Trade Secrets in the Age of AI: Can Internal AI Models Be Protected?

Introduction

AI ("Artificial Intelligence") has come a long way from being an experimental technology to one of the most valuable commercial assets of the modern economy. From automating processes to generating insights, from enhancing decision-making to personalizing customer experiences, from detecting fraud to optimizing supply chains, and from creating competitive advantages to what was once impossible, businesses across industries are increasingly turning to AI systems to accomplish these tasks. With billions of dollars being spent by organizations on building their own AI solutions, there is one basic legal quandary that has come to the fore: how to prevent competition from getting hold of those AI models ?

In the past, the protection of intellectual property has been linked to patents, copyrights, trademarks and industrial designs. AI systems pose special challenges, however, that cannot be classified into the conventional categories of existing IP rights. Copyright, generally speaking, protects expression and not functionality, although software code can be protected by copyright. Likewise, patents could cover some aspects of AI-related invention, but getting to patent protection will involve making technical information public, which businesses may wish to keep under a proprietary lock. As a result, companies are looking for trade secret law as an alternative means to protect their most valuable AI assets.

This is a significant issue that can be seen from the practices of some of the leading technology companies in the world. When organizations create sophisticated AI systems they rarely reveal details regarding their model architecture, training methods, datasets, reinforcement learning approaches, optimisation, or deployment. They do not simply rely on patents, but often consider such information to be confidential business assets. This is indicative of a shift toward understanding that the act of being secret might be a form of protection in itself in the AI world.

Let's look at the top AI firms and how they implement strategic secrecy. OpenAI has been silent on key aspects of GPT-4's architecture, training data, and model parameters, which they attribute to competitive and safety concerns. Likewise, Anthropic keeps a lot of its details of its Claude models confidential, and Google has long adhered to a rule of keeping much of its sophisticated AI development work under wraps. Meta, on the other hand, has taken a relatively open stance with its Llama family of models by making model weights available for download under certain licensing terms. The conflicting views demonstrate the current debate in the field of AI between open and closed research, open and closed innovation, and open and closed patents. Several major AI developers are still being very secretive, which indicates that the protection of trade secrets is an essential aspect of today's AI tactics.

Thus, the question comes up: Is an internal AI model a trade secret? If it is, what elements of the model can be safeguarded, what legal requirements need to be met and what problems may occur in preventing enforcement of protection? With the changing face of commerce around the world, businesses, legal professionals and policy makers alike have increasingly been asking these questions.

Understanding Trade Secrets

Unlike patents, trademarks and copyrights, trade secrets are valuable because they are kept a secret. In its broad sense, a trade secret is the information that is not generally known, which has commercial value and is the result of reasonable efforts to keep it secret.

Foreign trade secret protection is acknowledged worldwide through Article 39 of the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS). Within the context of this framework, information can be protected if it is secret, has commercial value because of its secretness and is the subject of reasonable measures to ensure its secrecy.

India does not have a specific trade secret law in place. However, Indian courts have always safeguarded confidential information through contractual means, equitable principles and breach of confidence suits. These can be manufacturing processes, formulas, customer databases, business methods, technical know-how, source code, algorithms, pricing information, or proprietary methods.

Lack of a specific statutory regime does not imply that trade secrets are not protected. Instead, Indian law has evolved a flexible policy favoring protection of commercially valuable confidential information where appropriate protections are put in place. The key issue, then, is whether internal AI models meet traditional trade secret requirements.

The United States has one of the most well-developed legal systems concerning trade secret protection. Most states have enacted the Uniform Trade Secrets Act ("UTSA"), which provides a fairly consistent definition of trade secrets and remedies for misappropriation. The federal Defend Trade Secrets Act, 2016 ("DTSA") established a new civil cause of action in every state for trade secret misappropriation. Both laws consider information a trade secret if it has independent economic value because it is not generally known and has been the subject of reasonable efforts to keep it secret. The wide scope of the DTSA and UTSA definitions are broad enough to cover contemporary technological assets such as algorithms, software systems, machine learning models, information like datasets, and other proprietary information related to AI. While conflict over AI keeps cropping up, U.S. trade secret law is becoming a more valuable tool for other jurisdictions, like India, that lack any specific legislation on trade secrets.

