AI Training, Copyright And The Limits Of Fair Dealing

• UPSC Syllabus Tags: GS Paper III—Science and Technology: developments and their applications and effects in everyday life
• Context: The article examines how the Delhi High Court’s interim ruling on AI training attempts to reconcile copyright protection, technological innovation and access to information.
• Source: “Is training an AI model ‘public interest’? Delhi HC verdict goes to heart of copyright debate,” The Indian Express, July 28, 2026.

Two-Minute Summary

The Delhi High Court refused an interim injunction against OpenAI, holding prima facie that storing ANI’s literary works for LLM training fell within fair dealing under Section 52(1)(a) of the Copyright Act.

The ruling distinguished statistical or non-expressive training use from outputs reproducing protected expression and found insufficient evidence that ChatGPT had memorised or substantially reproduced ANI’s works.

The decision reduces immediate uncertainty for AI developers but does not settle the main suit or India’s wider policy choice between unlicensed training, opt-outs and mandatory remuneration.

Legal Context

• ANI alleged that OpenAI stored its copyrighted material for training and generated infringing or falsely attributed outputs.
• The Court considered an application for temporary injunction under Order XXXIX of the Code of Civil Procedure.
• Its findings are expressly prima facie and have no binding effect on the final determination of the suit.
• DPIIT’s competing policy proposal, One Nation, One License, One Payment, favours blanket licensing with remuneration for right-holders.

Essential Background

• Section 14 of the Copyright Act grants specified exclusive rights, including reproduction and electronic storage.
• Section 52 identifies acts that do not constitute infringement, including fair dealing for specified purposes.
• Copyright protects original expression, not the underlying facts, ideas or linguistic rules.
• Interim relief requires a prima facie case, balance of convenience and likelihood of irreparable injury.

Key Terms

• Fair dealing: A statutory limitation permitting specified uses when their purpose, extent and market effect are considered fair. It is contextual rather than unconditional.
• Non-expressive use: Use directed at extracting patterns, relationships or statistical information rather than communicating the protected expression to an audience.
• Statutory licence: A legal mechanism permitting use of protected works subject to prescribed conditions and remuneration without separately negotiating with every right-holder.

Analysis

Training and Output Are Distinct Legal Stages

LLM development may involve collecting material, temporarily storing it, converting language into numerical representations and adjusting model parameters. A later output need not retrieve a verbatim copy of the training document.

The Court accepted that storing a protected work may initially engage the reproduction right. It nevertheless treated training as fair dealing on the present evidence because the material was used to extract linguistic and statistical relationships rather than republish ANI’s expression.

This finding does not immunise generated outputs. An output substantially reproducing protected material may still infringe copyright. A different model, dataset or evidentiary record could consequently produce a different result.

The Court’s Fairness Test

The Court examined whether OpenAI’s use was confined to training, whether it competed with or injured ANI’s legitimate market and whether the resulting system served the public interest.

ANI had not established loss of subscribers or injury to its syndication business. The Court also emphasised that factual information and news events cannot be monopolised merely because a publisher reported them first.

Public interest received substantial weight. The judgment reasoned that LLMs facilitate access to information and have applications across education, healthcare, finance, agriculture and skill development. Requiring separately negotiated licences from numerous sources could make model development economically difficult, especially for smaller Indian developers.

Commercial Use and Market Substitution

OpenAI operates commercially, but the Court did not treat commercial purpose as conclusive. Section 52 expressly restricts commercial use in some contexts but does not use the same limitation in Section 52(1)(a).

Commercial purpose nevertheless remains relevant to market harm. A system generating substitutes for paid articles, books, photographs or music raises stronger concerns than one extracting abstract patterns without making the protected work available to users.

Right-holders argue that models derive commercial value from human-created material while creators bear the cost of producing it. An uncompensated training exception could weaken investment in journalism, publishing and cultural production.

Creator Incentives and Licensing

Mandatory licensing may distribute some AI-generated value, but it could impose high compliance costs, entrench large collective-management bodies and disadvantage domestic start-ups. Opt-outs reduce compulsory transaction costs but shift the burden of exclusion onto creators.

Parliament could instead establish a specific text-and-data-mining framework covering training-data transparency, machine-readable reservations, collective remuneration and continuing liability for substantially reproducing outputs.

Editorial Lens

The article correctly warns against allowing a non-final order to become a substitute for settled copyright policy. Interim decisions influence platform behaviour even though they are based on incomplete evidence and do not finally determine rights.

The judiciary can interpret the existing statute and decide whether a claimant has established infringement. Designing an economy-wide remuneration system involving multiple sectors, technologies and classes of right-holders requires legislative and regulatory choices.

Public interest also requires careful definition. AI may generate substantial social value, but technological usefulness cannot excuse every unauthorised use. Public interest includes continued investment in reliable journalism and creative production.

The balanced conclusion is that the judgment supplies a fact-specific interim response, not a comprehensive AI-training exemption. Durable policy must distinguish input-side analysis from output-side substitution while distributing innovation benefits and compliance costs fairly.

Prelims Focus

• Facts and ideas are not ordinarily protected by copyright; original selection, arrangement and expression may be protected.
• India’s statutory expression is “fair dealing”; it is not identical to the open-ended United States fair-use doctrine.
• An interim order does not finally determine the rights of the parties.
• Retrieval-augmented generation supplies external material at query time and is distinct from pre-training.
• Copyright infringement and false attribution are separate legal questions.

Mains Answer Enrichment

• Legal anchor: Section 52 balances exclusive copyright with research, criticism, reporting, education and dissemination of knowledge.
• Case study: In July 2026, interim relief was refused because ANI did not establish substantial reproduction, market injury or irreparable harm.
• Balanced formulation: Innovation policy should prevent protected expression from becoming substitute content without granting control over facts, ideas and linguistic patterns.
• Reform: Enact a technology-neutral text-and-data-mining framework covering transparency, opt-outs, creator remuneration and output liability.