A Swiss biotech company engages an AI platform provider to accelerate lead compound identification. The platform ingests the company's proprietary screening data. It trains on public compound databases. It generates 50,000 candidate molecules. Medicinal chemists review the output. They select twelve candidates for synthesis. One shows promising in vitro activity against the target. The company files a patent application. Who is the inventor?
The question is not philosophical. It determines whether the patent is valid.
1. How Much Human Contribution Is Enough?
Patent law, globally, requires human inventors. The DABUS cases established this principle across major jurisdictions. The USPTO, EPO, UK IPO, and German Federal Court have all rejected attempts to name AI systems as inventors.1Thaler v Vidal, 43 F.4th 1207 (Fed Cir 2022); EPO J 8/20 (DABUS); Thaler v Comptroller General [2023] UKSC 49; BGH, X ZB 5/22. An inventor must be a natural person. This principle is not under serious dispute. The question that remains unanswered is different: how much human contribution is enough?
In February 2024, the USPTO issued guidance on AI-assisted inventions, requiring humans to make "significant contributions" to conception.2USPTO, Inventorship Guidance for AI-Assisted Inventions (13 February 2024) 89 FR 10043, rescinded and replaced by Revised Inventorship Guidance (28 November 2025). That February 2024 guidance applied the Pannu factors, derived from Pannu v. Iolab Corp., which require that each named inventor contribute in some significant manner to the conception of the invention as claimed.3Pannu v Iolab Corp, 155 F.3d 1344 (Fed Cir 1998). Under this framework, a person who simply follows instructions or exercises ordinary skill is not a co-inventor. The November 2025 revision rescinded this guidance and replaced it with a framework treating AI systems as "instruments used by human inventors, analogous to laboratory equipment." The revised guidance withdrew explicit reliance on Pannu for AI-assisted invention analysis but did not repudiate the underlying principle that conception requires intellectual contribution beyond mere selection or supervision. The human must still conceive the claimed invention; the question is how much AI assistance undermines that conception.
The principle is clear: AI cannot be an inventor. Everything that follows from that principle is not.
The EPO's April 2025 examination guidelines reinforce similar principles for Europe. AI and machine learning inventions must demonstrate "technical character" and solve a "technical problem."4EPO Guidelines for Examination (April 2025) G-II, 3.3.1. The underlying treaty provisions are clear: Art. 60(1) EPC provides that the right to a European patent shall belong to the inventor or his successor in title, while Art. 81 EPC requires that the European patent application designate the inventor.5Art. 60(1), 81 Convention on the Grant of European Patents (EPC). The EPO's interpretation that "inventor" means a natural person derives from these provisions read in light of the EPC's drafting history and subsequent practice. Where the technical effect depends on particular characteristics of training data, those characteristics must be disclosed, though the EPO clarifies that "in general, there is no need to disclose the specific training dataset itself." The challenge lies in identifying which characteristics matter when the model's decision-making process is opaque. Notably, the German Federal Court adopted a different framing in its June 2024 DABUS decision: while confirming that only natural persons can be inventors, the BGH held that a human can be designated as inventor even when AI was used, provided some human involvement in the inventive process can be identified.6BGH, X ZB 5/22 (11 June 2024), confirming natural person requirement while allowing human designation where AI was used. How much human involvement suffices, and what forms of involvement qualify, remains contested across jurisdictions.
Swiss patent law under Art. 3 and 5 of the Bundesgesetz über die Erfindungspatente (PatG), as interpreted, similarly requires that an inventor be a natural person who made an inventive contribution; the natural-person requirement is read into these provisions rather than stated in terms.7Art. 3, 5 PatG (SR 232.14); BVGer B-2532/2024 (26 June 2025) (Swiss DABUS). The Swiss Federal Institute of Intellectual Property (IGE) has not issued specific guidance on AI-assisted inventions, but applies the same natural person requirement as the EPO. Switzerland has its own DABUS authority. In B-2532/2024 (26 June 2025), the Federal Administrative Court confirmed that only a natural person may be designated as inventor, yet took a permissive view of human contribution, holding that a person who supplies and trains the data, recognises the AI output as a patentable invention, and pursues protection qualifies as the inventor; it ordered examination to resume with the human applicant named. The decision aligns Switzerland with the BGH's reasoning and suggests the Swiss threshold for sufficient human contribution may be lower than the broader inventorship debate implies.
