The Black-Box Dilemma: Reconciling Medical AI Inventions with Sufficiency of Disclosure under Indian Patent Law
This is an Award-winning Essay conferred with a Certificate of Rank-based Merit by LegalWiki as part of LegalWiki's 1st National Article Writing Competition.
The Quid Pro Quo Paradigm and the Issue with Deep Neural Networks
The basic tenet of patent law is that of a “bargain” that takes place between the inventor and the public, represented by the “quid pro quo” in a patent: The inventor is granted a temporary (usually 20 years) legal monopoly in exchange for making a full disclosure of the invention.[1] Section 10(4) of the Indian Patents Act, 1970, states that a complete specification should clearly describe the invention, its mode of operation and the ‘best mode’ of performing it.[2] This requirement ensures that at the end of the patent a PSITA (Person Skilled In The Art) will also be able to duplicate the teaching without undue experimentation.[2]
But this classical paradigm is disrupted by the application of AI across clinical healthcare, from automating diagnostics to exploring genomics and generative drug discovery.[3] The current medical AI innovations are based on deep neural network architectures, which are non-linear mathematical transformations over hidden parameters.[4] These models are trained from training data using stochastic gradient descent, so they are epistemologically opaque in the sense that they acquire statistical representations of the data without explicitly specifying the decision path.[5] Deep learning models make diagnostic predictions based on distributed weight matrices that cannot be explained in the source code or a flowchart.[4] This is unlike deterministic software, which has algorithmic logic that can be written down.[6] As a result, traditional specifications are unable to effectively meet the needs of Section 10(4) which creates tension between the opacity of these machines and legal requirements to disclose as per patent law.
Judicial Interpretation: Examining Motupalli and the CRI Guidelines 2025
Indian Patents Act establishes very strict enablement criteria for computer-implemented innovations. The Madras High Court’s decision in Caleb Suresh Motupalli v. Controller of Patents[7] is a crucial landmark case in the context of this discussion. The court dismissed an appeal against a proposal for an AI-human cognitive integration system, finding that the proposed system was neither sufficiently specific nor did it meet the statutory enablement requirements of Section 10(4)(a).[7] Justice Senthilkumar Ramamoorthy said that specifications should clearly state reliable and reproducible technical procedures to obtain the stated results and not an abstract concept or a black box feature.[7]
Motupalli also helped us understand certain crucial aspects of patentability of Artificial intelligence. The court held that an order-in-review, as may be issued under Section 77(1)(f)[8], is not directly appealable under Section 117-A[9] of Patents Act unless it is issued as a fresh order-in-original under Section 15.[10] As far as the PSITA standards in interdisciplinary domains are concerned, the court noted that PSITA should be considered as a team of experts from multiple fields of technology (AI, software, clinical practitioners) and not as a single technological expert.[11]
This judicial stance comes to the fore in the context of statutory subject-matter restrictions on computer programs under Section 3(k) and mental acts under Section 3(m)[12]. According to Ferid Allani v Union of India, software should have some tangible technical effect or interact with specific hardware.[13] Black box machine learning without any clear technical effect details is not in compliance with Section 3(k) eligibility and Section 10(4) enablement requirement.
In line with this judgement, the Indian Patent Office issued its Guidelines for Examination of Computer Related Inventions (CRIs) in 2025.[14] The Guidelines call for greater scrutiny of deep learning claims, with clear technical specifications of how the data input is converted into specific technical output.[14] This leads to a legal dichotomy for medical AI developers: if they simply disclose the functional outputs, it is insufficient as per law and disclosing millions of dynamic neural weights is extremely unfeasible.
The Threat of Innovation Flight: Trade Secrecy and Public Health Imperatives
If we impose these unadapted enablement requirements it could push Biomedical AI innovations into trade secrecy. As developers are coming to understand that deep neural networks do not satisfy Section 10(4) structural disclosure requirements, they are more likely to leave model architectures, training sets, and hyper parameter configurations as proprietary secrets. The table below clearly shows the different aspects of the patent framework and of the trade secrets framework.
Legal Dimension | Patent Framework | Trade Secrecy Protection |
Statutory Basis | The Patents Act, 1970, Section 10(4) | Common law breach of confidence and contracts |
Disclosure Requirement | Full public disclosure enabling PSITA replication | Complete non-disclosure; operational confidentiality |
AI Compatibility | Challenged by deep neural network opacity | Highly suited for unexplainable neural weights |
Public Health Impact | Promotes scientific verification and safety audits | Prevents external auditing, obscuring bias and risks |
Term of Protection | Statutory grant of 20 years from filing date | Indefinite term, while secrecy is maintained |
The advantage of trade secrecy is that it gives immediate commercial protection, but that is also is a huge public health disadvantage for this case specifically. In the clinical healthcare sector, black-box AI trade secrets could prevent independent auditing of algorithmic bias across various patient populations.[15] The absence of parameters or training procedures being available to the public would make it difficult for clinicians to detect demographic biases or failure modes in automated diagnostic systems.[16]
Stratified Framework: Tiered Disclosure Models and Regulatory Safe Harbours
In order to address the issue of opacity of Deep Learning (DL) under Indian patent law, patent examination should move from a structural enablement approach to a functional and stratified disclosure approach. The requirement of mechanical duplication of hidden neural layers is too strict a requirement and may discourage filings.
