownership has to run through people. Concretely, that means each named human inventor must execute an assignment to Heliotrope, and Heliotrope's invention-assignment agreements with its employees must be drafted to capture exactly this kind of work. A well-drafted present assignment ("hereby assigns") is materially stronger than a promise to assign in the future, a distinction the Supreme Court underscored in Stanford v. Roche, 563 U.S. 776 (2011), where a defective assignment clause cost a university its rights. For AI-heavy R&D, the agreements should sweep in inventions "conceived or reduced to practice with the assistance of any software or automated system," so there is no argument that an AI-assisted invention falls outside the grant. We work through the drafting in detail in employee invention assignment agreements and in our overview of popular legal documents for startups.

There is a subtler trap for multinationals. If a true inventor is omitted in the United States but the omission is uncorrected, that person never assigned—and an unassigned co-inventor can, in principle, license the patent to a competitor without the company's consent and without a duty to account. The clean-roster discipline urged above is therefore not just an invalidity-avoidance measure; it is a chain-of-title measure. A company that later wants to license its AI-derived portfolio will run straight into these questions in diligence, as we discuss in how to license your patent—from valuation to term sheet.

Fixing It Later: Correction of Inventorship

Because inventorship is determined by the law, not by the applicant's first guess, the system builds in a way to fix honest mistakes. Both pending applications and issued patents can have inventorship corrected: 35 U.S.C. § 116 governs correction in pending applications, and 35 U.S.C. § 256, implemented through 37 C.F.R. § 1.324, governs uncontested correction of issued patents. The procedure is, by patent standards, fast, inexpensive, and low risk; it can be invoked at almost any time, including during litigation.

One detail is especially relevant to AI work. Before the America Invents Act, § 256 correction was available only for errors that arose without deceptive intent—a requirement that could trip up a good-faith correction if the other side cried fraud. The AIA removed that "without deceptive intention" condition, so correcting inventorship no longer requires the parties to prove their innocence as a precondition. That is genuinely useful here: as AI tools and the governing standard both evolve, a company that, in good faith, named inventors under one understanding and later concludes the roster should change has a clean statutory mechanism to fix it. Uncontested correction generally requires the named inventors, any omitted true inventors, and the assignees to agree to the change (37 C.F.R. § 1.324). Contested inventorship—where the parties disagree about who invented—is resolved by a court under § 256, and that is exactly the kind of fight you want to avoid by getting the analysis right up front.

The lesson is not "name anyone; we can fix it later." It is that the safety valve exists for honest error, and that the smart play is a defensible roster from day one, with correction held in reserve for genuine surprises.

The Duty of Candor: Why You Cannot Just Hide the Machine

Running beneath all of this is an obligation that the 2025 guidance pointedly left untouched: the duty of candor and good faith owed to the USPTO by everyone substantively involved in prosecuting an application (37 C.F.R. § 1.56). Each inventor and each attorney must disclose information material to patentability. Breach the duty with intent to deceive, and the patent can be held unenforceable for inequitable conduct—a defense so devastating it is often called the "atomic bomb" of patent law, because it can render the entire patent unenforceable and, in egregious cases, expose the patentee to attorney-fee awards. The modern standard, set in Therasense, Inc. v. Becton, Dickinson & Co., 649 F.3d 1276 (Fed. Cir. 2011) (en banc), requires both materiality (but-for materiality, generally) and a specific intent to deceive—a high bar, but a real one. Our deep dives on inequitable conduct and on finding evidence of inequitable conduct explain the doctrine in full.

Here is why this matters for AI inventorship. The temptation the 2024 framework created—minimize the machine's role so the human's contribution looks larger—is exactly the temptation candor forbids. Naming the right inventors and being truthful about how the invention came to be are not in tension; they are the same obligation. An applicant who shades the record to obscure AI's contribution, or who names a human inventor who did not actually conceive the claimed subject matter, is courting an inequitable-conduct problem that no favorable inventorship standard can cure. The 2025 guidance makes the standard more accommodating; it does not make lying about how you cleared the standard any safer. If anything, the friendlier the standard, the less excuse there is to misrepresent the facts—because under a conception-centered test, an honest account of a real human contribution should usually suffice.

