Introduction

Disruptive technologies like AI provide fascinating solutions to long term problems but bring unique challenges with them. In the AI development industry, protection of commercial intangible property is one such problem. Companies working in cutting-edge AI are currently facing a dilemma when it comes to choosing a suitable legal framework to protect their commercially valuable know-how. While patent law provides strong protection, it demands complete disclosure, something which is not suitable for an industry like AI development where training data and algorithmic know-how are too commercially valuable to be disclosed completely. On the other hand, trade secret protection imposes no such disclosure requirement. But with a complete lack of a dedicated legal framework, one can only protect their IP under trade secrets and hope that nobody else makes independent discoveries of their own or reverse engineer the technology. In such a situation, the AI industry is left to grapple with the problem of optimally protecting innovation and not losing the competitive advantage in the market.

The aim of this blog is to explore the patent and trade secret regimes with respect to AI, examine the doctrinal and practical limitations with both, compare approaches from the US, EU, and China, and examine emerging proposals for India.

Outside the Silos

Most modern AI systems consist of different core components[1] including the system architecture or algorithm, the trained model parameters, the datasets on which the AI was trained as well as the deploying infrastructure, and the background know-how. The issue arising here is because of the novel nature of these components themselves- the characteristics of these components make them unsuitable to be protected under any traditional IP regime.

India’s Patents Act explicitly excludes “a computer program per se or algorithms[HT1] ” from patentability.[2] This has been consistently relied upon by the Indian Patent Office (IPO) to deny patents on pure algorithms without being tied to a tangible technical application and showing a clear “technical effect”.[3] While the Patent Office’s new guidelines on Computer Related Inventions (CRIs) give a welcome clarification that the focus is not on the mere presence of a software but its application and effect (meaning substance is being prioritized over the form of the invention), thus reducing uncertainty, this reasoning may take time to seep into claim evaluations by the IPO.  In such a scenario, a novel model architecture is likely to be considered an unpatentable abstract mathematical method. Thus, businesses are choosing trade secrets as the default protection regime[4]

When it comes to training datasets, Copyright law may provide some protection to curated datasets, but since most AI training datasets aim for comprehensiveness or representativeness rather than creative selectivity, a merely comprehensively compilation of training data would not be an original work that can be copyrighted. Patent law does not provide any relief either, as obviously there is nothing inventive about a dataset consisting of facts and information. Therefore, an AI training dataset in India is usually only protectable as a trade secret (kept confidential via contracts).  Similarly, the trained model weights, arguably the most valuable output of the training process, defy traditional IP categorization. These weights are nothing, but numerical parameters generated from training an AI model, and they hold considerable value. However, this very numerical nature of training model weights makes them ineligible to qualify either for copyright or patent protection. Thus, industry players are on their own to navigate through this legal uncertainty.

AI and Section 3(k): Patent Ambiguities

Patent law, in its current state, can raise several hurdles in protecting AI and associated intangible property. This is obvious in India due to Section 3(k)’s explicit bar and its consequent narrow interpretation by the Indian Patent Office[HT2] . Consequently, this bar means that the claimed invention must exhibit a clear “technical effect” beyond the mere algorithm or software. This becomes a mounting challenge as when it comes to AI specific and AI assisted inventions,  it is very difficult to just understand the reasoning employed by the AI, let alone sufficiently disclose how exactly did the AI performs such inventive tasks.  Although there have been rulings[5] by Indian courts in favor of broader[6] interpretation of Section 3, no positive developments have occurred in AI-specific contexts so far.