The commercial implications of Internal AI Models

Modern AI systems are much more complex than traditional software applications. They can be created after years of research, a huge amount of computation, a lot of expertise in engineering and access to valuable data.

The importance of an AI model is not just in one piece of it. Instead, it comes from many factors, like model architecture, training data, parameter design, fine-tuning, reinforcement learning algorithms, optimization methods, evaluation methods and deployment infrastructure.

For instance, a financial institution could train an AI system to identify fraudulent transactions with a high degree of accuracy. Likewise, a health care provider could develop a diagnostic model using their own clinical information gathered over years of use. Recommendation engines created by analyzing consumer behaviours over a long period of time may be used by e-commerce platforms.

In every of these cases, the AI model is a huge competitive benefit. A competitor that is granted access to the underlying methodologies is able to repeat years and years of research and development – without the same expense. This means that organizations face a significant business imperative to keep important information regarding their AI systems confidential.

The assets associated with artificial intelligence (AI) continue to grow in value and are now regarded as valuable business assets like traditional trade secrets like manufacturing formulas or proprietary industrial processes.

Does an internal AI Model meet the requirements of a Trade Secret?

Secrecy

The first element needed for trade secret protection is that the information is not generally known or easily ascertainable. There are a number of internal AI models that easily meet this need. Most organisations don't want to release the models, training methods, code, hyperparameters or proprietary data sets. This information is typically only accessible to a select few engineers, researchers and executives. Importantly, it may not only be a matter of the model, but also of the surrounding. The parts of the AI ecosystem can be considered eligible for protection independently, such as:

• Model architecture

• Training database

• Fine-tuning methodologies

• Reinforcement learning processes

• Hyperparameter configurations

• Evaluation metrics

• Internal documentation

• Source code repositories

• Deployment frameworks; and

• Proprietary workflows

If these elements are not available publicly and not accessible by the public, it is likely that it will meet the requirement for confidentiality.

Commercial Value

The second requirement is related to commercial value in terms of secrets. Competitors often don't have access to internal AI models, which often bring significant value. An advantage in the economy created by an AI system can be contingent upon the lack of competition for that system.

For instance, if an algorithm can predict the trajectory of the market more effectively than other systems, the owner of the system can get a lot of money. Similarly, an AI model that has been constructed based on a company's own healthcare information might have diagnostic functions and characteristics that are not found in other models.

In such cases, the secret of the model itself adds directly to the value of the economy. This is similar to the basis of trade secret protection.

Reasonable Measures to Maintain Secrecy

The most difficult may be the third requirement. Typically, businesses are required to show evidence of taking reasonable steps to maintain confidentiality. Just saying that information is confidential is not enough. It is crucial that organizations put measures in place to actively defend the information with concrete and legal precautions.

Examples of these measures are:

The most difficult may be the third requirement. Typically, businesses are required to show evidence of taking reasonable steps to maintain confidentiality. Just saying that information is confidential is not enough. It is crucial that organizations put measures in place to actively defend the information with concrete and legal precautions.

Examples of these measures are:

• Non-disclosure agreements

• Confidentiality clauses in employment contracts

• Restricted access controls

• Encryption mechanisms

• Cybersecurity safeguards

• Internal confidentiality policies

• Employee training programs; and

• Lack of such measures may severely limit any trade secret rights.

Choosing Between Patent and Trade Secret Protection for AI Models

A crucial strategic consideration for any artisan who is developing an AI is whether it is better to seek patent protection or to keep it as a trade secret.

There are several benefits of patents. They give them exclusive rights which can be enforced against third parties, and there is no need for perpetual secrecy. But for patent protection, the invention needs to be disclosed to the public. When patent applications are published, other parties have the opportunity to obtain technical information that could be useful for further innovations.

AI firms face a big hurdle in complying with this disclosure requirement. Training methods, model architectures, and optimization techniques are seen as commercially sensitive information for many organizations that want to keep this information confidential.