A Problem Hiding in Plain Sight: Section 112 and the Black Box

Inventorship gets the headlines, but AI-assisted inventions face a quieter doctrinal hazard that can be just as fatal: the disclosure requirements of 35 U.S.C. § 112. A patent must contain a written description showing the inventor possessed the claimed invention, and an enabling disclosure teaching a person of ordinary skill how to make and use it without undue experimentation. The Supreme Court reinvigorated enablement scrutiny in Amgen Inc. v. Sanofi, 598 U.S. 594 (2023), holding that broad functional claims must enable the full scope of what they claim.

AI strains both requirements in a particular way. Machine-learning systems are often "black boxes": the engineers know the training data, the model architecture, the inputs, and the outputs, but the actual decision logic the model applies is, in a meaningful sense, unknown even to its creators. If a claim recites what an AI-derived system does in functional terms but the specification cannot explain how to reproduce it across the claimed scope, the patent risks invalidity for lack of enablement or written description—exactly the Amgen problem, transplanted to software. Two practical consequences follow for Heliotrope. For the molecule, § 112 is manageable: the claimed compound has a definite structure, and the specification can teach synthesis and use in the ordinary way—the fact that HELIX proposed it does not make it harder to describe. For claims directed to the discovery method or the model itself, the black-box problem bites, and the application must be drafted with unusual care to disclose enough architecture, training methodology, and parameters to satisfy § 112 across the claim's full breadth. This is one more reason to think hard, asset by asset, about what to claim and whether to claim it at all—a question that leads directly to trade secrets.

Practical Strategies for Companies Deploying AI in R&D

The framework is unsettled—a court decision, fresh guidance, or legislation could move it—so the right posture is a set of practices that protect the company under current law while staying robust to change. None of what follows is exotic. It is disciplined patent hygiene, adapted for a world where one of the lab's most productive contributors is a piece of software.

Document Human Contributions Rigorously

The single most important practice is rigorous, contemporaneous documentation of human contributions throughout the research process. Good records support inventorship during prosecution, defend the patent against later validity and inventorship challenges, and preserve institutional knowledge across personnel turnover. The record should capture not just what the AI produced but what humans did to direct, interpret, refine, and apply those outputs: formulating the problem and the target specification; selecting and curating training data; designing prompts, queries, and constraints; evaluating outputs and choosing which to pursue; modifying AI proposals; testing and validating; and integrating outputs into a complete, working invention. Lab notebooks, dated emails, meeting notes, and version-controlled logs of AI interactions all build the file.

The discipline matters more under the 2025 guidance than before, not less. Because the inquiry is now whether the human conceived the claimed invention, the file should tell a coherent story of conception—Dr. Okafor's hypothesis, her design choices, her selection and refinement of the lead—rather than a bare log of machine outputs with a name attached at the end. A good rule of thumb: imagine a litigator deposing Dr. Okafor years from now and asking, "When did you form a definite and permanent idea of this molecule, and how do we know?" The record should answer that question cleanly. (The deposition is not hypothetical in spirit; named-inventor depositions are a standard feature of patent litigation, and inventors are routinely cross-examined on precisely when and how they conceived.)

Update Invention Disclosure Processes

Most companies funnel potentially patentable work through invention-disclosure forms. Those forms should now address AI involvement explicitly: whether AI systems were used; which tools; what prompts or queries were submitted; what the AI generated; how the human inventors evaluated and selected among outputs; what modifications they made; and how the final, claimed invention differs from the raw output. This information lets patent counsel assess whether the proposed inventorship will withstand scrutiny, draft an application that characterizes the human contribution accurately, and build a contemporaneous record to consult if inventorship is later challenged—where, recall, improper inventorship can render a patent unenforceable. Our practical guide to preparing an invention disclosure for your patent attorney provides a template that adapts readily to AI-assisted work; the only real change is adding a candid "AI involvement" section and resisting the urge to soft-pedal it.

