Section 3(k) for the model, Section 17(b) for the training data, contract IP for the customer deployments. AI/ML in India sits at three IP intersections.
Indian AI and ML startups face the most rapidly-evolving IP environment of any current technology sector. The patentability of an AI invention turns on whether it satisfies the Section 3(k) 'technical effect' test; the copyright in training data and model outputs is contested across multiple legal systems; the trade-secret protection of model weights and architecture is the principal commercial protection for many startups; the contract IP in customer-facing AI deployments raises new questions about who owns the customisations and the fine-tuned models. The Indian legal framework on most of this is at most settling — and AI/ML startups operate in advance of the case law that will eventually clarify it.
This guide is for Indian AI/ML startups building large language models, computer-vision systems, predictive analytics platforms, AI-driven SaaS products and AI agents. The IP framework decides what can be patented, what stays in trade secret, what gets licensed to customers, and what exposure inbound — particularly from training-data sources — needs to be managed.
Three filings cover most of the IP risk on day one. Each is a standalone service and each links to a deeper walkthrough.
Section 3(k) of the Patents Act excludes computer programmes per se. The Ferid Allani technical-effect test from 2019 governs how AI/ML applications are examined. Practical guidance:
Indian AI patent grants have been issued — for image-processing improvements, signal-processing methods, security applications, computational efficiency improvements. The drafting decides the outcome.
For most Indian AI/ML startups, the principal IP protection is trade secret on model weights, training data composition, and proprietary fine-tuning approaches. India has no Trade Secrets Act; protection runs through contract and equity. The India trade secrets framework applies — NDAs with employees and vendors, classification of confidential materials, exit protocols capturing returned information.
For AI startups specifically, the standard stack includes: tight access controls on model weights, structured deployment pipelines that prevent customer extraction, model-watermarking where feasible, and assignment of all training-data preparation work to the company.
The use of copyrighted materials to train AI models is one of the most contested IP issues globally. Indian law has not yet ruled definitively. The relevant frameworks:
Indian AI startups training on internet-scraped data carry undefined exposure. The risk profile of doing so for the Indian market is a strategic question — global precedent will inform Indian outcomes when matters reach the courts.
Indian copyright law currently does not recognise non-human authorship. Outputs generated by an AI without human creative input do not attract copyright. Where human creative input is significant — the prompt, the curation, the refinement — copyright vests in the human contributor. The Indian Copyright Office has issued mixed signals on AI-generated works.
For commercial AI-output deployments, the operational fix is contractual. Customer agreements should specify output ownership, customer's licence to use, and the AI provider's reservations on aggregated learning from customer use.
AI SaaS contracts typically allocate IP in three layers:
Indian AI startups should template these allocations carefully. Default 'all rights to the customer' positions undermine the provider's portfolio; default 'all rights to the provider' positions deter enterprise customers concerned about data leakage. The right allocation depends on the deployment model.
AI startup figuring out what to patent, what to keep secret, and how to draft the customer IP allocation? Send us the stack — we'll map the IP layers and the priority filings.
WhatsApp our team →Possibly. Section 3(k) excludes computer programmes per se, but the Ferid Allani test permits AI patents that produce a technical effect — improved performance, reduced latency, lower memory, better compression, more robust security. The drafting around the technical contribution decides the outcome.
Currently uncertain. Indian fair-dealing exceptions under Section 52 may cover some research-and-study training; commercial-scale training on copyrighted material is not clearly within the exceptions. The safer route is explicit licensing of the training corpus.
Indian copyright requires human authorship. Pure AI-generated outputs may not attract copyright. Where human creative input is significant — prompts, curation, refinement — copyright vests in the human contributor. Commercial deployments typically address this contractually in customer agreements.
Trade secret on model weights, training data composition and proprietary techniques, anchored through NDAs, access controls and structured deployment pipelines. Patents on the technical contributions, trademarks on the brand, and contractual IP allocation in customer agreements complete the stack.