A “tiered disclosure” model provides an organised regulatory structure that could meet Section 10(4) requirements. The assessment of enablement for medical AI could be done at three functional levels:
Data Provenance and Curation: The detailed specifications of sources of the training data, pre-processing methods, selection criteria for inclusion and exclusion, and characteristics of the validation data set will need to be disclosed.[17]
Architectural Topology: Complete description of network topology to show how devices (nodes) and wires (links) are arranged in a computer network, including layer types, activation functions, loss landscapes, hyperparameter tuning protocols and optimization algorithms.[18]
Explainable AI (XAI) Integration: Transparency regarding post-hoc attribution techniques (like SHAP values, Integrated Gradients or Saliency maps) which will reveal feature relevance and clinical predictability should also be disclosed to test model robustness.[18]
The Indian Patent Office could also create a regulatory safe harbour and deposit regime similar to the biological deposit provisions of the Budapest Treaty[19], in addition to tiered disclosure mentioned above. Applicants could be given the option to deposit/upload their trained model weights, source code and validation binaries into a secure repository within the Patent Office to be verified during examination. Meanwhile, public specifications would reveal architectural structures, sources of data, and XAI metrics for functional understanding.
Balancing Machine Opacity with Statutory Integrity
A modern approach towards interpreting Section 10(4) in the context of the functional reality of AI is required to successfully tackle the black-box medical AI problem under Indian patent law. As seen in Caleb Suresh Motupalli v Controller of Patents and the 2025 CRI Guidelines, abstract claims that lack any technical implementation cannot be granted any statutory protection under Patent law. But having medical AI innovations as a trade secret is a risk for public health transparency and scientific advancement. Hence, a multi-tiered disclosure regime coupled with the use of Explainable AI metrics, along with mechanisms for model deposit, could possibly help protect medical AI innovations while maintaining the ‘bargain’ that we started our discussion with, under Section 10(4) of Patents Act, 1970.
References
1. Jeffrey A. Lefstin, The Formal Structure of Patent Law and the Limits of Enablement, 23 Berkeley Tech. L.J. 1141, 1141 (2008).
2. Patents Act, No. 39 of 1970, § 10(4), India Code (1970).
3. Yosri A. Fahim et al., Artificial Intelligence in Healthcare and Medicine: Clinical Applications, Therapeutic Advances, and Future Perspectives, 30 European Journal of Medical Research 848 (2025), https://doi.org/10.1186/s40001-025-03196-w.
4. Anichur Rahman et al., Machine Learning and Deep Learning-Based Approach in Smart Healthcare: Recent Advances, Applications, Challenges and Opportunities, 11 AIMS Public Health 58 (2024), https://doi.org/10.3934/publichealth.2024004.
5. Ying Tian, Yating Zhang & Haoyu Zhang, Recent Advances in Stochastic Gradient Descent in Deep Learning, 11 Mathematics 682 (2023).
6. Shigeru Yamada, Software Reliability Modeling: Fundamentals and Applications, vol. 5 (Springer 2014).
7. Caleb Suresh Motupalli v. Controller of Patents, 2025:MHC:293, C.M.A. (PT) No. 2 of 2024 (Madras H.C. Jan. 29, 2025).
8. Patents Act, No. 39 of 1970, § 77(1)(f), India Code (1970).
9. Patents Act, No. 39 of 1970, § 117-A, India Code (1970).
10. Garima Sethi & Rachita Jain, Caleb Suresh Motupalli v. Controller of Patents: Maintainability of Appeal and Patentability Under the Patents Act, 1970 (July 10, 2025), C&CIP.
11. Pragya Saigal, Fueling Innovation: Madras High Court Clarifies Patentability of Technical Methods in Bosch Case, IIPRD (July 30, 2025), https://www.iiprd.com/fueling-innovation-madras-high-court-clarifies-patentability-of-technical-methods-in-bosch-case/.
12. Patents Act, No. 39 of 1970, § 3(k), 3(m), India Code (1970).
13. Ferid Allani v. Union of India, W.P. (C) No. 7 of 2014, 2019 SCC OnLine Del 11867 (Del. H.C. Dec. 12, 2019).
14. Office of the Controller General of Patents, Designs & Trade Marks, Guidelines for Examination of Computer Related Inventions (CRIs) 2025 (2025), https://ipindia.gov.in/storage/uploads/docs-operator/62449276-74f6-438d-92fc-14b2cf3fa204.pdf.
15. Stacey Kusterbeck, Growing Demand for Ethics Expertise to Evaluate AI Tools Before Use in Healthcare, 41 Medical Ethics Advisor 1 (2025).
16. Shubhada R. Kanetkar & Vaishali M. Thorat, Ethical Challenges and Guidelines for AI Deployment in Healthcare, in Internet of Medicine for Smart Healthcare (Anand Kumar, Nitin Vyas, P. S. Rathore, A. Anand & P. Dixit eds., Wiley 2024), https://doi.org/10.1002/9781394272266.ch21.
17. Kelvin K. L. Wong et al., From Trust in Automation to Trust in AI in Healthcare: A 30-Year Longitudinal Review and an Interdisciplinary Framework, 12 Bioengineering 1070 (2025), https://doi.org/10.3390/bioengineering12101070.
18. Jeevana Jyothi Pujari et al., Advances in Explainable Artificial Intelligence Evaluation and Ethics from 2023 to 2025, 6 Discover Artificial Intelligence (2026), https://doi.org/10.1007/s44163-026-01473-z.
19. Budapest Treaty on the International Recognition of the Deposit of Microorganisms for the Purposes of Patent Procedure art. 3, Apr. 28, 1977, 1861 U.N.T.S. 361.
Comments