Structure Workflows to Preserve—and Demonstrate—Conception

Beyond documentation, companies can design their AI-assisted research to maximize and surface human conception. Rather than handing the AI an open-ended problem and passively accepting whatever comes back—the doctrinally weakest posture—researchers should engage iteratively, providing specific direction and exercising judgment throughout. At Heliotrope, Dr. Okafor does not ask HELIX to "find a drug." She begins with a defined hypothesis about the biological target and the properties an effective inhibitor must have; uses HELIX to screen candidates against that specification; personally evaluates the most promising structures; identifies specific improvements; and directs the model to explore particular variations. Each iteration layers human conception onto the result. Recall the conception doctrine: posing a problem is not conceiving the solution, but settling on a specific, operative structure is. A directed, iterative workflow is the difference between "I asked a machine and it answered" and "I conceived this, using a machine to help me realize it." The bonus is that this approach usually produces better science, because current AI excels at optimization within well-defined parameters but lacks the broad understanding and creative flexibility humans bring to defining the problem in the first place. This is a recurring theme in how counsel approach the legal landscape of AI-driven development.

Consider Trade-Secret Protection Strategically

Not every innovation needs a patent—and for some AI outputs, a patent is the wrong tool. Trade-secret protection requires only that the information derive independent economic value from not being generally known and that it be subject to reasonable secrecy measures (the standard under the Defend Trade Secrets Act, 18 U.S.C. § 1836, and the state Uniform Trade Secrets Act). It carries no fixed term, demands no public disclosure, and—crucially for our purposes—does not care who or what conceived the information. Trade-secret law sidesteps the inventorship question entirely.

That makes it especially attractive for AI assets that are hard to reverse-engineer from a shipped product: a training methodology, a model architecture, a fitness function, a curated dataset, a manufacturing process the product does not reveal. For those, secrecy may dominate patenting—and it neatly dodges both the inventorship question and the § 112 black-box problem discussed above (you need not enable what you never disclose). The calculus flips for Heliotrope's molecule, whose structure will be disclosed in regulatory filings and is readily analyzable once on the market; secrecy is no protection for a compound the world can characterize, so patent protection remains essential there. The right answer is asset-by-asset, and the two regimes are complementary, not rivals—the model stays secret, the molecule gets patented. We cover the operational side in building a trade-secret protection program from scratch, the distinctive challenges of distributed research in trade secrets in the age of remote work and cloud computing, and the full menu of options in legal protection of software—copyrights, patents, trade secrets, and contracts.

Mind the Training Data—and the Outputs You Feed In

A risk specific to AI R&D, and easy to overlook in an inventorship discussion, is the provenance of the inputs. If HELIX was trained on data scraped without rights, or if proprietary third-party information found its way into a prompt, the resulting invention can carry latent legal liabilities that have nothing to do with who invented it—copyright exposure on the training corpus, trade-secret contamination, or contractual breach. The wave of copyright infringement litigation against generative-AI developers over training data, and the law of data scraping after hiQ v. LinkedIn, both bear directly on what data is lawfully available to train an R&D system like HELIX. Clean inputs are part of a clean invention.

Audit Foreign-Filing Strategies for Inventor Consistency

Divergent national treatment of AI inventorship complicates multinational strategy, and the 2025 USPTO guidance adds a specific trap: benefit and priority claims require inventor continuity with a natural person. A U.S. application generally cannot properly claim priority to a foreign application whose only named inventor is an AI system, because there is no human inventor to carry the priority chain. Companies should therefore review the inventorship requirements of every target jurisdiction before filing, ensure consistent identification of the same human inventors across the family, and—if an application was filed in a jurisdiction that tolerated an AI designation (recall South Africa)—consider whether correction is possible before filing where human inventors are required. The goal is a single, consistent inventorship story told in every office: same humans, same account of conception, everywhere. These consistency questions also surface in licensing diligence, as we note in licensing your patent from valuation to term sheet, and they are part of the broader discipline of global patent litigation strategy.