Consider the practical implications. If a pharmaceutical company's scientists set parameters, review outputs, and select candidates for further development, but the AI performed the generative work of designing the molecular structures, inventorship analysis becomes fact-intensive and uncertain. If the scientists contributed to conception of the specific compound claims, they may be proper inventors. If they merely supervised an AI that autonomously generated the claimed structures, the patent may lack a valid inventor. The consequences of inventorship defects vary by jurisdiction; some systems permit correction under specified conditions, others treat misdesignation as grounds for invalidity or unenforceability. In the United States, incorrect inventorship can render a patent unenforceable if deceptive intent is established; in Europe, the focus shifts to entitlement disputes. The jurisdictional variation creates additional uncertainty for multinational patent portfolios.
2. Who Owns the Trained Model?
AI platforms do not generate drug candidates from nothing. They learn from data. Public databases. Proprietary screening results. Published literature. Licensed compound libraries. When a pharmaceutical company provides its proprietary data to train or fine-tune an AI model, questions multiply. Who owns the trained model? Who owns improvements derived from that training? If the AI platform uses insights from Company A's data to improve performance for Company B, has something been transferred that the contract should have addressed?
The typical collaboration structure separates three categories of intellectual property: background IP that each party brings to the collaboration, foreground IP created during the collaboration, and improvements to background IP. This framework, developed for traditional R&D partnerships, strains when applied to AI-driven discovery. The AI platform's algorithms constitute background IP. The pharmaceutical company's screening data constitutes background IP. But what is the trained model? It incorporates both. It is neither purely one party's contribution nor clearly joint work.
The trained model incorporates both parties' contributions. Standard IP allocation frameworks do not address this.
If the pharmaceutical company's data improves the AI platform's general capabilities, not just its performance on the specific project, the platform provider captures value that extends beyond the collaboration. Whether this constitutes misappropriation, a breach of confidentiality, or simply the natural consequence of machine learning depends entirely on what the agreement says. Many agreements drafted before 2023 say nothing about it.
Data provenance creates additional complications. If an AI model trained on multi-source datasets generates a drug candidate, and one data source later claims joint IP rights, the resulting patent may face ownership disputes that the collaboration agreement never anticipated. The risk is more than theoretical: commentators increasingly anticipate that data providers will assert joint-IP or ownership claims in compounds discovered using AI models trained on their data. Whether data contribution creates an ownership interest in AI-generated outputs, or merely a licensing claim, depends on contractual language that many early agreements failed to address. Whether existing data licensing language adequately separates compensation rights from IP claims depends on distinctions these agreements were rarely drafted to address.
These data ownership complexities do not exist in isolation; they interact with the broader collaboration structure to determine how IP risks concentrate and who bears them when challenges arise. The choice of collaboration model determines not only who owns resulting compounds, but who bears the inventorship risk and who controls the evidence necessary to defend patent validity.
3. How Do Collaboration Agreements Address AI-Generated IP?
Deal Structure Models
AI drug discovery collaborations typically follow one of several models, though the R&D partnership frameworks examined in Insight 01 predate the complications AI introduces. The pharmaceutical company may license access to the AI platform, retaining all IP rights in compounds discovered. The AI company may retain platform ownership while granting exclusive licenses to discovered compounds. The parties may co-own resulting IP with defined commercialization rights. Major deals illustrate the range: Insilico Medicine's collaboration with Sanofi, potentially worth up to USD 1.2 billion, grants Sanofi rights to AI-discovered compounds while Insilico retains its platform; Exscientia, which combined with Recursion Pharmaceuticals in November 2024, structured its arrangements so that the AI company retained platform ownership while pharmaceutical partners secured exclusive rights to specific compounds under defined milestone and royalty terms. Each structure creates different risk allocations, and different vulnerabilities to the inventorship problem.
