AI and patents, trade secrets and training data in India

Four questions, four different regimes

An AI project in India touches four bodies of law at once. Patent law asks whether the technique can be protected. Inventorship rules ask who may be named. Confidence law protects what stays secret. Copyright decides what may go into the training set.

They do not line up neatly. A choice that helps under one can hurt under another. Publishing a patent specification, for instance, ends secrecy over whatever it describes.

Patents: section 3(k) and technical effect

Section 3(k) of the Patents Act, 1970 removes certain subject matter from the definition of an invention:

a mathematical or business method or a computer programme per se or algorithms

The words "per se" carry the weight. Software is not excluded as a class. A claim is excluded when it is nothing more than the program or the algorithm.

The Delhi High Court read the provision this way in Ferid Allani v Union of India (2019). Prathiba Singh J, on 12 December 2019, directed the Patent Office to re-examine the application and treated technical effect or technical contribution as central. A computer-related invention is not barred if it shows one.

The same court returned to the point in Microsoft Technology Licensing, LLC v Assistant Controller of Patents and Designs (15 May 2023, C.A.(COMM.IPD-PAT) 29/2022). The application concerned authentication using two cookies to reach sub-locations of a network. The Controller had refused it as a computer program per se. The court set that aside. As reported, it said that running algorithms in a fixed sequence on a conventional computer does not by itself remove a technical effect, and that the examiner must look at substance, not claim form. The matter is reported to have gone back to the Controller to decide novelty and inventive step.

A further Delhi High Court decision of 2 August 2024, Ab Initio v Assistant Controller, concerned divisional applications 6500/DELNP/2011 and 6501/DELNP/2011. It applied the same technical effect test. It found a technical effect in features such as faster dependency analysis and reduced computation. The Controller had not dealt with inventive step, so the court sent the matter back for that.

What the 2025 guidelines say about AI

The Patent Office published draft revised Guidelines for Examination of Computer Related Inventions on 25 March 2025, with comments invited until 15 April 2025. The Patent Office website lists a final set titled Guidelines for Examination of Computer Related Inventions (CRIs) - 2025. The page gives no date or version number for it. Commentary reported its release in August 2025. This note relies on secondary summaries of the contents of both the draft and the final set.

As summarised, the guidelines treat an AI or machine learning model as abstract in itself. It can become patentable when it produces a practical application with a demonstrable technical effect. The listed effects include higher speed, reduced memory access time, better control of a robotic arm, improved signal reception and security enhancement in networks.

Disclosure gets specific attention:

• Claims should describe a concrete implementation, not a speculative one.
• Where the model's internal logic is not disclosed, the specification should explain how input becomes output, with the processes, parameters and variables.
• For a claim built on a trained model, the applicant should identify the dataset, say why it is relevant, and show how it produces effective results.
• Where the inventive step lies in the data's characteristics, such as origin, labelling or diversity, those characteristics should be disclosed unless obvious. The full dataset is usually not required.

That last point bears on secrecy. The guidelines ask for the architecture, training method and parameters, supported by validation results. A patent specification written to that standard will disclose much of the training pipeline. Model weights themselves are not asked for in the summaries read.

Inventorship: the DABUS refusal

Section 6(1) of the Patents Act allows an application by a person claiming to be the true and first inventor, by an assignee, or by a legal representative. Section 2(1)(y) defines "true and first inventor" and excludes the first importer into India and a person to whom the invention was first communicated from abroad.

Dr Stephen Thaler filed in India naming DABUS, his AI system, as the inventor. The application was for a food container with fractal features. According to published commentary of April 2026, the Assistant Controller refused it. The commentary does not give the order date or the application number. The reasoning had two parts. First, an AI system cannot be recognised as the true and first inventor under sections 2, 6, 7 and 10, because only natural persons qualify, and owning the machine does not supply the proof of right that section 7 needs. Second, the claims lacked inventive step over prior art on containers, and no data showed an unexpected advantage.

The same commentary reports that the PCT route did not help. Section 138(4) still requires a formal proof of right. Whether Dr Thaler appeals is not confirmed.

The guidelines summary points the same way. AI-generated inventions are not patentable because an AI cannot be an inventor. AI-assisted inventions may be, if a human contributes and the technical contribution is clear. The practical record question is who conceived the claimed features, and whether that can be shown.

Trade secrets: contract and equity

India has no trade secret statute. Protection rests on contract and on the equitable duty of confidence. Delhi High Court authority such as John Richard Brady v Chemical Process Equipments (Delhi High Court, 6 July 1987) is the usual starting point, with courts granting injunctions where information with a confidential character was passed in circumstances that imposed a duty.

The Law Commission of India's 289th Report (dated 5 March 2024) annexed a draft Protection of Trade Secrets Bill, 2024. The report and commentary on it indicate that:

• the report's discussion of TRIPS lists three tests for a trade secret: secrecy, commercial value from secrecy, and reasonable steps to keep it secret;
• it notes that individuals may use skills and knowledge gained in earlier jobs, and that post-employment restraints should not be permitted;
• remedies under the new law should be civil only. Criminal action may still follow under other laws;
• whistleblowers should get immunity, not just a defence, where they act in good faith and in the public interest.

