Ownership and Protection of AI Model Weights in India

Introduction : In the process of a company developing an AI product, the software that actually brings the product to market is often not the training code or paper outlining the architecture. A good team can reconstruct, publish or copy both of those within a couple of weeks. The trained weights themselves, the numerical parameters derived from weeks or months of compute, curated data and iterative tuning, are something that can't be reproduced without massive expenses. Give a competitor your architecture and training scripts, minus the weights, and they have a recipe, but no dish. Give them the weights and they are set.

This imbalance is now being seen in the context of fundraise calls, M&A discussions, disagreements with existing employees and in the terms and conditions of contracts with contractors and cloud service providers, even though Indian intellectual property laws were in existence before anyone trained a neural network. It's a very specific, but ever-hotter, question that is raised in this article: What, if anything, in Indian law protects model weights, and who claims them when development is distributed across employees, contractors, open-source components and cloud infrastructure? It ends with a checklist of ownership tips for Indian AI firms to take notes from.

Legal Provisions

A. Copyright Act, 1957

Section 2(o) defines ‘computer programme’ as a literary work and section 2(ffc) defines a computer programme as a set of instructions expressed in words, codes, schemes or any other manner that can make a computer perform a task. This would comfortably include source code and object code. A trained weight matrix is not a command to the computer, but it is a number that is generated when instructions are run over data. Section 17(c) does give works created in the course of employment the automatic protection of the employer, but only when the output itself is capable of copyright protection in the first place, which is a question workers struggle to meet.

B. Patents Act, 1970

Mathematical methods, algorithms and computer programmes are not patentable per se under section 3(k). At the end of the day, a trained weight matrix is just a giant matrix of numbers that were trained via some method to minimise a loss function and that's the same category into which the provision was drafted to prevent it from falling. No matter how it's trained, a trained weight matrix is essentially: a big matrix of numbers, which is pretty much what the provision was written to exclude from the patent system.

C. Information Technology Act, 2000 (act 214)

Section 72A makes disclosure of information obtained under a lawful contract, without the consent of the person concerned, a criminal offence if done "with the intention of causing wrongful loss or gain. The result is that this provides companies with a statutory remedy that complements the common law breach of confidence action and is especially significant in cases where weights are passed on from the cloud infrastructure provider, non-disclosure agreement (NDA) holding contractors, or third-party auditors.

Legal Analysis

A. Confidential Information: The Best Fitting Doctrine

There is no separate trade secret law in India, the rights to trade secrets are provided through contract and the equitable doctrine of breach of confidence. The precedent of the English case of Coco v A.N. Clark (Engineers) Ltd, [1969] RPC 41, where the claimant would be required to establish (1) the fact that the information was of a confidential nature, (2) that it was communicated in circumstances that import an obligation of confidence, and (3) that it was used by the claimant without authorisation to their detriment, has been followed in Indian courts.

Weights make better sense under this system than any registered IP category, as their market value is directly linked to their secrecy. A published novel can still be valuable after the publication and can be copyrighted for decades before the copying of the published portion can occur; a model's weights become devalued as soon as they leak or are published, and an opponent can simply load and use them. The doctrine's Achilles heel is also structural: it can be as strong as the access barriers, NDAs and contract chains which a company actually establishes; and it does not protect against a competitor creating an equivalent model without any cooperation.

B. Weights as Software: A Weak Fit

Another question that is quite similar is if weights can be considered as part of the ‘software’ and protected in the same manner as code. In Eastern Book Company & Ors. v. D.B. Modak & Anr., (2008) 1 SCC 1, the Supreme Court ruled that the requirement of originality in Indian copyright law cannot be reduced down to the level of mechanical work, but requires ‘skill and judgment' to be performed. Even this small threshold is hard for weights to overcome: no one chooses, organises or evaluates the numeric values produced by gradient descent. The skill and judgment that one has to apply to design the architecture, to select the hyperparameters and to curate the training data is real, but so is the process, not the matrix, and Indian courts have so far kept the two in separate categories. The notion of authorship in Section 2(d)(vi) refers to the computer producing a computer-generated work, such as an AI-generated image or a text, which awkwardly overlies weights, which are not the same thing.