IP Allocation and Inventorship Risk
If the pharmaceutical company retains all compound IP, it bears the full risk of inventorship challenges. If the scientists named as inventors contributed insufficiently to conception, the patents fail. The AI company, having licensed its platform, has no stake in patent validity. If the AI company retains or co-owns compound IP, it must identify human inventors among its own personnel: the engineers who designed the algorithms, the scientists who curated training data. Whether these contributions qualify as inventive contributions to the specific compound claims is uncertain.
Contractual Risk Management
Contractual approaches to this uncertainty are emerging. Some agreements require the AI provider to warrant that its personnel made sufficient inventive contributions to support valid inventorship claims. Others allocate risk through indemnification provisions: if patents are invalidated for inventorship defects, specified parties bear the resulting losses. Still others require documentation of human-AI collaboration workflows, creating contemporaneous records intended to support inventorship determinations if later challenged.
The questions multiply across each contractual dimension. If the collaboration agreement assigns compound IP to the pharmaceutical company, whether it also assigns the inventorship risk, and what representations the AI provider makes about the patentability of AI-generated outputs, determines where exposure concentrates when patents are challenged. If the AI provider's engineers are named as co-inventors, what ongoing obligations they have and whether they can be compelled to cooperate in patent prosecution and enforcement shapes the practical enforceability of rights that depend on their participation. If inventorship is later challenged and the patent invalidated, which party bears the loss of exclusivity (the party that filed the patent, the party that funded the research, or the party that provided the AI platform) may not have been contemplated when the agreement was drafted, leaving allocation to litigation. And whether confidentiality provisions adequately address model training, including whether insights derived from one collaboration can be used to improve platform performance for competitors, may determine competitive value that outlasts any specific compound.
4. What Happens to the Model When the Collaboration Ends?
When pharmaceutical Company B engages AI platform provider A to discover compounds for Indication X, Company B may insist that the model trained on its data be segregated from A's general platform. This "sequestration" requirement prevents A from using learnings derived from B's proprietary data to benefit other clients, including B's competitors. The sequestered model exists on dedicated infrastructure. It incorporates B's data. It generates outputs specific to B's program.
What happens when the collaboration ends? If Company B developed a compound using the sequestered model and that compound advances to clinical trials, B needs continued access to the model, or at least to the compound-specific outputs, for regulatory filings, manufacturing optimization, and potential follow-on discovery. If Company A retains exclusive control of the model and the collaboration terminates acrimoniously, B's regulatory and commercial program may depend on infrastructure it does not control.
The question of model fate arises at termination. If the agreement requires destruction of the sequestered model, institutional knowledge disappears: the optimizations, the failed approaches, the compound-specific learnings vanish with certified deletion. If the agreement permits transfer of the model to Company B, the AI provider's background algorithms may be exposed, raising questions about what exactly transfers and what license rights accompany it. If the agreement requires maintained access that survives termination, the parties remain in uncomfortable dependency, a terminated relationship that continues through infrastructure neither fully controls. Each approach creates different risks. None eliminates them.
