Artificial intelligence in drug discovery: legal status of AI-generated inventions
Alison Leroy
UGGC Avocats, Paris
a.leroy@uggc.com
Artificial intelligence (AI) is quickly becoming a major driver of innovation in the pharmaceutical industry, fundamentally reshaping how drugs are discovered, designed and developed. Over the past decade, the widespread adoption of AI-driven tools and platforms has dramatically accelerated the identification and optimisation of new therapeutic compounds, while reducing development times and costs, and increasing the chances of clinical success.
At the same time, the rise of AI brings significant challenges, especially regarding the ownership, protection, and commercialisation of AI-generated inventions. As AI systems gain the ability to autonomously create patentable innovations, the adequacy of current intellectual property protections is being questioned, leading to a thorough examination of how legal frameworks should evolve to address the realities of the AI era.
AI as a tool for innovation in drug discovery
The evolution of AI in pharmaceutical research
The application of AI in healthcare dates back well before the term was even coined,[1] beginning with early mathematical models that could simulate basic medical reasoning. A pivotal moment occurred in 1943, when Warren McCulloch and Walter Pitts introduced the first mathematical model of the ‘formal neuron’. Their work laid down the foundation for modelling human thought processes and ultimately propelled the development of AI. In the 1950s, this breakthrough ushered sparkling new ideas about machine intelligence. Alan Turing’s landmark paper, ‘Computing Machinery and Intelligence’, posed the fundamental question: ‘Can machines think?’ He introduced the renowned Turing test,[2] which evaluates whether a machine can convincingly imitate human intelligence.
These advances established both the theoretical and practical groundwork for AI, marking the start of a new era in applying AI to drug discovery. They paved the way to the first physical neural networks (SNARC),[3] the first conversational agents (ELIZA), the first artificial neural network designed for pattern recognition and decision-making (Perceptron),[4] and the earliest expert systems in medical AI (MYCIN).[5]
Impact of AI on the innovation process
Despite these substantial advances, AI truly began to show its full potential in the pharmaceutical industry with the arrival of big data, enhanced computing power, and the development of machine learning and deep learning techniques.
Today’s most advanced AI platforms, such as Chemistry42 (Insilico Medicine), Atomwise or DeepMind (Google), leverage enormous amounts of biological and chemical data to predict pharmacokinetics, toxicity and possible drug interactions. These platforms can refine the design of drug candidates, forecast adverse effects, generate pharmacophores and predict protein structures (AlphaFold). AI is also pivotal in identifying new therapeutic targets through gene-disease association analyses, thereby paving the way for precision medicine.
The practical impact of these technological breakthroughs is evident in major industry collaborations, including the partnership between Amgen and Amazon Web Services, as well as Novartis’ development of the Data42 platform. Such initiatives highlight how AI is being integrated across all stages of the drug development cycle: from target identification and clinical trial optimisation to patient monitoring, supply chain management and counterfeit drug detection.
However, human expertise remains essential at every step of the process, whether interpreting AI-generated results, designing experimental protocols, or conducting in vitro and in vivo validations. In this context, AI serves as a powerful complementary tool, enhancing rather than replacing the vital contributions of researchers, chemists, biologists and clinicians.
Legal status of AI-generated inventions
The issue of inventorship for AI-generated inventions
In light of this AI boom, a significant legal question is arising: how should inventions generated or co-developed by AI be legally classified, and to what extent can they benefit from appropriate protection under current intellectual property (IP) law? Patent law, both in Europe and in the United States, now faces the unprecedented challenge of handling ‘invention without a human inventor’, or at least the gradual diminishing of the human contribution in the inventive process.
United States
In the US, the issue was crystallised in several emblematic cases, DABUS and Thaler v Vidal,[6] where the United States Patent and Trademark Office (USPTO) and courts have strictly interpreted the law[7] to uphold the principle that only a human being can be recognised as the inventor of a patentable invention, and confirmed that AI cannot be designated as inventor or co-inventor. According to the USPTO, inventor status required a human contribution that involves creative and intellectual input specific to the conception of the invention.
Applicants are therefore required to identify all human inventors who have made a significant contribution to at least one claim of the patent. Making a false declaration exposes the inventor to criminal sanctions (fine and/or imprisonment),[8] while an inaccurate designation that goes uncorrected may cause the USPTO to reject the application, or even render the patent invalid or unenforceable.
To support applicants, the USPTO published, in February 2024, a guide called Inventorship Guidance for AI-Assisted Inventions. This guide clarifies how to determine inventorship when AI plays a role in the inventive process, and outlines what constitutes a genuine inventive contribution – ie, a contribution that goes beyond simply using, overseeing or validating AI-generated results.
Additionally, the Pannu case law[9] has established a three-part test to determine whether a person can be qualified as a co-inventor in a patent application:
- the co-inventor must contribute significantly to the conception or development of the invention;
- provide a contribution that is not insignificant to the claimed invention; and
- do more than merely explain known concepts or the state of the art.
These criteria are also relevant in the context of AI-generated inventions.
Europe
In Europe, the position of the European Patent Office (EPO) is similar when it comes to refusing to recognise an AI as an inventor. However, there are significant differences compared to the US legal framework, especially regarding the type of oversight and the consequences for incorrect inventor designation.
