As machine learning systems increasingly transform healthcare, patent law faces a fundamental challenge: how should it protect inventions whose inner workings even their creators cannot fully explain? Dr. Gunjan Chawla Arora and Nidhi Krishna examine how India’s patent framework grapples with the rise of “black-box” medical AI, and whether existing disclosure requirements are compatible with the realities of modern machine learning. Dr. Arora is an Assistant Professor of Law & Head, Centre for Intellectual Property Rights, Institute of Law, Nirma University, and has over a decade of experience as an academician in Intellectual Property Rights law. Nidhi is a fourth-year B.Com LL.B. (Hons.), is a student at the Institute of Law, Nirma University, with a keen interest in Intellectual Property Law, Competition Law, and Arbitration.
Black-box Medicine and Indian Patent Law: Why India’s Disclosure Framework Struggles with Black-box AI
By Dr. Gunjan Chawla Arora and Nidhi Krishna
Artificial Intelligence (AI) based clinical predictive models are rapidly being utilized in the healthcare sector. Some of them rely on modern machine learning techniques. Machine learning (ML) systems work by identifying patterns in data and applying those patterns to new situations.
Typically, ML involves three stages:
- Human programmers design the AI model.
- They train the model using algorithms and large datasets.
- The trained model is applied to new data to generate outputs.
These systems are usually trained, and not fully programmed. They generate outputs through internal decision-making processes that develop on their own during training, based on large amounts of data. Modern deep neural networks are non-deterministic, and their internal logic is opaque. This means an identical code, architecture, and training data produce different outputs across trial runs due to random weight initialization, order of the data, random initialization, or random regularization. The name “Black-box AI” emerges for this system as the developers themselves cannot ascertain how the AI reached a certain conclusion, and their lack of transparency raises serious concerns for patent law.
This post examines why India’s current patent framework struggles to accommodate Black-box medicine. It argues that while India’s cautious stance protects transparency and enablement, it may also unintentionally discourage innovation in high-risk, high-reward medical AI.
Disclosure and Enablement: A TRIPS-Grounded Problem with India-Specific Dimensions
The Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS) under Article 29 requires applicants to clearly and completely disclose their invention so a Person Having Ordinary Skill In The Art (PHOSITA) can reproduce it. The best-known method of carrying it out may also be required at the filing or priority date.
The disclosure and enablement challenges that Black-box AI faces are not unique to India. Regulations for patenting in India, the U.S., the EU, and China, for example, satisfy the TRIPS Article 29 baseline. However, in India, failure of the “best mode obligation” under Section 10(4)(b) remains a ground for revocation u/s. 64(1)(h). The US amended 35 USC §282 has virtually removed any consequences for not disclosing the “best mode.” China has no explicit best mode requirement, though Rule 17(5) of its Implementing Regulations (as revised January 2024) requires disclosure of “the optimally selected embodiments contemplated by the applicant,” while the EP has no such requirement at all.
India’s Patent Framework and Core Challenges for Black-box AI
Black-box medical AI in India would be regulated under the Patents Act, 1970, read alongside the Computer Related Inventions (CRI) Guidelines. Section 3 of the Patents Act, sets out subject-matter exclusions for patenting, which poses a challenge for Black-box medicine. Section 3(k) deals with the exclusion from patentability of computer programme per se, algorithms, mathematical and business methods.
In Ferid Allani v. Union of India and Ors the Court held that patentability for AI inventions is based on whether they deliver a genuine technical effect or technical contribution. When a CRI results in a technical effect that improves the system’s functioning and efficacy, or provides a technical solution to a technical problem, it would surmount the exclusion. The CRI Guidelines too, provide for a broad scope of what could possibly qualify as a technical effect.
However, meeting Section 3(k) requirements is challenging. The inability to scrutinize the inner workings of black-box algorithms makes it highly uncertain whether a true technical effect has actually occurred. Therefore, the risk of rejection due to narrow claim scope remains high.
Additionally, a significant risk of rejection persists under Section 10, where the reproducibility and best-mode disclosure requirements impose obligations that black-box AI systems structurally struggle to satisfy.
Section 10 of the Patents Act requires a complete specification fully and particularly describing the invention. This includes explaining how it works, and disclosing the best method of performing it known to the applicant. Failure to satisfy any of these requirements is an examination objection and a ground for both pre-grant and post-grant opposition. Moreover, failure to fulfill the best mode criteria is a ground for revocation.
Section 10(4)(a) requires that the description enable a PSITA to replicate the invention. CRI Guidelines state that in assessing sufficiency of disclosure for AI-related inventions, applicants should disclose how the AI system converts inputs into outputs, including the training data, to achieve the claimed technical effect. Training data transforms generic AI models into specialized, problem-solving systems. Thus, training data is essential for achieving a technical effect.
Black-box training datasets typically consist of sensitive patient data. While the CRI Guidelines only mandate disclosing the defining characteristics of sensitive data, such descriptions may still carry legitimate re-identification risks. Research indicates that as few as 15 demographic attributes suffice to re-identify 99.98% of individuals. Thus, the concern of patient privacy is significant.
Yet the same may still fall short of what a person skilled in the art (PSITA) demands. In Caleb Suresh Motupalli v. Controller of Patents, the Court, relying on No-Fume Ltd v. Frank Pitchford & Co Ltd, held that a specification lacking sufficient teachings or working examples is insufficient if it improperly forces a PSITA to engage in undue experimentation or exercise inventive faculty.
There exists a disclosure paradox. The very data whose characteristics must be disclosed to satisfy enablement also risks patient re-identification. This leaves applicants structurally unable to satisfy both obligations simultaneously.