C. Database Rights: A Weaker Analogy Still

Unlike the EU Database Directive (96/9/EC) which gives rise to a sui generis database right, in India, only copyright in a ‘selection or arrangement’ reflecting ‘genuine creative effort' exists, as stated in Burlington Home Shopping Pvt. Ltd. v. Rajnish Chibber, 1995 (35) DRJ 210, which held that a compiled mailing database was eligible for copyright protection. That doctrine assumes facts or records that have existed beforehand and have been managed by a person in some sort of judgment. The weights do not come from a set of pre-existing data points, but are parameters an optimisation algorithm pulls out of the data, and no human effect on the single values. This is too much even in a jurisdiction where there is a sui generis right, and it's a dead end without one in India.

D. Trade secret protection in practice 

Combined, the above analysis suggests that the only plausible legal resting place for weights was confidential-information and trade-secret doctrine, with the criminal remedy in section 72A of the IT Act available in cases where there is a contractual relationship. The Delhi HC, in the case of John Richard Brady v. Chemical Process Equipments (P) Ltd., AIR 1987 Delhi 372, had ruled that technical and engineering information shared in a business deal can also enjoy confidentiality; a doctrine that is easily transferrable to weights to be shared with cloud vendors or contractors in the context of NDA. So the protection of weights is not dependent on any registered or statutory right, but upon the measures which a company can take to protect them in its contracts, access and internal hygiene.

Ownership questions

A. Employees

Works created in the course of employment are subject to section 17(c) of the Copyright Act, which puts the output of an employee in the ownership of the employer, provided that the output is itself an original work of authorship, and as explained above, weights may not be. Employers relying on a ‘works made in the course of the employment' clause when assuming that a model's weight falls under the employer are making an assumption which is not supported by the statute. When the Delhi High Court had to consider the question of ownership of a product of the working relationship in Diljeet Titus vs. Alfred A. Adebare & Ors. 130 (2006) DLT 330, it was found necessary to scrutinize the terms of the working relationship, rather than assuming a transfer of ownership, a strict requirement which is even more important in the context of an asset which is not covered by copyright at all like weights. Therefore, when drafting employment contracts, the following terminology should be explicitly assigned: ‘model parameters’, ‘trained weights’ and ‘outputs of training and fine-tuning processes’, not ‘works’ or ‘inventions’.

B. Contractors and Freelancers

The section 17(c) default does not apply to independent contractors, even for copyrightable material, and there are no similar defaults for weights. The very thing the client thinks he's buying, the resulting weights and checkpoints, might be something the ML engineer or the boutique vendor has a real stake in or, at least, a lot of negotiating leverage. This gap should be filled before the first training run, at the engagement-letter stage, not afterwards when the contractor has the only copy of the weights.

C. Open-Source Components

‘Open weight’ is not the same as public domain. The terms and conditions of community licences are often easy to miss when a start-up has optimized a base model and starts to consider it their own. The terms and conditions of community licences are often easy to overlook once a start-up fine-tunes a base model and starts to view it as their own. Typically, derivative weights will also be subject to the base model's licence, and so a company's ownership rights over its modified weights are only as secure as the base model's ownership rights; this should be monitored on a model-by-model basis, not at the time of intake.

D. Cloud Providers

Weights are generally not housed on the company's infrastructure, so it raises a new question: who has technical access to these weights and what are the confidentiality and data-processing commitments that are binding to the cloud provider? Standard cloud terms of service are targeted at generic customer data, not an asset whose entire value rests on its concealment, and don't often cover what happens to model artefacts at the end of the term, in the event of a backup by the provider or during a security incident. The problem with these gaps is that the trade-secret doctrine above can only be as effective as the access controls that surround the information and a provider's typical terms may not provide them.