The EPO’s review of inventor designation is essentially formal and does not involve a thorough investigation into the actual contribution of the designated inventor. As a result, it is possible to designate a ‘front’ inventor without this automatically leading to the rejection of the application. In addition, the means for challenging inventor designation in Europe are limited and often ineffective, unless there is clear evidence of fraud or a harmed human inventor. This situation contrasts sharply with the US system which conducts more rigorous checks and imposes clear penalties for inaccurate or fraudulent designation of the inventor.
Furthermore, the concept of ‘inventor’ has not yet been harmonised across Europe. There are no directives or guidelines specifying the criteria for determining which natural person should be designated as the inventor when AI is involved in the inventive process. Each Member State therefore applies its own standards. Notably, in a DABUS decision dated 11 June 2024, the German Federal Court (Bundesgerichtshof) adopted a more flexible approach than that of the US. It ruled that any human contribution, no matter how minor, is sufficient to justify the designation of an inventor for AI-generated inventions.
The challenges of inventive steps and methods of protection
Beyond the question of inventorship for AI-generated inventions, another significant challenge lies in demonstrating the inventive step, especially when AI can autonomously generate new content or solutions without human intervention and actively participates in the inventive process. Frequently, these AI systems operate based on so-called ‘black box’ deep learning models, which obscure the path leading to invention. This lack of transparency makes it difficult to clearly identify the specific human contribution to the invention, which is essential if patentability requirements are to be met: if the applicant fails to adequately explain how the invention was conceived, the validity of the patent may be called into question.
In this context, the patent application must clearly articulate the invention’s technical contribution, detailing both the nature and extent of human involvement and explaining how AI was utilised during the inventive process. The description must be thorough enough for a person skilled in the art to reproduce the invention. Comprehensive and precise documentation of human choices, interventions and decisions throughout the inventive process is therefore essential to support the justification of inventor status and the validity of the patent.
For AI-generated inventions whose results or methodologies are difficult to fully describe, or where public disclosure could undermine their value, trade secret protection can be a useful alternative or complement to patent protection. This is especially pertinent for certain aspects of AI-generated inventions (such as proprietary algorithms, learning methods or intermediate results) which may not be eligible for patent protection.
Trade secrets offer indefinite protection as long as the information remains confidential and is protected by appropriate security measures.[10] However, this approach has limitations: it offers no protection against independent discovery or reverse engineering and may be less suitable when disclosure is necessary for regulatory or marketing authorisation.
Conclusions
AI is transforming drug discovery, unlocking unparalleled opportunities to accelerate innovation, cut costs and enhance patient outcomes. However, the emergence of AI-generated inventions is disrupting the foundations of IP law, especially in terms of inventorship, patentability and disclosure requirements. While today’s legal frameworks continue to prioritise human creativity and involvement, the growing autonomy of AI may ultimately necessitate legislative updates to ensure that IP protection keeps pace with technological advancement.
For the pharmaceutical industry, the key to securing and maximising the value of AI-generated inventions lies in adopting proactive and flexible protection strategies. This means drafting patent with precision and transparency, strategically leveraging trade secrets, or combining multiple forms of IP protection and thoroughly documenting every aspect of human involvement throughout the inventive process.
Notes
[1] The expression AI was introduced in 1956 at the Dartmouth Conference by John McCarthy, who defined it as the science of ‘making machines intelligent’. Marvin Minsky completed this definition by emphasising the design of programs capable of performing tasks requiring complex mental processes, such as learning and critical reasoning.
[2] Three participants: a human judge, another human being and a machine. The judge, isolated in a room, dialogues in writing with the two interlocutors without seeing them. The machine tries to answer in such a way that the judge cannot easily distinguish the machine from the human being. If the judge cannot make the distinction after five minutes of dialogue, the machine passes the test, demonstrating its ability to communicate in a similar way to a human being.
[3] The Sthocastic Neural Analog Reinforcement Calculator was designed to simulate a rat navigating through a maze: the device tracked the rat’s virtual movements across a light panel, visually demonstrating how the machine gradually learned to find its way out of the maze.
[4] An artificial neural network designed to automatically classify data into two distinct categories. It learns to differentiate between two types of data from a set of examples, and then, once trained, is able to autonomously assign new data to one or other of the categories.
[5] MYCIN relies on a database of clinical rules, enabling it to recommend treatments based on the symptoms presented by patients. It has the ability to provide relevant recommendations in emergency situations, even in the absence of full data on the pathogen, while adjusting dosage according to the patient’s weight.
[6] Thaler v Hirshfeld, United States District Court for the Eastern District of Virginia, decision of 2 September 2021, no 1:20-cv-00903; Thaler v Vidal, United States Court of Appeals for the Federal Circuit, decision of 5 August 2022, no 2021-2347; Decision of the USPTO of 13 August 2020 (Application no 16/524,350).
[7] Title 35 USC, Sect 115
[8] Title 35 USC, Sect. 115(i) – Title 37 CFR s 1.63
[9] Pannu v Iolab Corp, 155 F.3d 1344 (Fed Cir 1998)
[10] TRIPS Agreement, art 39, s 2; Defend Trade Secrets Act; Directive (UE) No 2016/943 dated 8 June 2016, art 2.