Artificial intelligence has created a set of intellectual property questions that existing statutes were not written to answer. Indian law has not been rewritten for AI, and no comprehensive AI-specific IP legislation is in force. What exists is the Patents Act, 1970, the Copyright Act, 1957, and the Designs Act, 2000, applied to a technology their drafters did not contemplate.
This article sets out where the settled law is clear, where it is genuinely unsettled, and what an inventor or business working with AI should do in the meantime. Where a question is open, we say so rather than offering false certainty. Anyone with a specific matter should take specific advice, because several of these issues turn closely on facts.
Patenting AI inventions in India
The first question is almost always whether an AI invention can be patented in India at all. The answer is that it depends on how the invention is characterised, and the analysis runs through Section 3 of the Patents Act.
Section 3(k) and what it excludes
Section 3(k) provides that a mathematical or business method, a computer programme per se, or algorithms are not inventions within the meaning of the Act.
Every word there does work. The exclusion for computer programmes is qualified by "per se", and Indian practice has developed around what that qualification means. The general position that has emerged is that a claim directed to an algorithm as such, or to a computer programme as such, falls within the exclusion, while a claim to a technical solution producing a technical effect beyond the ordinary operation of the computer may not.
The Indian Patent Office publishes guidelines on the examination of computer-related inventions, which have been issued and revised over the years and which examiners apply in practice. Applicants and their advisers should work from the current version of those guidelines, since the approach has shifted more than once and the version in force at the time of examination is what matters.
Indian courts have also considered Section 3(k) in a number of decisions, and the case law has developed towards an approach focused on technical contribution and technical effect rather than on the form of the claim. The details of that jurisprudence are beyond the scope of a general article, and the practical point for an applicant is that outcomes turn on the specific technical effect asserted and how well the specification supports it.
What this means for drafting
The practical consequences for anyone preparing an AI patent application in India:
- Identify a technical problem and a technical solution. An application framed as "a better prediction using a neural network" is weaker than one framed around a specific technical problem in a technical field and how the claimed method solves it.
- Describe the technical effect concretely. Reduced memory footprint, lower latency, improved sensor accuracy, reduced power consumption, improved control of a physical process. Effects that are only commercial or informational are less likely to assist.
- Do not hide the AI, but do not lead with it. The invention is the technical solution; the model is part of how it is implemented.
- Support the claims with detail. Architecture, training approach, data handling, and how the components interact. Sufficiency objections under Section 10 are common in AI applications, because the specification often describes the outcome without enabling the skilled person to achieve it.
- Include fallback positions. As with any Indian application, amendments are limited to matter disclosed in the specification as filed, so anything you may later need as a narrowing limitation must be in the document from the start.
Note also that Section 3 contains other exclusions that may be engaged depending on the field, including in respect of methods of treatment, and that these are assessed independently of Section 3(k).
Our patent practice handles computer-related inventions, including drafting for the Indian exclusions alongside foreign filings where the exclusions differ.
Can an AI be an inventor?
This has been litigated in several jurisdictions following applications naming an AI system as inventor, and the consistent outcome in the major patent systems that have decided the question has been that an inventor must be a natural person. The reasoning varies by statute, but the direction of travel has been uniform.
The Indian position follows the structure of the Patents Act, which is framed around a "true and first inventor" and contemplates a person. An application naming a machine as inventor faces a formal difficulty on that basis.
The practically important point is different, and it is often missed. Almost no real AI invention is made by an AI alone. Inventions are made by teams using AI tools, and the humans who identified the problem, selected and structured the approach, designed the architecture, curated the data and recognised the significance of the output are the inventors. Documenting who did what, contemporaneously, is worth doing, because inventorship disputes are easier to prevent than to resolve.
Where AI tools have been used substantially in the inventive process, the sensible course is to keep records of the human contributions rather than to speculate about the legal status of the tool.
Copyright in AI-generated and AI-assisted works
The Copyright Act, 1957 protects original literary, dramatic, musical and artistic works, cinematograph films and sound recordings. Authorship is central to the scheme, and the Act's definition of author for computer-generated works refers to the person who causes the work to be created.
Two situations should be separated.
AI-assisted works, where a human makes the creative choices and uses AI as a tool. These are the ordinary case and are treated much as any other work, with the human as author.
Works generated with minimal human creative input, where a person supplies a short prompt and the system produces the output. Whether such output attracts copyright, and if so who the author is, is genuinely unsettled in India as in most jurisdictions. The requirement of originality, which Indian courts have addressed in terms of skill and judgment rather than mere effort, sits awkwardly with output produced substantially by a system.
Businesses relying commercially on AI-generated content should therefore not assume they hold enforceable copyright in it. Practical mitigations include ensuring meaningful human creative input and recording it, relying on contract and confidentiality where copyright is uncertain, and protecting the surrounding assets, brand names through trademark and distinctive product appearance through design registration, which do not depend on the same authorship analysis.
Training data
Training data raises questions under copyright law that are being actively litigated internationally and that Indian law addresses only through general principles.
The Copyright Act contains fair dealing provisions in Section 52 covering specified purposes including private or personal use, research, criticism and review, and reporting. Indian fair dealing is enumerated rather than open-ended, which makes it structurally different from the fair use doctrine in some other systems. Whether and how these provisions apply to large-scale ingestion of protected works for model training has not been settled by Indian courts, and the position should be treated as open.
For a business, the practical risk management measures are unglamorous and effective:
- Keep records of data provenance, including licences and terms of the sources used.
- Read the terms of use of scraped or third-party sources, since contractual restrictions apply independently of copyright.
- Check the licence terms of any pre-trained models and datasets used, including whether they restrict commercial use or impose obligations on outputs.
- Consider indemnity positions in supplier contracts where models are procured rather than built.
- Note that data protection law applies independently where training data contains personal data, and that this is a separate compliance analysis from copyright.
Trade secrets and the alternative to filing
India has no dedicated trade secrets statute, and protection rests on contract, equitable principles of confidence, and practical security measures. For AI work this is frequently the most useful instrument, because model weights, training methodologies, data curation pipelines and hyperparameter choices are often better kept confidential than published in a patent specification.
The trade-off is the usual one. A patent requires disclosure and gives a time-limited exclusive right. A trade secret requires secrecy and lasts as long as secrecy does, with no protection against independent development or reverse engineering. For a fast-moving model architecture that will be superseded before a patent grants, secrecy is often the better answer. For a durable technical method embodied in a product, a patent may be.
What we would advise doing now
- Classify each AI asset by what it actually is: a technical method, a model, a dataset, content, or a brand.
- Apply the appropriate instrument to each rather than defaulting to patents.
- Where patenting, draft around technical problem and technical effect, with sufficiency in mind.
- Document human contributions to inventions and creative works contemporaneously.
- Maintain data provenance records from the beginning, since reconstructing them later is usually impossible.
- Treat the unsettled questions as unsettled, and structure contracts so that your position does not depend on how they are eventually resolved.
Talking it through
AI IP questions rarely have a clean general answer, and the useful conversation is usually about a specific system, a specific dataset and a specific commercial objective.
Our innovation and startups practice works with technology companies on exactly these decisions, and a free consultation is available for an initial discussion. You can also reach us directly through the contact page.