Insolvency adds another dimension. If the AI platform provider enters bankruptcy, the pharmaceutical company's access to sequestered models and trained systems may be at risk. Standard software licensing protections (escrow arrangements, technology access agreements, pactum de non cedendo provisions restricting assignment) require adaptation for AI-specific assets. Under Swiss law, Art. 211a of the Schuldbetreibungs- und Konkursgesetz governs continuing-obligation contracts in the licensor's bankruptcy: rather than a licence automatically surviving, the bankruptcy administration may elect whether to continue performance, and the treatment of licence agreements specifically remains unsettled.8Art. 211a Bundesgesetz über Schuldbetreibung und Konkurs (SchKG) vom 11. April 1889 (SR 281.1). Even where performance continues, its practical value depends on whether the pharmaceutical company can actually operate the licensed technology without the licensor's ongoing support. The model is not simply source code that can be deposited in escrow. It includes trained weights, hyperparameters, and potentially the training data itself. Replicating the model from escrow materials may be technically infeasible without the platform provider's expertise, expertise that bankruptcy may have dispersed.
5. Strategic Considerations
The legal uncertainty is real. Patent offices have clarified that AI cannot be an inventor. They have not clarified the minimum human contribution required when AI performs substantial generative work. Courts have not yet addressed AI-assisted drug discovery patents in adversarial litigation. The first wave of such litigation, likely within two to three years, will establish precedents that do not yet exist.
Does Patent Protection Remain the Right Strategy?
Swiss pharmaceutical and biotech companies entering AI collaborations face decisions that neither the November 2025 USPTO revised inventorship guidance nor the EPO's April 2025 examination guidelines resolve. Where patent protection faces inventorship uncertainty, alternative protection strategies warrant consideration. Trade secret protection under the Art. 6 Bundesgesetz gegen den unlauteren Wettbewerb (UWG) and the EU Trade Secrets Directive does not depend on inventorship; the information need only be secret, have commercial value, and be subject to reasonable protective measures.9Art. 6 Bundesgesetz gegen den unlauteren Wettbewerb (UWG) vom 19. Dezember 1986 (SR 241); Directive (EU) 2016/943 on trade secrets [2016] OJ L157/1. For drug candidates, the tension between trade secret duration and regulatory disclosure timing creates a protection gap that patent protection would ordinarily fill. Data exclusivity under Art. 10(1) of Directive 2001/83/EC and corresponding Swiss provisions provides regulatory protection independent of patent status: eight years of data exclusivity plus two years of market protection in the EU, ten years total data protection in Switzerland.10Art. 10(1) Directive 2001/83/EC; Art. 11a–11b Heilmittelgesetz (HMG) vom 15. Dezember 2000 (SR 812.21). Whether regulatory exclusivity adequately substitutes for patent protection when inventorship claims are uncertain depends on the compound's commercial timeline and competitive landscape, an assessment that interacts with the inventorship analysis rather than replacing it.
Can Workflow Documentation Support Inventorship Claims?
Whether documentation requirements exist for human-AI collaboration workflows, and whether they are sufficient to support inventorship claims if later challenged, shapes defensibility that cannot be reconstructed after the fact. The practical analysis begins with whether the collaboration agreement clearly defines ownership of the trained model, not just the resulting compounds, a distinction that traditional pharmaceutical agreements rarely needed to draw.
Where Does Risk Concentrate in the Agreement?
The AI provider's warranties about patentability, its indemnification exposure if those warranties prove incorrect, and the ongoing obligations that attach to co-inventorship all converge in the risk allocation provisions of the collaboration agreement. These provisions interact in ways that standard technology licensing frameworks do not anticipate: the patentability warranty depends on the inventorship analysis, which depends on the collaboration workflow, which depends on the platform architecture that the AI provider controls. The resulting interdependencies mean that risk allocation in AI drug discovery collaborations cannot be assessed dimension by dimension; each provision shapes the others.
Why Do Template Answers Fail Here?
These questions have no template answers. Each collaboration presents unique combinations of platform architecture, data contributions, human involvement, and commercial objectives. IP allocation, risk distribution, and termination consequences arise in a legal environment where the rules are still being written. What worked for traditional R&D collaborations does not work here. The involvement of AI in the generative process has introduced questions that existing contractual frameworks were not designed to address.