India's Copyright Act, 1957 has no text-and-data-mining exception and no settled case law on whether training an AI model on copyrighted text is infringement or fair dealing. For Indian startups, this means training-data risk cannot be resolved by reading a statute — it has to be managed contractually, especially where the scraped data is also personal data under DPDP.
Indian courts have not decided a single AI-training copyright case, unlike the US where New York Times v. OpenAI is actively testing the fair-use defense — leaving Indian startups to guess at a standard that does not yet exist while still carrying real commercial risk.
What Changed
- The US "fair use" defense for AI training is being tested in court (NYT v. OpenAI, Andersen v. Stability AI) with no final ruling yet — India has no equivalent litigation to draw guidance from.
- India's Copyright Act, 1957 offers only a narrow "fair dealing" exception (Section 52), which is stricter and more use-specific than the US "fair use" doctrine — it was not written with model training in mind.
- Where training data is also personal data (resumes, chat logs, KYC documents), the DPDP Act applies independently of copyright, creating two separate legal exposures from one dataset.
- Global AI vendors' indemnification terms (or lack thereof) are becoming the de facto risk allocation mechanism for Indian buyers, since local law offers no clear floor.
The Details
The Core Conflict: Is Training "Fair Dealing"?
The fundamental question — does scraping the internet to train a model constitute permissible use — is unresolved globally and untested in India specifically.
The global fair-use argument: AI labs argue training is transformative, akin to a student learning language patterns rather than copying text. NYT v. OpenAI challenges this directly, alleging ChatGPT can recite Times articles verbatim.
Why India's "fair dealing" is narrower: Section 52 of the Copyright Act permits specific enumerated uses — private study, criticism, review, research — rather than the US's open-ended "transformative use" test. A court asked to fit "training a commercial LLM" into one of these categories would be doing real interpretive work, not applying settled precedent.
The DPDP Overlap: Two Regimes, One Dataset
For Indian startups, the practical risk is rarely pure copyright — it is the overlap with personal data law. Scraped resumes, support transcripts, or public social media posts can simultaneously be:
- Copyrighted expression (the specific wording, images, or creative arrangement), and
- Personal data under DPDP (if it identifies a living individual)
A dataset can be copyright-clean (public domain, licensed) but still violate DPDP if it contains personal data collected without lawful purpose. Conversely, DPDP-compliant data collected with consent can still infringe a third party's copyright in the underlying content. Treat these as two separate checklists, not one.
Output Ownership: Who Owns What the Model Generates?
If an Indian team uses a generative tool to produce a logo, article draft, or code snippet, authorship is genuinely unclear under current law.
- India's position is undefined. The Copyright Act requires an "author," and no amendment or ruling has clarified whether AI-assisted work with minimal human input qualifies.
- The US position (persuasive, not binding): the US Copyright Office has repeatedly refused registration for works lacking meaningful human authorship.
- The safer contractual assumption: treat AI-generated output as protectable only where a human has made substantive creative choices and can document that "chain of custody" if ownership is ever challenged.
What This Means for Indian Founders and CTOs
- Separate your training-data checklist from your DPDP checklist. Run both audits — copyright provenance and personal-data lawful basis — on any corpus before it touches a training pipeline.
- Insist on IP indemnification from your model vendor. Enterprise tiers of major APIs (Azure OpenAI, ChatGPT Enterprise) typically indemnify commercial customers against training-data IP claims; free and consumer tiers usually do not.
- Document human editing of AI drafts if you intend to claim copyright in the output — a visible "chain of custody" from AI draft to human-edited final is your best evidence of authorship if challenged.
- Prefer licensed or "clean" datasets for anything you plan to fine-tune internally, especially for customer-facing outputs where infringement risk is commercially material.
- Flag GitHub Copilot–style code suggestions for open-source license contamination (GPL) before merging into proprietary Indian codebases — this is a live risk regardless of copyright's AI-training uncertainty.
Frequently Asked Questions
Does Indian copyright law allow AI training on copyrighted text?
It is untested. The Copyright Act, 1957 has a narrow "fair dealing" exception, not the broader US "fair use" doctrine, and no court has yet ruled on whether AI training qualifies.
If training data has personal information, does DPDP apply on top of copyright law?
Yes. Copyright protects the rights-holder's expression; DPDP protects an individual's personal data. A scraped resume or chat log can trigger both regimes simultaneously.
Who owns the output of a generative AI tool under Indian law?
Unclear. The Copyright Act requires an "author" and India has not clarified whether AI-generated work without meaningful human authorship qualifies for protection.
Should Indian startups use enterprise AI models over free-tier APIs for IP protection?
Yes for production use. Enterprise tiers (Azure OpenAI, ChatGPT Enterprise) typically include IP indemnification that free or consumer tiers do not.
Related reading: How India's DPDP Act Affects AI Training Data, AI Liability in India, AI Regulation in India: A Business Guide, Open Source vs. Regulation, and the AI Compliance Starter Kit.
