AI Drug Discovery Tests the Limits of Patent Law

Artificial intelligence (AI) already helps research teams design novel molecular structures, predict protein interactions, and identify therapeutic candidates, dramatically increasing the speed of drug discovery. The economic implications are substantial: AI drug discovery is projected to generate $13.4 billion by 2035, up from $1.8 billion in 2024.¹

But there’s a catch. Pharmaceutical economics depends on patents, and patents require human inventors who actually conceived the inventions. Courts have held that AI systems cannot be inventors, yet they haven’t clarified how much human contribution suffices when AI does the generative work. Companies may therefore struggle to demonstrate that an individual substantially contributed — leaving breakthroughs unpatentable. The commercial viability of AI drug discovery thus hinges on legal standards that don’t yet exist.

Despite the uncertainty, companies are already using AI to develop drugs — more than 500 U.S. Food and Drug Administration submissions from 2016 to 2023 included AI components.²

Given competitive dynamics, companies can’t wait for clarity. How they structure AI workflows to preserve human involvement, and how they document contribution, could determine which discoveries they can protect.

The Inventorship Problem

The Federal Circuit has long held that conception — “the formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention”³ — is the touchstone of inventorship. For chemical compounds, conception requires both an appreciation of the molecular structure and an understanding of how to make it. Generative models can propose structures and synthetic routes that satisfy preset constraints. When a scientist instructs an AI model to design a kinase inhibitor with specified binding properties and the system produces a novel compound that meets those requirements, who conceived the invention?

Existing case law offers little guidance. Courts have ruled that merely posing a research goal or validating a solution is not inventorship. But modern workflows involve more: Human teams frame problems, select data, set constraint and reward functions, tune models and hyperparameters, interpret rankings against biological context, and decide which candidates to test. Which of these steps amount to a “significant contribution” to the claimed compound remains untested.

Two edge cases illustrate the stakes. First, automation-heavy generation: A team sets broad constraints and presses “go,” then the AI system outputs a structure later validated in vitro. If courts deem the human role too attenuated, no one may qualify as an inventor — placing exclusivity at risk. Second, human-directed iteration: Scientists iteratively refine constraints based on mechanistic hypotheses, reject multiple outputs for biological reasons, and direct additional model runs until a structure with a predicted binding mode and synthesis plan emerges. In this case, documented human choices and rationales may better support conception. The legal line between these scenarios is precisely what early litigation will have to draw.

Proving Human Contribution

Until courts clarify the standard, companies must work backward from what might satisfy it. Establishing human contribution is necessary but not sufficient: Companies must show evidence of that contribution and describe the invention clearly enough for others to reproduce it. The practical questions are simple: What to document, and what to disclose.

What to Document

Companies should document how scientists framed the problem and chose targets; why particular constraints, endpoints, or safety screens were used; and how data was selected, cleaned, or set aside when bias appeared. The record should trace how — and why — prompts and parameters changed over successive runs; explain why some candidates advanced while others were dropped; and capture chemist input on whether a molecule could actually be made and how routes were adjusted. 

This isn’t paperwork for its own sake. A contemporaneous record both establishes human contribution — showing that people did the inventive work of defining, interpreting, and deciding — and builds credibility with examiners, investors, and partners long before any courtroom test.

What to Disclose

Patents involve a trade-off: Patent holders can get exclusive rights if they teach others how to make and use their inventions without excessive trial and error. For a compound, that traditionally meant describing the structure, a workable route to synthesis, and its use. AI complicates this because the path to the structure may pass through a model that functions as a black box. 

The goal isn’t to hand over the platform. It’s to show a reproducible path to the molecule or class — enough detail on inputs, constraints, scientific rationale, and validation steps for a skilled scientist to follow the same approach.

Companies should avoid falling into two common traps. First, reliance on black box explanations. “The model said so” isn’t a teaching. If there is no intelligible path from inputs and constraints to the claimed structure, the disclosure could fail patent law’s enablement test.

Second, over disclosure. Don’t include model or data details that aren’t needed to teach someone how to make and use the compound. Extra specifics only help competitors replicate the approach without improving enablement. The practical test is not “Could a competitor rebuild the platform?” but “Could a skilled person make and use the claimed compound or class based on the description without undue experimentation?”

These two traps create a narrow corridor. Disclose too little, risk invalidation. Describe model architecture and training data in detail, publish a blueprint for competitors. Case law hasn’t defined where companies should land, leaving them to make high-stakes decisions with incomplete guidance.

Workflow Design as Strategy

Without clear legal standards, the best companies can do is build processes that hedge against multiple outcomes. Require scientists to justify hypotheses before generation begins. Create human-in-the-loop checkpoints that record why candidates were rejected and how human judgment informed each iteration. Tie rankings to explicit scientific criteria. If courts ultimately require substantial human contribution, these practices document it. If they focus on enablement, the same records supply disclosure without revealing proprietary methods. The workflows don’t solve the uncertainty — they create defensible positions within it.

The New Economics of Inventorship

Behind the documentation strategies lies a more fundamental question about what AI actually is. Research tools — whether microscopes, genomic sequencers, or computational chemistry software — have historically been treated as instruments. But when AI generates innovative insights, can it still be called merely a tool? And if it is something more, does the current framework for allocating credit and economic returns still make sense?

The answers will influence more than drug patents. They will affect how companies structure research, how investors value intellectual property, and whether industries capture the full promise of AI-accelerated discovery. If patent uncertainty undermines the economic incentive, society may lose the opportunity to accelerate drug development with AI.

The first wave of patent litigation on AI-generated drugs — likely just two or three years away — will establish precedents that could either sustain pharmaceutical innovation or force its fundamental restructuring.

The challenge isn’t to preserve an old definition of inventorship but to decide what kind of innovation society wants to encourage. And that decision can’t wait for perfect clarity — it’s being made right now, in research labs and patent offices, by teams documenting their workflows one compound at a time.

*  *  *

Check out the next article, “Cybersecurity’s New Power Dynamics,” explore our full collection of insights into legal dynamics disrupting modern business strategy, or subscribe to Forces of Law to receive our latest publications.

---

1. ^[1] “AI in Drug Discovery Market,” Roots Analysis (October 2025). ↩

2. ^[2] “2024: The year AI drug discovery and protein structure prediction took center stage—2025 set to amplify growth,” Drug Discovery and Development (November 2024). ↩

3. ^[3] “2138 Pre-AIA 35 U.S.C. 102(g) [R-01.2024],” United States Patent and Trademark Office (October 2024). ↩

This informational piece, which may be considered advertising under the ethical rules of certain jurisdictions, is provided on the understanding that it does not constitute the rendering of legal advice or other professional advice by Goodwin or its lawyers. Prior results do not guarantee similar outcomes.