Worked Examples: Where the Line Falls

Doctrine is easier to feel than to state. Three short hypotheticals—all clearly hypothetical—show how the conception standard sorts cases.

Hypothetical 1: The bare prompt (likely no human inventor). A researcher types into a general-purpose model, "Design a more efficient heat exchanger," and patents whatever it returns unchanged. Under both the old and new frameworks, this is the weak case. Posing a research goal is not conceiving a complete and operative invention; the human supplied the problem, the machine supplied the solution, and no natural person formed a definite and permanent idea of the claimed exchanger. There may be no valid inventor at all.

Hypothetical 2: Dr. Okafor's molecule (likely a human inventor). Dr. Okafor defines a specific biological target and a profile of required properties, curates training data toward that target, sets constraint parameters, runs an iterative screen, personally selects a lead from among candidates, identifies a structural modification to improve potency, and directs HELIX to explore that variation—arriving at the claimed compound. She can tell a coherent story of forming a definite and permanent idea of the operative molecule, using HELIX as the instrument. Under the 2025 conception standard, she is well positioned to be named inventor; the machine is doing the work of an extraordinarily capable lab.

Hypothetical 3: The serendipitous output nobody understands (the hard middle). HELIX returns an unexpected structure that works, but no human can explain why it works or claims to have conceived that particular structure before seeing it. Here the cases—and the two USPTO frameworks—diverge most. A defensible inventorship claim depends on whether a human, after seeing the output, recognized its significance and then formed a definite and permanent idea of the complete and operative invention (the theory Switzerland's court accepted), and on how much human refinement and selection went into the claimed subject matter. This is the zone where contemporaneous documentation is decisive: the difference between a valid patent and an invalid one may come down to whether the file shows a human mind settling on the claimed solution, or merely a human hand reaching for whatever the machine produced.

The pattern across all three: the more the human defines, selects, refines, and settles, the safer the patent; the more the human merely prompts and accepts, the more precarious it becomes.

Legislative Proposals and the Future of AI Inventorship

The current framework—human inventors, AI as tool—reflects interpretations of statutes written before anyone imagined a molecule-designing machine, not a considered legislative choice about AI innovation policy. Congress could rewrite the rules, and the debate is genuinely two-sided.

Proponents of recognizing AI inventorship, or of a new sui generis protection for AI-generated innovations, advance three arguments. The patent system exists to promote innovation, and refusing protection to AI-generated inventions may dampen investment in AI research tools or, worse, push their outputs into secrecy rather than disclosure—undermining the very "open exchange of information" the patent bargain is meant to encourage. The current framework also creates a perverse transparency problem: if candid disclosure of AI's role might jeopardize validity, applicants are tempted to minimize the machine's apparent contribution, degrading the accuracy of the public record. And as AI grows more capable, the line between human-conceived and AI-conceived inventions may come to seem arbitrary—an artifact of how a given lab structured its workflow rather than any meaningful difference in creative contribution.

Opponents answer with equal force. The patent system has always rewarded human ingenuity, and extending inventorship to machines would transform that understanding without evidence the change is needed. Patents confer a twenty-year power to exclude; granting that power over machine-generated innovations—potentially produced at massive scale and trivial marginal cost—could concentrate economic power and impede progress rather than promote it. Skeptics also question whether current systems "invent" in any meaningful sense: they process training data through mathematical operations defined by human programmers, and the creativity, if any, may reside in the people who built and directed them. Finally, inventorship carries duties an AI cannot bear—the duty of candor, assignment obligations, amenability to legal process, the capacity to take an oath. A bare extension of "inventor" to machines would leave those roles unfilled.

On the legislative front, the most consequential pending effort is the Patent Eligibility Restoration Act (PERA), reintroduced in 2025—but note carefully that it addresses eligibility, not inventorship. By replacing the judge-made exceptions to § 101 with a narrow statutory list, PERA would make many AI methods and systems easier to patent, easing the § 101 hurdle discussed above, without changing who may be named as inventor. Congress has held hearings on AI and intellectual property, and the bipartisan Senate AI Working Group's 2024 roadmap recommended further study of AI's effect on the patent system. But as of this writing, no bill has been introduced that would redefine "inventor" to include machines or create a dedicated regime for AI-generated innovations. Given the pace of change, that legislative caution may be wise: premature codification could lock in rules poorly matched to capabilities that do not yet exist.