Relevant Case Laws

In Eastern Book Company & Ors. v. D.B. Modak & Anr., (2008) 1 SCC 1, the Supreme Court had set the Indian originality threshold at ‘skill and judgment’ which cannot be met by the trained weight values since no individual chooses or arranges the numbers.

The three-part test, quality of confidence, obligation of confidence, unauthorised detrimental use, which Indian courts apply to breach of confidence claims, including technical information, such as weights, is provided by Coco v. A.N. Clark (Engineers) Ltd., [1969] RPC 41 (England).

John Richard Brady v. Chemical Process Equipments (P) Ltd., AIR 1987 Delhi 372: granted confidentiality protection to technical and engineering information submitted to a business associate, with which weights might have been exchanged with contractors or vendors.

Burlington Home Shopping Pvt. Ltd. v. Rajnish Chibber, 1995 (35) DRJ 210 (Delhi): recognised copyright in a compiled database based on skill and judgment in selection and arrangement, a test weights do not meet since they are not selected from pre-existing data points.

In Diljeet Titus v. Alfred A. Adebare & Ors., 130 (2006) DLT 330 (Delhi), the court stressed that the actual employment/engagement relationship must be taken into account to decide as to who is the owner of the work product, and cautioned against assuming that the employer is automatically the owner.

Practical Implications

At every point where an outside party comes in contact with the model, no registered right protects weights directly, so ownership must be developed and documented in a contract. So the acquirers and investors are testing exactly what the company can demonstrate and document, that it has all the inputs, training data, the base model licence, the contractor agreements, the employee assignments, and that it owns them. The deal risk and value of a deal will immediately reflect one weak link in that chain. An Indian AI company's ownership checklist should comprise of:

1. Employment contracts with explicit assignment clauses covering ‘model parameters’, ‘trained weights’ and ‘outputs of training and fine-tuning’, not generic ‘works’ language.

2. Previously signed contract or freelancer agreements assigning specific weights or checkpoints created.

3. An inventory of all open-source base models licence types used, with monitoring and reporting against field-of-use, redistribution and revenue-threshold conditions.

4. Terms of confidentiality, data-processing and access commitments with cloud providers, not only generic customer data, reviewed.

5. Access-control documentation on the inside for who has access to the weights and audit logs of access to the weights.

6. NDAs signed with each party, with each party granted meaningful access to the weights.

7. Documentation of training data origins, to control downstream copyright exposure which may contaminate the learned weights.

Conclusion

Indian IP law was drafted without any intention of a ‘large matrix of trained numerical parameters' and any attempt to insert weights into copyright, database rights and patent doctrine falls short. The value of what weights actually are, is entirely based on their secretiveness and upon the contracts and equitable doctrines in favour of their secretiveness, and nothing upon any registered or statutory right.

While that's correct, it's a light answer. It has nothing against independent development, nothing when the weights are shared, and nothing if the contracts in the first place were not properly drafted. The hard part for Indian companies in AI is not a new legal concept, it's just simple contract law: Approach weights as a trade secret right from the start, don't rely on presumptions from statutes that may not exist, and be prepared to demonstrate to an investor, acquirer or court exactly how each one of the inputs to those weights was acquired.

Author: Shivangi in case of any queries please contact/write back to us via email to content@khuranaandkhurana.com or at  Khurana & Khurana, Advocates and IP Attorney

Endnotes

1. Copyright Act 1957, ss 2(o), 2(ffc), 17(c) and 2(d)(vi).

2. Patents Act 1970, s 3(k).

3. Information Technology Act 2000, s 72A.

4. Eastern Book Company v DB Modak (2008) 1 SCC 1.

5. Coco v AN Clark (Engineers) Ltd [1969] RPC 41.

6. John Richard Brady v Chemical Process Equipments (P) Ltd AIR 1987 Delhi 372.

7. Burlington Home Shopping Pvt Ltd v Rajnish Chibber (1995) 35 DRJ 210.

8. Diljeet Titus v Alfred A Adebare 130 (2006) DLT 330.

9. Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases [1996] OJ L77/20.