India is not the only jurisdiction to treat algorithms and mathematical methods as unpatentable abstract ideas. The U.S. Supreme Court’s Alice test, for example, says that a software or AI concept must include a specific “inventive concept” or technological improvement to be patent-eligible, otherwise it is deemed an unpatentable abstract idea[7], leading the Courts throughout the US to invalidate patents that merely apply generic machine-learning techniques to conventional activities without a concrete technical advance.[8] European patent law similarly excludes algorithms “as such” from patentability, allowing AI-related patents only when a technical effect is achieved in a field of technology.[9]

Even when an AI invention is patent-eligible in principle, the innovator faces the disclosure trade-off inherent in patenting. The patent application must disclose the algorithm or system in enough detail that a skilled person could reproduce it,[10] which is challenging for complex AI models – often functioning as opaque “black boxes” whose inner logic can be difficult to fully explain or replicate outside the original development context.[11] This is further exacerbated by the speed (or lack thereof) of the patent process itself.  Procedural delay may extend for years before a patent is granted, while technology in consideration evolves exponentially. This raises questions about the value of patenting fast-moving AI advancements[12] and often makes the 20-year term of a patent less compelling in practice.

Do secrets pay off?

Businesses are increasingly shifting towards the comparatively uncharted waters of trade secret law over the public disclosures required by patents.  While this strategy does seem powerful, it is fraught with legal uncertainty and practical enforcement challenges, particularly in India. The scope of what can be protected as a trade secret is uncertainly broad, especially in the AI context where several components have commercial value. Apart from the training datasets, these components also involve specific implementation of model architecture and code, as has been evident by the closely guarded secrets of OpenAI’s GPT-4[13] and DeepMind’s strategic shift towards commercial secrecy with AlphaFold3[14]. However, this protection is not without conditions; the owner must take reasonable measures to maintain secrecy, such as using NDAs and access controls. This legal regime varies significantly across jurisdictions, creating a complex international landscape. In India, the lack of a dedicated statute is the focal point of the uncertainty that the industry face, as protection is often grounded in common law principles like breach of contract, increasing the need for iron-clad agreements. While the Law Commission of India’s report[15] and draft bill suggest a way forward, the current framework shows a reality opposed to the robust, harmonized frameworks of the US, with its Defend Trade Secrets Act (DTSA) of 2016, and the EU’s Trade Secrets Directive. Yet, even with strong laws, relying on trade secrets brings significant enforcement headaches. The detection of theft is often not fast enough, the burden of proof in Indian courts is high and a considerable litigation risk of inadvertently exposing the secret you are trying to protect persist. Ultimately, while trade secrets are a vital tool, the practical difficulties of enforcement mean that prevention, through a fortress of contracts and internal security, remains a far better strategy than the cure.

Choosing the Right Path

The choice between patenting an AI innovation or guarding it as a trade secret represents a critical strategic calculus for any developer, often hinging on the innovation’s vulnerability to public scrutiny and reverse engineering.[16] A patent becomes the logical safeguard for any technology that will be publicly accessible or must be disclosed for regulatory purposes, as secrecy offers no protection against independent replication. In contrast, for systems used exclusively in-house or as opaque cloud services—and particularly in a field evolving faster than the patent prosecution timeline—trade secrecy offers a nimbler and more effective shield. In practice, this is rarely a binary decision. Many sophisticated companies adopt a hybrid approach, patenting a foundational inventive concept to block competitors while keeping the crucial implementation details, such as proprietary datasets, specific source code, and tuning techniques, as closely guarded trade secrets.[17] However, such a strategic decision carries considerable implications for public interest. The patent regime is designed to foster innovation by compelling disclosure in exchange for a limited monopoly, thereby enriching the public domain with technical knowledge. If secrecy becomes the developers’ default, a legitimate concern about the stagnation of the collective progress of the AI field arises. This mandates that policymakers engage in a delicate balancing act. Narrow & rigid patent laws excluding many software-based inventions can inadvertently push innovators towards secrecy, undermining the very purpose of the patent system.

