AI Is Inventing Molecules: Why U.S. Patent Law Must Allow Owners of Inventions to File Patent Applications Without Naming Inventors

Artificial intelligence is generating novel molecular structures, identifying new therapeutic uses for approved drugs, and redesigning proteins to perform beyond their natural limits. In a new article in IDEA, the authors explain that U.S. patent law has not kept pace—and that the gap between doctrine and technological reality is creating a category of patents that may be invalid from the day they issue.

Under current law, represented by Thaler v Vidal, AI cannot legally be listed as an inventor on a U.S. patent application or patent. What to do, then, when AI contributes significantly to a claimed invention? The USPTO’s 2025 Revised Inventorship Guidance treats AI as a sophisticated laboratory tool and instructs that, because AI cannot legally be a joint inventor, there is no joint-inventorship question to analyze. The USPTO’s framework, however, collapses two distinct inquiries: who conceived the invention, and who may lawfully be named inventor. It leaves the harder question unaddressed: what happens when the human’s conception is incomplete, or absent altogether?

AI Is Already Inventing

The article describes four real-world examples in which AI played a key role in developing new molecules and methods for treating certain diseases or conditions.

First, a research team asked AI to scan the full universe of FDA-approved drugs—roughly 24,000 products—to identify the best candidate for treating idiopathic multicentric Castleman’s disease (iMCD). The AI identified adalimumab as the best candidate. The research team then administered adalimumab to a treatment-refractory iMCD patient, and the patient entered long-term remission. Who (or what) invented this new therapeutic use of adalimumab?

Second, human scientists asked AI to generate novel antibiotic compounds with activity against Staphylococcus aureus. The AI first generated 30 million candidates, and after filtering and selection, produced a list of 22 candidates. The humans synthesized and tested the 22 candidate compounds, identifying one—DN1—with potent antibiotic activity. Who (or what) invented DN1?

Third, humans used AI to identify a therapeutic target for treating pulmonary fibrosis and a therapeutic agent to hit that target. The AI identified a kinase known as TNIK, and a novel inhibitor of that kinase, rentosertib. Rentosertib entered Phase II clinical trials for idiopathic pulmonary fibrosis less than 18 months later. Who (or what) invented rentosertib and its use?

Fourth, AI generated, synthesized, and tested various chitinase enzyme sequences, ultimately producing an improved chitinase with significantly better thermostability compared with natural chitinase. Who (or what) invented the new chitinase protein and its uses?

These are not hypothetical examples; they are scientific studies published in peer-reviewed literature. Applying the USPTO’s current framework for AI-assisted inventions such as these creates a structural trap.

The Structural Trap

The authors survey existing human inventorship case law, addressing two categories of inventorship disputes: (i) when conception is non-existent or incomplete; and (ii) when conception is complete and additional contributions are insignificant. As the case law makes clear, merely identifying a problem is not a complete conception (e.g., Ex parte Smernoff; University of Colorado Foundation, Inc. v. American Cyanamid Co.), and conception must be more than a research plan—it must extend to the specific claimed subject matter (e.g., Board of Education ex rel. Board of Trustees of Florida State University v. American BioScience, Inc.). The case law also shows that simply confirming the activity of a molecule that someone else conceived does not confer inventorship to the tester (e.g., MacMillan v. Moffett; Burroughs Wellcome Co. v. Barr Laboratories, Inc.; Vanderbilt University v. ICOS Corp.). However, identifying the specific active structure where no prior conception of the molecule existed (e.g., Bergendahl v. Sasai), or supplying a specific synthesis route for the molecule (Falana v. Kent State University) are not insignificant contributions and may confer inventorship.

Applying this framework to modern-day AI-assisted inventions creates a trap. Where AI merely verifies, the laboratory-tool analogy holds. Where AI supplies the structure or method that completes a partial conception or even no conception at all, it does not. And under Pannu v. Iolab, whether that contribution is “not insignificant” is a question of fact. 35 U.S.C. § 256(b) forgives inventorship errors only when they can be corrected. Recently, Fortress Iron, LP v. Digger Specialties, Inc. confirmed that an uncorrected inventorship error for failure to add a missing coinventor invalidated the patent. A missing human inventor can at least in principle be located and added; AI is categorically ineligible. The resulting uncertainty will push innovators – when commercially feasible – toward trade secrecy; increase risk and dampen investment in AI-driven drug development; and cede ground to jurisdictions that adopt more flexible approaches.

A Targeted Legislative Fix

Rather than litigate whether a machine is a legal person, the article proposes three amendments that build on a distinction already embedded in 35 U.S.C. §§ 100(f)–(h). The law already distinguishes between the narrow term “individual” (which refers to a human being), and the broader term “person” (which includes both humans and corporate entities). First, replace “individual” with “inventive entity” in 35 U.S.C. §§ 100(f), (g) and 135(a)(1). Second, repeal § 115 (requiring identification of inventors and inventor oaths or declarations) and remove the oath requirement from § 116(a). Third, amend § 118 to permit any legal or equitable owner of an invention to file a patent application, not only an assignee taking title from a human inventor.

This proposed remedy preserves the quid pro quo disclosure bargain at the heart of the patent system while eliminating the uncorrectable-inventorship problem: validity would turn on ownership and disclosure rather than on naming every conceptual contributor. There should be no constitutional bar to the proposed legislative solution. The Federal Circuit’s holding in Thaler v. Vidal that an “inventor” must be a human being rested on statutory interpretation, not constitutional mandate. Similarly, the D.C. Circuit’s decision in Thaler v. Perlmutter affirming that an “author” of a copyrighted work must be a human also rested on statutory grounds and pointedly left the policy question to Congress. The article also considers, but rejects, an alternative amendment to 35 U.S.C. § 282(b)(3) (addressing presumption of validity and defenses) that would simply immunize AI-related inventorship defects. Such an amendment would invite human applicants to overstate – or even misrepresent – their own contributions and undermine a transparent patent system.

Patent law has adapted to transformative technologies before. The authors urge Congress to address this one directly, rather than continuing to treat AI as though it were simply another laboratory tool.

This summary shares insights published in 67 IDEA: The Law Review of the Franklin Pierce Center for Intellectual Property, 1 (Summer 2026).