The Mirror in Copyright—and the Rest of the IP Map

The inventorship debate is not an island. Its mirror image runs through copyright, and the parallel is exact. The U.S. Copyright Office has consistently held that copyright requires human authorship, refusing registration for works generated autonomously by AI while allowing registration where human creativity substantially shaped a work that contains AI-generated elements. The D.C. Circuit affirmed the human-authorship requirement in 2025 in Thaler v. Perlmutter—litigation brought, fittingly, by the same Stephen Thaler, this time over an image he attributed to his AI. The structural lesson is identical across both regimes: machines cannot be authors or inventors, but human-plus-AI collaboration can yield protectable works and patentable inventions, judged by the human contribution. Our copyright FAQs and overview of copyright vs. trademark vs. patent vs. trade secret place these regimes side by side.

The ripples extend outward. The emerging law of digital likeness—deepfakes, AI avatars, and celebrity likeness—presents yet another facet of the same underlying strain: legal frameworks built for human creativity stretching to accommodate machine-generated content. AI inventorship is one chapter of a much larger story about how intellectual property law metabolizes the machine.

International Harmonization Challenges

The global character of modern innovation makes coordination on AI inventorship increasingly important, and increasingly difficult. A company developing AI-generated drug candidates needs protection in markets worldwide, and a patchwork of inconsistent rules raises both cost and risk. WIPO has convened its Conversation on Intellectual Property and Frontier Technologies, gathering national offices, academics, and stakeholders to explore convergence, but it has no power to mandate uniform rules. The IP5 offices—the USPTO, EPO, and the Japanese, Korean, and Chinese offices—have engaged through expert roundtables that identify common concerns without resolving the underlying divergence. The EPO, for its part, addresses AI inventions primarily through patentability in its Guidelines for Examination, requiring technical character and a contribution to a technical problem, as for other computer-implemented inventions—an approach that accommodates many AI inventions without confronting the deeper question of whether machines "invent."

For companies operating globally, the practical mandate is the one urged above: a filing strategy that respects jurisdictional differences while keeping the core inventorship story consistent across every office. Experienced multi-jurisdictional patent counsel is not a luxury here; it is the difference between a coherent global portfolio and an unenforceable one. The same lessons we draw in advising on standard-essential patents and FRAND licensing—early strategic planning, consistent documentation, attention to national variation—apply with full force to AI-assisted inventions.

Ethics and Innovation Policy

Step back from doctrine and the debate becomes one about incentives and human welfare. The patent bargain assumes that exclusive rights for a limited term encourage innovation by letting inventors recoup their investment before facing competition. If AI can generate innovations at massive scale with minimal human input, the bargain may need rethinking. Society benefits from innovation, but it also benefits from the rapid spread of knowledge; if AI-generated inventions can be produced cheaply and abundantly, the welfare-maximizing policy might favor open availability over twenty-year monopolies. Yet AI research itself demands heavy investment in computing, data, and system development—as it plainly does today—so some exclusivity may be necessary to justify the spend. The open question is whether traditional patents are the right instrument, or whether a new regime, shorter in duration or narrower in scope, would better balance incentive against access.

These questions have no easy answers, and thoughtful people disagree. The current framework—human inventors, AI as tool—may be a pragmatic compromise that preserves incentives for AI development while keeping intellectual property human-centered. Or it may prove an inadequate response to technological change. What seems clear is that the choice should be made deliberately, by legislatures weighing the policy stakes, rather than backed into by accident from statutory language drafted before anyone imagined a machine that could design a molecule.

Frequently Asked Questions

Can an AI be named as the inventor on a U.S. patent? No. Under 35 U.S.C. § 100(f) and the Federal Circuit's decision in Thaler v. Vidal, an inventor must be a natural person. Every major patent system to consider the question has reached the same conclusion, though by different routes.