The technical reality of this limitation is illustrated by T 0161/18, where merely specifying input data covering a broad patient spectrum was found insufficient for a PSITA to reproduce the network without undue experimentation. Consequently, relying on the CRI Guidelines to provide only data characteristics creates significant rejection risks.
For instance, Patent Application No. 201611017772 was rejected under Section 10(4) because the applicant claimed an exemplary implementation determining “cultural similarity” yet omitted the specific details of how those cultural attributes were evaluated. Consequently, the complete specification was found to have failed to adequately teach the invention. While no Indian decision has directly assessed black-box AI disclosure, the standard applied in Patent Application No. 201611017772 suggests such applications would face similar scrutiny. Ultimately, the inscrutability of generative processes due to their inherent opacity renders sufficient disclosure difficult to achieve.
Section 10(4)(b) of the Patents Act, 1970, requires an applicant to disclose the best method of performing the invention known to them. Non-disclosure of this method constitutes a distinct ground for revocation under Section 64(1)(h).
Tata Global Beverages Limited v. Hindustan Unilever Limited clarified that this requirement does not obligate the applicant to disclose every possible or imaginable method of implementation. Instead, the precedent practically refined the standard by tying the requirement strictly to the applicant’s subjective knowledge and functional sufficiency.
The opacity of black-box medicine is an unavoidable scientific reality rather than deliberate concealment. Following the precedent in Tata Global Beverages, enablement should be viewed through a lens that accounts for practical limitations.
Where the internal functioning of an AI model cannot be fully explained, applicants may rely on functional claiming by defining the invention through the result achieved rather than the exact method used. For example, a claim may describe a system diagnosing sepsis risk with specified sensitivity and specificity levels without disclosing the model’s internal architecture or learned weights. While functional claiming is not prohibited, the CRI Guidelines 2025 require that such claims remain tied to a specific technical purpose and be supported by definite descriptions of the hardware, firmware, or software performing the function.
However, functional language does not cure insufficiency under Section 10(4)(a). The specification must still disclose how the claimed technical effect can be reliably achieved. Vague or non-implemented concepts in the specification cannot be rescued by the framing of claims alone. Functional claiming therefore represents a partial and imperfect tool.
Thus, AI disclosure remains challenging due to its reliance on algorithms that identify covert patterns to arrive at diagnostic conclusions. Under the Patents Act, 1970, the right to apply for a patent vests in the “true and first inventor” or their assignee. The Act does not define “inventor,” but there is no express statutory provision in India prohibiting AI as an inventor or applicant. However, in the prominent case of DABUS Patent application, the Indian Patent Office held that an inventor must be capable of possessing and assigning legal rights, providing nationality and address details, executing assignments, and assuming legal responsibility. These attributes are, however, absent in an AI system. The Assistant Controller reasoned that because inventive step is assessed from the perspective of a person skilled in the art, “the conception of invention should also originate from a human mind.” Thus, a black-box AI system cannot legally be recognized as an inventor for its prescription.
Can India Nurture Black-box Medical Innovation?
Firstly, AI models can be protected through trade secret law, instead of patents, as invention is claimed without revealing the “recipe” behind the model. It does have a drawback. Trade secrets do not protect information if it is independently discovered or reverse-engineered. Thus, a hybrid model (patent + trade secrets) could be implemented.
However, in medicine, this approach conflicts with regulatory transparency and patient safety. Trade secrecy also fails to promote knowledge diffusion through enablement, which is an essential goal of the patent system. This would undermine the ethical and public-interest foundations of medical innovation. In the medical industry, public benefit through innovation and competitive pricing is essential.
As legal scholar W. Nicholson Price II highlights, traditional patents and regulatory exclusivities fail to effectively protect new uses or dosing protocols discovered by these systems. Under Section 3(d), one cannot patent a known substance’s new form, property, or use without added efficacy, nor an existing process unless it yields a new product or reactant.
This blocks patents on algorithmic outputs from the black-box as it merely customizes a prescription for an individual. This shifts the focus of intellectual property protection to the underlying process itself. For process patents to work under Indian law, the strict disclosure requirements of Section 10 must be adapted to accommodate inherently opaque computational systems.
As Price observes, fragmenting data into corporate silos yields weaker medical insights, harming the entire field. To avoid private monopolies, the Black-box process should operate as state-owned public infrastructure, fostering a competitive, multi-player ecosystem. This multi-player ecosystem preserves innovation incentives without surrendering the infrastructure to monopoly.
This public-benefit model is critical during pandemics, neutralizing risks like the licensing bottlenecks and overcharging targeted by the WTO Pandemic Agreement to ensure equitable healthcare access. Such an approach would align with India’s commitment to public health while acknowledging the realities of modern AI research.
Conclusion
India’s compliance with TRIPS Article 29 creates structural friction for nondeterministic black-box AI, as the technology’s unpredictable nature cannot satisfy classical disclosure and enablement requirements. This mismatch is compounded because India still retains the obligation to disclose the best method while examining a patent application as a patent. While this strictness protects transparency, it might hinder the objectives that seek to shape the future of medicine responsibly, with the help of Black-box AI. India’s IP framework must evolve, adapting foundational patent principles to technologies that no longer operate like traditional inventions.
Addressing this challenge does not require abandoning foundational patent principles, but rather rethinking how they apply to technologies that no longer operate like traditional inventions. The question is no longer whether Black-box medicine will shape healthcare, but whether India’s IP system will evolve quickly enough to shape its development responsibly.