The Patent Disclosure Trap: Will AI Undermine Biopharmaceutical Innovation?

Artificial intelligence could help deliver the next generation of lifesaving medicines. It could also weaken the economic foundation that makes discovering those medicines possible. Both propositions deserve serious attention, and neither is mutually exclusive. Celebrating faster discovery tells only part of the story if the same technology allows competitors to appropriate the commercial opportunity created by pioneer research or obtain patent claims that obstruct the innovator’s next steps.

The concern is an emerging class of super-fast followers, defined as competitors positioned to mine newly published patent applications and scientific literature, identify opportunities around disclosed inventions, and rapidly pursue related candidates and patent filings. The most troubling scenario for innovators is a competitor using AI to generate proposed variants and hypothetical examples, filing before doing the work necessary to establish that it has actually contributed a patentable invention or conceived anything. The speed of the filing can then become a business weapon against the company whose research made the opportunity visible after engaging in pioneering research and development.

This is the issue we will explore during The AI Arms Race: How AI is Impacting R&D and Enabling Super-Fast Followers at IPWatchdog’s 2026 Life Sciences Masters™ program on November 9. Our panel pre-call conversation exposed substantial disagreement about where legitimate competition ends, and a systemic incentive problem begins. My concern is straightforward: if we want more biopharmaceutical innovation, we must preserve a compelling reason to finance the original, expensive, uncertain research and development, which will typically cost billions of dollars before any drug ever reaches the market.

The Cost of Creating the Opportunity

A pioneer does more than disclose a molecule or sequence. It determines which biological questions are worth pursuing, explores approaches that fail, develops methods, and generates knowledge that narrows a highly unpredictable and uncertain field. A successful disclosure can represent the visible result of years of work, including failures competitors will never need to repeat. The knowledge is valuable precisely because the pioneer has reduced uncertainty for everyone who follows, and today that work is ultimately rewarded with patent protection and exclusive rights for a limited time.

While it is certainly true that a patent disclosure is intended to spread technical knowledge, the quid pro quo of the patent bargain depends on meaningful protection for the inventor. A system that extracts useful disclosures while leaving the disclosing party unable to justify its investment will eventually produce fewer of those disclosures, particularly in a marketplace where the desired innovation costs billions of dollars to achieve. The danger is particularly acute in the biopharma industry when early research must be followed by substantial spending on development, manufacturing, and regulatory approval before a medicine generates any revenue.

The patent system does not promise freedom from competition, and a patent’s legal term is not a guaranteed period of commercial exclusivity. Nevertheless, investors must make assumptions about the time and opportunity available to recover capital. If AI materially shortens that opportunity, those assumptions will necessarily change. And while a discovery tool can improve scientific productivity, it can also simultaneously weaken the business case for research programs in the first place.

From Fast Followers to Super-Fast Followers

Consider a hypothetical sequence. A company publishes an application describing a promising research direction. A competitor feeds that disclosure into an AI-enabled workflow, identifies related structures or sequences, and generates proposed modifications. Within days, it files an application directed to selected variants, supported largely by predicted performance and proposed experiments. The pioneer may still be conducting the work needed to determine which variants actually matter.

The point is not that the competitor has developed an approvable medicine in days; identifying a candidate, filing an application, conducting clinical trials, and reaching patients are different milestones. The concern is that the competitor may secure an early filing position around future development opportunities before committing comparable resources to establish their value, and without having any actual data or experiments to support the AI generated hypotheticals, some of which may be hallucinations, others that may be wanting for human conception.

Long before modern generative AI, competitors studied patent disclosures and designed around claims. The difference is the potential scale, compression of time, and lack of genuine human conception. A process requiring sustained scientific and legal effort may become much cheaper to repeat across many disclosures, but the patent framework developed without today’s capacity to industrialize that process largely overnight. This suggests that the patent incentive structure should be examined against the resulting competitive dynamics, rather than assuming a familiar category of conduct must have familiar consequences.

What the China Reporting Establishes

An April 10 Wall Street Journal article by Peter Loftus, “Big Pharma Is Turning to China for the Newest Drug Ideas,” provides context. It reports that a Chinese microblog analyzed a Novartis paper and related patent filings within four days. The article also describes GluBio Therapeutics, with operations in Shanghai and San Diego, improving its own candidates after studying Western companies’ patents. GluBio’s chief executive said those lessons “helped us further improve our compounds.”

The reporting does not establish that GluBio used AI, filed unsupported applications, or lacked its own inventive contribution. Its account describes ongoing research and improvements over months. Nevertheless, the article illustrates how quickly public information can become useful to competitors. It is not proof of the AI-enabled filing scenario, and as an innovator themselves one could easily conclude that there is nothing wrong with what GluBio has done, or that this is precisely the purpose of patent disclosures in the first place. But the speed at which Novartis’ paper and related filings were dissected with actionable pathways presents an opportunity to ask broader questions.

While an innovator learning, no matter how fast, from competitors seems entirely within the designed patent framework, what if this happens with a company or individuals who are not innovators who are just sitting at a computer terminal and using AI tools to identify potential gaps and then without real human conception rely on AI generated hypotheticals alone? Or what if there is human conception? Regardless of whether the entity decoding years of painstaking R&D is an innovator, introduces human conception, and/or relies exclusively on AI generated hypotheticals, from the perspective of the original innovator, a competitor can exist in a few days, and that speed is not something that previously has been possible and was never envisioned by the current patent bargain.