Are AI-assisted inventions patentable at all? Yes. Both the USPTO's 2024 guidance and its 2025 revised guidance confirm that AI's involvement does not categorically disqualify an invention. What matters is whether a natural person conceived the claimed invention; the AI is treated as a tool the human used.

What changed when the USPTO revised its guidance in 2025? The 2024 guidance analyzed AI-assisted inventions through the Pannu joint-inventorship factors, asking whether the human's contribution was "significant" relative to the AI's. The November 2025 Revised Inventorship Guidance rescinded that approach, reasoning that Pannu governs joint inventorship among people and has no role where the only candidate inventor is human. The revised guidance returns the inquiry to traditional conception and treats AI as an ordinary instrument, like lab equipment or software.

What does "conception" actually require? A natural person must form "a definite and permanent idea of the complete and operative invention"—a specific, settled solution, not a goal or research plan (Burroughs Wellcome v. Barr; Hybritech v. Monoclonal Antibodies). Posing a problem to an AI is not conception; settling on a specific, operative solution is.

If I just prompt an AI and patent whatever it returns, am I the inventor? Probably not. Merely presenting a problem and accepting an unmodified output is the weakest posture under both frameworks, because no human formed a definite and permanent idea of the claimed invention. The more you define the problem, set parameters, select among outputs, and refine the result, the stronger your inventorship claim.

Who owns an AI-assisted invention? The human inventors own it by default; the company acquires ownership through assignment. You cannot own an invention merely by owning the AI that helped produce it—the UK Supreme Court rejected exactly that theory in Thaler, and U.S. law reaches the same place because the AI is not an inventor and has nothing to assign. Clean assignments from each human inventor are essential. See employee invention assignment agreements.

What happens if we name the wrong inventors? An inventorship defect can render a patent invalid or unenforceable and is a common litigation attack. Honest errors can be corrected—§ 116 for pending applications, § 256 and 37 C.F.R. § 1.324 for issued patents—and the AIA removed the old "without deceptive intention" precondition. But correction is a safety valve for good-faith mistakes, not a substitute for getting the roster right.

Could naming a human to hide the AI's role get us in trouble? Yes. The duty of candor (37 C.F.R. § 1.56) survives the 2025 guidance untouched. Misrepresenting how an invention came about, or naming a human who did not actually conceive the claimed subject matter, risks rendering the patent unenforceable for inequitable conduct under Therasense. Truthfulness and correct inventorship are the same obligation. See inequitable conduct in patent prosecution.

Should we patent our AI model, or keep it secret? Often, keep it secret. Trade-secret protection (under the DTSA, 18 U.S.C. § 1836, and state law) requires no disclosure, has no fixed term, and ignores who conceived the information—so it sidesteps both the inventorship question and the § 112 black-box disclosure problem. It suits assets hard to reverse-engineer, like training methods and model architectures. Patenting still makes sense for assets that will be publicly disclosed anyway, like a drug's molecular structure. See building a trade-secret protection program from scratch.

Does the U.S. position match the rest of the world? On the narrow question, yes—no jurisdiction's highest authority allows machine inventors (South Africa's registration grant aside). But the routes and the edges differ: the UK rejected the ownership-by-accession theory, Germany allows the AI to be mentioned alongside a human, and Switzerland accepted a human who recognizes an AI output's significance as the inventor. File with the differences in mind. See AI and inventorship—global perspectives.

Is Congress about to change any of this? Not the inventorship rule, as far as anyone can tell. The pending Patent Eligibility Restoration Act addresses § 101 eligibility, not inventorship, and no bill has been introduced to make machines inventors. Expect study and hearings before any structural change.

Conclusion: It Comes Down to Conception

The question of who owns what the machine creates has no single, tidy answer—but it has a center of gravity, and that center is conception. Courts and patent offices worldwide have settled the narrow point: under current law, AI systems cannot be inventors, and human involvement remains essential. The DABUS cases nailed that down, with the Federal Circuit, the UK Supreme Court, the EPO Legal Board of Appeal, and courts in Australia and beyond converging on the result through varied reasoning.