Secondly, patent prosecution is a time-consuming and costly process. On the other hand, AI technologies are rapidly changing. The technology itself could be significantly altered by the time the patent is granted.

An alternative is trade secret protection. However, as long as there is secrecy, protection may theoretically be continued indefinitely. For over 100 years, the Coca-Cola formula has been a trade secret.

This open-ended time frame is a great benefit for many AI developers. But, trade secret protection can be exposed to independent invention, reverse engineering and inadvertent disclosure. In contrast with patents, trade secrets do not prevent others, without the owner's knowledge, from developing the same technologies. Organizations, therefore, should be mindful to assess if the commercial goals are better served by patent protection or by trade secret protection or a combination of both.

In practice, many organizations are increasingly employing a hybrid protection approach. Some innovations that can be reverse engineered or are likely to meet the patentability criteria can be patented; sensitive information like training methodologies, proprietary datasets, model optimization techniques, and deployment mechanisms, can be kept as trade secrets. In this way, businesses can enjoy both types of protection without having to endure the drawbacks of either. Hybrid IP solutions are likely to be the standard approach, as the sophistication of AI systems increases.

The Risk of AI Trade Secret Misappropriation – Employee Mobility.

One of the biggest risks to AI trade secrets is that employees move on. There are very few places to hire away from the talent competition in the AI industry. Engineers move often between rival organizations, along with researchers and developers. This mobility opens up the possibilities of unconsciously or intentionally giving out confidential information to others.

An outgoing worker might have a comprehensive understanding of architectural design, training methods, optimization strategies and proprietary data sets. In some situations, the employee's knowledge may affect his or her ability to work later with a competitor without the information being moved.

Indian courts have been called upon to decide several cases relating to confidential information and rosters. The Delhi High Court, in American Express Bank Ltd. v. Priya Puri, reiterated that although an employee may not be prevented from using the general skills and knowledge he learns in the course of his work, but if it is “confidential information” it is protected.

Likewise, in John Richard Brady v. Chemical Process Equipment's Pvt. Ltd. the Delhi High Court held that the protection of confidential business information from unauthorized disclosure is important.

Artificial Intelligence and Cybersecurity Trade Secrets

Trade secret protection depends solely on making the trade secret a secret. For these reasons, cyber security has become an important element of trade secret management. AI models are often deployed in cloud systems, on private computers, and over decentralized systems. Unauthorized access to these systems can be a source of code, model weights, datasets, and proprietary documents being stolen.

Intellectual property protection has been the target of more sophisticated cyberattacks. Authorities, companies, and cybercrime syndicates could be interested in gaining entry to valuable AI technologies. The repercussions of these violations may be dire. If the disclosure of critical AI components results in the loss of the trade secret, then the entire property is lost and there are no legal rights to protect.

Cybersecurity needs to thus be seen as a part of an organization's IP strategy instead of just an IT problem.

The Challenge of Reverse Engineering

Another big challenge is related to reverse engineering. AI systems can have direct interactions with users, unlike traditional trade secrets which are only physically inaccessible. The competitors are allowed to make inferences about the architecture of models, describe patterns in behaviour and analyse outputs.

This concern is reflected in the emerging area of model extraction attacks. By engaging in multiple interactions with deployed AI systems, sophisticated actors could try to build up a picture of how these systems work.

In the future, courts will likely be asked to decide whether information acquired through these means is likely to affect trade secret status. Such concerns have not been addressed sufficiently and will continue to be a major challenge for the protections of AI systems in the legal system.

India's Position on Protecting AI Models

This aspect has not been directly explored by Indian Courts but the judgments available suggest that this is the possibility. India already has a legal framework in place to safeguard confidential information, commercially important information and methodologies of a business and proprietary information. The principles of trade secret and breach of confidence cases are complex and adaptable enough to new technologies.

India has no specific legislation in the area of trade secrets and this has been seen as a lacuna in the IP regime in India. Unlike patents, trademarks and copyrights, trade secrets are not protected mainly by statute, but by contractual requirements and equitable principles that have been developed by the courts. However, Indian law has consistently shown an inclination in safeguarding confidential business information when parties have exercised reasonable acts to keep business information confidential. This has given the courts a degree of flexibility to ensure that they can keep up to date with the new technology and apply the same rules to new technology as they did to traditional technology.