Prophetic Examples and the Demand for Substance

Prophetic examples are not inherently illegitimate. Patent law does not impose a requirement for an innovator to conduct an experiment before filing, and an application can be enabling without working examples. See MPEP 2164.02. A predicted embodiment may rest on sound science. Conversely, a plausible-looking AI output may be entirely unsupported, or it could be completely fabricated. As long as there is human conception, the relevant inquiry concerns what the disclosure actually teaches and supports, not whether a computer assisted its preparation.

The MPEP explains that the claims should be drafted in a manner that assists readers in differentiating between actual working examples and prophetic examples. In other words, fabricated experimental results are not supposed to be presented as actual claimed achievements. That distinction must survive the ease with which software can generate convincing technical prose and may need to become more explicit through appropriate U.S. Patent and Trademark Office (USPTO) guidance.

Meanwhile, written description and enablement remain central safeguards, but can patent examiners actually identify fundamental and foundational problems when text drafted by AI can and does sound so compelling, even when it is entirely fabricated? Section 112 requires substantive support for the claimed invention and examiners should evaluate the technical teaching, the breadth sought, and the experimentation required to practice the invention. The danger is allowing the appearance of scientific completeness to substitute for the substance that the law requires, but, practically, how will that work? Examiners already struggle to do the work in the time that is allotted. Given that 101 and 112 issues are only challengeable in post grant review (PGR) proceedings, will innovators need to become more vigilant than ever and resort to PGR to help the USPTO police inadequate, unsupported AI disclosures?

The Pressure on Pioneer Claim Scope

When the United States Supreme Court issued its decision in Amgen, Inc. v. Sanofi et al, the Court determined that the 26 antibody examples and detailed instructions for generating additional antibodies within the genus—which covered some 400 pages and included a CD Rom of the x-ray crystallography coordinates of Amgen’s “anchor” (or lead) antibodies 21B12 (Repatha) and 31H4—was insufficient to satisfy the enablement requirement. Unfortunately, the Supreme Court did not explain what level of detail would have been sufficient, nor did the Court provide direction to the industry on what more is expected to satisfy the enablement requirement. Still, the decision has become understood to stand for the proposition that if you are claiming broadly then you must have examples that cover the entire breadth of the claims.

Ironically, the Supreme Court’s ruling in Amgen may create a practical squeeze. A pioneer seeking broad coverage may need more supporting work—and certainly more across the entire breadth of the innovation—before it can reasonably claim commercially relevant variants. Meanwhile, its disclosure can help a competitor identify specific opportunities to pursue very quickly. The policy question becomes whether innovators can realistically protect enough of the value they create when the work required to support coverage proceeds more slowly than a competitor’s ability to generate proposed hypothetical alternatives.

The patent system will need to be on guard for this disclosure trap. One cannot reasonably demand rigorous support from the pioneer while accepting speculation from the follower merely because it is focused on a subset of a broad innovation. Consistent examination will be essential. Yet even flawless examination would not settle the separate investment question. A valid, independently invented alternative can still arrive fast enough to reduce the expected return on the original research. Thus, patent quality and the economics of research incentives are related problems, not interchangeable ones.

Preserving the Reason to Invest

A handful of troubling applications that fast-follow an innovator with a purely speculative disclosure that may, or may not, actually be the byproduct of human conception should support a different response from a widespread strategy that systematically and quickly cannibalizes pioneer research to block consequential innovation pathways. For that reason, the immediate institutional priority for the USPTO must be reliable examination of claimed inventions, including clear treatment of predictions, genuine scrutiny of disclosure support, and rigorous application of the human inventorship requirement. Meanwhile, policymakers and agency leadership need to monitor real-world filing patterns, development timelines, investment decisions, and evidence about ongoing innovation levels by pioneers.

The system can cope with several questionable filings, or patents, but the systematic AI-gaming of patent acquisition cannot be allowed to result in fewer medicines. If a structural incentive problem emerges, the answer should be designed around that evidence, and always through the lens of maximizing the continued research, development, and commercialization of new medicines. To do this we must be willing to examine whether existing protection adequately rewards the research that creates new therapeutic opportunities, or whether the system has tilted to reward AI-enabled fast followers at the expense of the capital intense hard work historically done by innovators.

The measure of success must be the research we induce, not the volume of patent applications generated. A system that makes it easier to stake speculative claims around someone else’s discovery will produce more patent documents while undermining the investment needed to produce more medicines. That would be a loss for the system, and society, and such a result would need to be confronted before it becomes an established business model.

AI should help us obtain more of the difficult, risky research that opens new possibilities for patients. It should not make that work an increasingly unattractive investment while rewarding the ability to surround its results with rapidly generated claims. The essential question is whether the pioneer retains a commercially meaningful reason to begin. Without that incentive, there will be fewer breakthroughs for anyone to follow, and in that world society loses.

Image Source: Deposit Photos
Author: ariteguhas@gmail.com
Image ID: 657443750