The live question is the one that decides commercial cases, and it remains genuinely in motion. How much human contribution an AI-assisted invention requires has already swung once in the United States—from the 2024 Pannu-factor approach to the 2025 return to traditional conception—and further change, administrative, judicial, or legislative, is entirely possible. But for a company like Heliotrope, the practical guidance is stable even where the doctrine is not. Document human contributions comprehensively and contemporaneously. Structure workflows so that human conception is real, directed, and visible in the record. Update invention-disclosure processes to capture AI involvement candidly rather than hide it. Get clean assignments from every true inventor, and keep the inventorship roster consistent across every jurisdiction you file in. Weigh trade secrets as a complement to patents, asset by asset. And never let the temptation to flatter a human inventor's role override the duty of candor.

In the end, Dr. Okafor's patent will rise or fall not on what HELIX did, but on what she can truthfully show she conceived. That is the through-line of this entire field: the machine can do the heavy lifting, but the law still asks for a human mind that formed a definite and permanent idea of the complete and operative invention. Build your records, your workflows, and your patents around that sentence, and you will be ready for whatever the framework becomes next.

This article is general information, not legal advice. The law governing AI-assisted inventions is evolving rapidly—the operative USPTO guidance changed as recently as November 2025—and specific situations call for qualified patent counsel. For assistance with patent strategy for AI-related innovations, please contact our intellectual property practice.

Related Articles

Selected Authorities

Constitution, statutes, treaties, and rules: U.S. Const. art. I, § 8, cl. 8; 35 U.S.C. § 100(f) (defining "inventor"); 35 U.S.C. § 101 (eligibility); 35 U.S.C. § 112 (written description and enablement); 35 U.S.C. § 116 (correction in pending applications; joint inventors); 35 U.S.C. § 256 (correction of issued patents); 37 C.F.R. §§ 1.56, 1.324; Defend Trade Secrets Act, 18 U.S.C. § 1836; UK Patents Act 1977, §§ 7, 13; Patents Act 1990 (Australia); European Patent Convention, arts. 60, 81, and Rule 19(1).

Cases and decisions: Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), cert. denied, 143 S. Ct. 1783 (2023); Mohamad v. Palestinian Authority, 566 U.S. 449 (2012); Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998); Burroughs Wellcome Co. v. Barr Labs., Inc., 40 F.3d 1223 (Fed. Cir. 1994); Hybritech Inc. v. Monoclonal Antibodies, Inc., 802 F.2d 1367 (Fed. Cir. 1986); Therasense, Inc. v. Becton, Dickinson & Co., 649 F.3d 1276 (Fed. Cir. 2011) (en banc); Bd. of Trs. of Leland Stanford Junior Univ. v. Roche Molecular Sys., Inc., 563 U.S. 776 (2011); Amgen Inc. v. Sanofi, 598 U.S. 594 (2023); Thaler v. Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49; EPO Legal Board of Appeal, J 8/20 and J 9/20 (Dec. 21, 2021); Commissioner of Patents v. Thaler [2022] FCAFC 62 (Austl.); Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. 2025) (copyright human-authorship requirement).

Agency and executive materials: USPTO, Inventorship Guidance for AI-Assisted Inventions, 89 Fed. Reg. 10,043 (Feb. 13, 2024) (rescinded); USPTO, Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54,636 (Nov. 28, 2025); Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence" (Jan. 23, 2025); USPTO, Inventing AI (Oct. 2020); EPO Guidelines for Examination (AI and computer-implemented inventions).

Secondary sources: Congressional Research Service, Artificial Intelligence and Patent Law, LSB11251 (updated Dec. 2025); WIPO, Technology Trends 2019: Artificial Intelligence (2019); Senate Bipartisan AI Working Group, Driving U.S. Innovation in Artificial Intelligence (2024); contemporary practitioner analyses (2025–2026) of the revised USPTO guidance.