AI-Generated Works and Indian Copyright: Authorship, Training and Ownership

The prompt is not the whole story

A musician asks a generative AI system for a melody; a designer requests a poster; a lawyer uses a drafting assistant. The resulting work may be shaped by the user's instructions, the system's training, iterative selection and later human editing. It is tempting to answer ownership with a simple slogan: the person who entered the prompt owns everything. Indian copyright law does not support such a universal shortcut.

Section 2(d)(vi) of the Copyright Act, 1957defines the author of a computer-generated literary, dramatic, musical or artistic work as the person who causes the work to be created. The difficult question is what the phrase means for contemporary, probabilistic generative systems. Does the relevant human contribution lie in the design of the system, control of the process, selection of outputs or substantial creative editing? The statutory text supplies a starting point, not a fully litigated answer to every generative-AI scenario.

Training data and outputs must be analysed separately

Copyright owners generally possess exclusive rights concerning protected works, subject to statutory exceptions. TheCopyright Act's Section 52 exceptionsinclude specified forms of fair dealing, but India does not have a blanket, express exemption authorising every commercial AI-training use. Whether a particular act of copying or processing is protected will turn on the use, the rights implicated and the applicable statutory defence.

Training creates a different question from the final output. A model may learn patterns without routinely producing verbatim copies, yet reproductions may occur during data ingestion or in some outputs. Conversely, an output that resembles a pre-existing work does not necessarily infringe it: substantial similarity, protectable expression, independent creation and the relevant rights matter. These questions require facts about the technology and particular works, not assumptions based on the word'AI'.

The contractual layer

Platform terms may allocate certain rights between developer and user, but a contract cannot necessarily grant a user more statutory copyright than the contracting party can lawfully convey. Organisations should therefore examine the provider's terms, permissions for confidential inputs, output-use restrictions and risk allocation. A commercial licence to use a system does not mean the output is guaranteed original, non-infringing or exclusively owned.

For creators, maintaining records of prompts, drafts and significant human revisions may help explain the creative process and originality. For companies, provenance checks are especially important when an output incorporates recognisable text, images or music. Licensing and consent remain practical ways to reduce uncertainty.

Five questions that should be asked separately

First identify whether copyright subsists in the claimed material. A question about an original literary or artistic work begins with the statutory category and originality requirement. A document consisting of factual information, a human arrangement of that information and an AI-produced image may involve different claims within the same file.

Next identify the author. Section 2(d)(vi) addresses computer-generated works by reference to the person causing creation. Its application to present-day generative systems should be reasoned from the actual creative process. It does not confer legal personality on an AI model.

First ownership is another question. Employment, commissioning arrangements, statutory exceptions and assignments can affect who holds rights even where authorship is known. Contractual permission to use an output should then be examined separately from a claim to statutory exclusivity. Finally, infringement asks whether somebody has exercised a protected right without authorisation or an applicable exception.

Human editing: what the record can establish

Consider a designer who generates a background image, then draws original figures, changes composition and writes the accompanying text. The assessment should identify those contributions rather than claim the entire package solely because a prompt was entered. Original human expression and the uncertain status of generated elements may coexist.

Retain version history, source layers and a record of substantial revisions. Those materials can show which expressive choices were made by a person. They cannot guarantee protection: the legal standard still applies, and changing a file format or selecting an output is not automatically equivalent to creating original expression.

The same caution applies when a generated output includes a recognisable work. A human contribution to one layer does not authorise copying another person's protected expression in another layer. Ownership claims and infringement exposure must be assessed independently.

Why access to a work does not settle training rights

A work being available online answers an access question. It does not necessarily establish permission for every reproduction, extraction or commercial use. A training dispute needs facts about what was copied, where the acts occurred, the rights held and the statutory or licensed basis claimed.

Section 52 contains purpose-specific exceptions. An analysis should test the actual exception relied upon rather than import another country's broad fair-use language into Indian law. Nor does an absence of verbatim output by itself resolve whether a relevant reproduction occurred earlier in the process.

For commercial users, practical questions include permitted input material, confidentiality, output restrictions, warranties and indemnities. Where a supplier offers an indemnity, examine its exclusions and claims procedure. That contractual allocation can affect who bears a loss between the parties, but it does not determine what rights a third-party creator holds.

A framework for the next dispute

Courts will need to distinguish authorship, subsistence, first ownership, contractual allocation and infringement rather than collapse them into one question. Legislators and regulators, meanwhile, must consider the effects of any new exception or licensing mechanism on innovation, authors' bargaining power and access to knowledge.

A reliable assessment identifies the work claimed, the human contribution, the material used and the particular exclusive right in dispute. A prompt log can help explain that record, but it cannot supply copyright that the statute does not recognise.