Contractual safeguards are of particular importance when there is no legislation for protection. Confidentiality must be explicitly stated in employment contracts, consultant contracts, research contracts, vendor contracts and licensing agreements.

But the existing framework might come under pressure due to the increasing use of artificial intelligence in India. In the near future issues like stealing AI models, unauthorized access to datasets, employee turnover and misappropriation of proprietary AI algorithms across regions are likely to escalate. Existing breach of confidence rules may be able to deal with some of these issues, but the nature and scope of potential conflicts involving artificial intelligence may, someday, require a more thorough statutory approach to trade secret protection. Such a law could provide increased clarity in implementation of remedies for, and confidentiality to modern technological assets.

The rise of AI usage is sure to bring with it conflicts dealing with model theft, data set misappropriation, employee mobility, and unauthorized disclosure, which Indian courts will have to grapple with. Such cases will probably continue to influence the trajectory of trade secret protection in the future in relation to artificial intelligence in India.

Conclusion

AI is one of the most important pieces of intellectual capital in today's economy. With significant investments needed to create cutting-edge AI systems, protecting them is an important concern for businesses.

As such, as long as it can be applied to, it's like a patent, but the trade secret law is a very interesting avenue for protection for an internal model, which has not been disclosed, and it does have the same level of enforcement, without losing the confidentiality. The models developed, proprietary datasets, training methodology, fine tuning technique, optimization strategy and deployment framework may each have elements that would traditionally be considered as trade secrets and of commercial value.

The protection isn't a given, though. To keep the trade secret confidential, organizations must have good data security protocols, cyber security protocols, contractual protocols and governance protocols.

New questions will arise in the legal field in the years to come regarding who is the owner of what, confidentiality, reverse-engineering, employee mobility and digital misappropriation. In this ever-evolving context, the most useful AI assets might not be those covered by registration certificates but those safeguarded using astute and well-managed seclusion.

The question now is whether or not an internal AI model could be deemed a trade secret. In most cases it will, if it is demonstrated that the organization can keep it confidential, that it has commercial value and a reasonable effort is made to ensure the confidentiality of the information. The next question is, do companies have the governance, contractual protections and cybersecurity to ensure protection is upheld? As AI becomes a highly competitive sector, with a need for unique models and approaches, it is likely that this will prove to be one of the most important legal tools for AI builders. Solutions to the protection of AI may not be in patented products, but in something that's not being shared in the conference room of some enterprises.

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

References

1. Agreement on Trade-Related Aspects of Intellectual Property Rights art. 39, Apr. 15, 1994, Marrakesh Agreement Establishing the World Trade Organization, Annex 1C, 1869 U.N.T.S. 299.

2. John Richard Brady v. Chemical Process Equipments Pvt. Ltd., 1987 SCC OnLine Del 372.

3. World Intellectual Property Organization, WIPO Technology Trends 2023: Generative Artificial Intelligence (2023).

4. Robert P. Merges & John F. Duffy, Patent Law and Policy: Cases and Materials 1127–29 (8th ed. 2022).

5. David S. Almeling, Seven Reasons Why Trade Secrets Are Increasingly Important, 27 Berkeley Tech. L.J. 1091 (2012).

6. Sharon K. Sandeen & Elizabeth A. Rowe, Trade Secret Law in a Nutshell 5–6 (West Academic, 2d ed. 2017).

7. American Express Bank Ltd. v. Priya Puri, 2006 SCC OnLine Del 638.

8. John Richard Brady v. Chemical Process Equipments Pvt. Ltd., 1987 SCC OnLine Del 372.

9. World Economic Forum, Global Cybersecurity Outlook 2025 18–22 (2025).

10. Ashish Bharadwaj et al., Artificial Intelligence and Intellectual Property: Emerging Issues in India, 15 Indian J.L. & Tech. 1 (2019).

11. V.K. Ahuja, Law Relating to Intellectual Property Rights 1043–45 (LexisNexis, 4th ed. 2022).