The report is a recommendation. Nothing in the material read shows that the government has acted on it or that Parliament has enacted it.

What to hold as secret across an AI stack

The "reasonable steps" element is likely to matter in any court, statute or not. Different layers of an AI system carry different risk.

• Model weights. These are the most concentrated asset and the hardest to patent. They are usually kept secret, with access controls and logging.
• Training pipeline. Data filtering rules, hyperparameters, curriculum and evaluation harnesses often hold more value than the architecture, which is usually published research. A patent specification may force part of this into the open.
• Training data. The curated and labelled set can qualify as confidential even when its parts are public. Its source terms may limit what can be said or shared.
• Prompts and system instructions. These leak easily through outputs and are weak candidates.

Contracts do the work that a statute would. Staff and contractor agreements, vendor terms and access policies are the evidence of reasonable steps. A restraint on an employee after employment faces section 27 of the Indian Contract Act, 1872, so confidentiality clauses carry more weight than non-compete clauses.

Copyright: training data and outputs

Two separate questions arise. May a developer copy protected works to train a model? And who owns what the model produces?

Training

On 24 July 2026, Amit Bansal J of the Delhi High Court decided the interim application in ANI Media (P) Ltd v OpenAI OPCO LLC (CS(COMM) 1028/2024), reported as 2026 SCC OnLine Del 5291. The court refused an injunction. On a prima facie view, storing ANI's works to train a language model fell within section 52(1)(a) of the Copyright Act, which covers fair dealing for private or personal use including research. The court applied a three-part inquiry: whether the use was confined to training, whether it harmed ANI's legitimate commercial interests, and whether it served the public interest.

On outputs, the court found that ANI had not shown memorisation or regurgitation. The responses relied on appeared to come from retrieval, which ANI had not pleaded as infringement. The court also held, prima facie, that it had territorial jurisdiction even though the training servers are abroad. The suit continues. Every finding is interim and the court said it does not bind the final decision. Commentary expects an appeal. Whether one has been filed is not confirmed in the material read.

A court may take a different view at trial, or a division bench may. Treat the ruling as an indication only.

Policy

The DPIIT committee's working paper (Part I), issued in December 2025, proposed a different route. Under its hybrid model, a mandatory blanket licence would cover lawfully accessed works used in training, and creators would receive a statutory right to remuneration. A government-designated non-profit body would collect and pass on royalties. The rate would be a percentage of the AI system's global gross revenue, set by a government committee. The comment period was reported as extended to 6 February 2026. Part I is reported to leave the copyright status of outputs to a later part. Part II was not found when this note was checked.

Outputs

Section 2(d)(vi) names the author of a computer-generated literary, dramatic, musical or artistic work as:

the person who causes the work to be created

That text predates generative models. Whether a user, a developer or neither fits "causes" is untested in the material read. For now, ownership of outputs is a matter to settle by contract.

A worked example

The numbers below are assumptions to show the mechanics, not a forecast. The royalty rate in the DPIIT paper has not been fixed.

Assume a developer's AI product earns global gross revenue of INR 500 crore in a year. Assume the rate committee sets a royalty at 1 per cent of global gross revenue. The annual royalty would be INR 5 crore (500 crore x 0.01).

Now assume the product is trained on a corpus of which only a share is registered with the body. The paper makes only registered works eligible. The pool still depends on the full revenue figure, not on that share. How the pool is divided among eligible works is a design question the paper leaves open.

The point of the arithmetic is the base. A revenue-linked royalty grows with success of the product, regardless of how much of the corpus was protected. Compare that with the ANI ruling, where the court noted ANI had offered a licence for USD 7.5 million and treated that as a sign any harm could be met in money.

Across the four sectors

Technology and services. Software and platform companies are the main users of the technical effect test and the disclosure guidance. They also face the sharpest tension between patenting a pipeline and keeping it secret. Open-source and customer-data terms bear on training-data confidentiality.

Manufacturing. AI that controls machinery, vision inspection or process optimisation sits closest to the technical effect examples in the guidelines, such as robotic control. Claims tied to a physical process are usually easier to present as technical than pure analytics. Sensor data from plants is often confidential and covered by supplier contracts.

Healthcare and pharma. Diagnostic and drug discovery models raise inventorship questions, because a human contribution must be shown. Training data may be patient data, so confidentiality duties run alongside the Digital Personal Data Protection Act, 2023. The Law Commission report also touches test data disclosure in regulatory settings.

Financial services. Credit, fraud and trading models are usually kept as trade secrets, since algorithms and business methods fall within section 3(k) in the first place. Regulatory expectations of explainability can sit uneasily with secrecy over model logic. Licensed news and market data feeds are covered by contract and by copyright, so the ANI ruling reaches them only indirectly.

Limits of this note

This note is general information on the law as at 10 October 2026. It is not legal advice and does not address any reader's own circumstances. Several points rest on interim rulings, draft guidelines and committee proposals that may change. Anyone relying on a point should check the primary text first.