Conclusion

Artificial Intelligence raises an unprecedented challenge to the foundations of the traditional IP regime, raising urgent calls for policy reform. To resolve this tension, innovative yet untested solutions are being discussed to deal with issues like the “black box” nature of many AI systems, including permitting confidential deposits of source code with regulatory bodies or employing cryptographic methods like zero-knowledge proofs to validate an invention’s efficacy without revealing its core algorithm.[18] Additionally, frameworks in which disclosures are made to authorized independent auditors under agreed upon mandatory confidentiality are also being considered.[19] Going forward, reform is most likely through a multi-pronged approach which clarifies standards for patent AI in unambiguous legal terms and provides statutory recognition to trade secret protection regimes, thereby maintaining the much-needed balance between proprietary interests and transparency & accountability requirements for AI development.

---

[1] Sam Mitchel, Nichol Bashor & Heather Ehlers, Risk Management Magazine – protecting AI Innovation: Why trade secrets are outpacing patents in IP portfolios Risk Management Magazine (2025), https://www.rmmagazine.com/articles/article/2025/07/17/protecting-ai-innovation-why-trade-secrets-are-outpacing-patents-in-ip-portfolios#:~:text=Unlike%20patents%2C%20trade%20secrets%20do,that%20flexibility%20can%20be%20invaluable. (last visited Jul 18, 2025).

[2] Section 3(k), Patents Act, 1970.

[3] DPS Parmar, India – protecting “Smart algorithms” under patent law. Conventus Law (2024), https://conventuslaw.com/report/india-protecting-smart-algorithms-under-patent-law/ (last visited Jul 18, 2025).

[4] Chloe Xiang, OpenAI’s GPT-4 is closed source and shrouded in secrecy VICE (2024), https://www.vice.com/en/article/openais-gpt-4-is-closed-source-and-shrouded-in-secrecy/#:~:text=The%20report%2C%20whose%20sole%20author,” (last visited Jul 18, 2025).

[5] Ferid Allani v. Union of India, W.P.(C) 7/2014.

[6] Microsoft Technology Licensing v. Assistant Controller, C.A.(COMM.IPD-PAT) 29/2022.

[7] Alice Corp. Pty. Ltd. v. CLS Bank Int’l, 573 U.S. 208 (2014).

[8] Recentive Analytics, Inc. v. Fox Corp., 10 F.4th 927, 931–33 (Fed. Cir. 2021).

[9] Convention on the Grant of European Patents, art. 52(2)(c)–(3), Oct. 5, 1973, 1065 U.N.T.S. 199.

[10] 35 U.S.C. § 112(a) (2018); Marrakesh Agreement, Annex 1C, Trade-Related Aspects of Intellectual Property Rights art. 29(1), Apr. 15, 1994, 1869 U.N.T.S. 299.

[11] Supra Note 1

[12] Ibid.

[13] Supra Note 4

[14] Dr Rose Hughes, Alphafold: From Nobel prize to drug-discovery gold mine? The IPKat (2024), https://ipkitten.blogspot.com/2024/11/alphafold-from-nobel-prize-to-drug.html#:~:text=Trade%20secrets%20were%20never%20the,24 (last visited Jul 18, 2025).

[15] Law Commission of India, Report No. 255: Protection of Trade Secrets (2023) (proposing a Trade Secrets Bill 2024)

[16] R Mark Halligan, Patents versus trade secrets: Factors to consider | reuters Reuters (2024), https://www.reuters.com/legal/legalindustry/patents-versus-trade-secrets-factors-consider-2024-04-09/ (last visited Jul 18, 2025).

[17] Matthew Kohel, Trade secrets may offer the best protection for AI Innovation Law360 (2023), https://www.law360.com/ip/articles/1577741/trade-secrets-may-offer-the-best-protection-for-ai-innovation (last visited Jul 18, 2025).

[18]Enrico Bonadio, Eduardo Alonso & Vansh Tanyal, The Ai Black Box Issue and Patent Disclosure City Law Forum (2025), https://blogs.city.ac.uk/citylawforum/2025/05/06/the-ai-black-box-issue-and-patent-disclosure/?utm_source=chatgpt.com (last visited Jul 18, 2025).

[19] Ibid.

Authored By: Mr. Ashutosh Jha, NALSAR University of Law