Global Copyright Rules for AI Training Diverge Sharply. India Needs Clarity.
As AI companies train large language models on copyrighted material, four major jurisdictions have adopted starkly different legal approaches. The United States relies on case-by-case fair use judgments. The European Union has written explicit statutory exceptions into law. China enforces strict compliance requirements. India has yet to clarify where its copyright law stands on the issue.
The legal uncertainty matters. In India's first major test case, ANI v. OpenAI, the Delhi High Court is still deciding whether training ChatGPT on copyrighted news content violates Indian copyright law. The outcome will shape how AI developers operate in the country.
How the US Handles It: Flexibility With Inconsistency
The United States has no specific rule for AI training on copyrighted works. Instead, courts apply a "fair use" doctrine that examines four factors: the purpose of use, the nature of the work, how much was used, and the impact on the market for the original.
Courts have reached conflicting conclusions. In Bartz v. Anthropic, a judge held that training on legally obtained works was permissible because it was "transformative." But in Kadrey v. Meta Platforms Inc., a different court upheld fair use even when pirated books were involved, again citing transformation.
The US Copyright Office has expressed a preference for voluntary licensing arrangements rather than government intervention. But without a clear statutory rule, developers face legal risk. Recent cases show courts treating market impact as increasingly critical to their decisions.
The EU Model: Explicit Exceptions and Opt-Out Rights
The European Union took the opposite approach. Its 2019 Copyright Directive and 2024 AI Act provide explicit statutory exceptions for text and data mining.
The framework operates in two tiers. Scientific research can use copyrighted material for mining. Commercial use is also allowed, provided the content was lawfully accessed and rights holders have not opted out through machine-readable signals like robots.txt.
The AI Act goes further. It requires AI providers to implement compliance policies, respect opt-out mechanisms, and disclose summaries of training datasets. This creates enforceable obligations rather than relying on judicial interpretation.
China: Compliance-First, With Judicial Flexibility
China mandates that AI service providers use only data from lawful sources and avoid intellectual property infringement. Its Copyright Law does not provide a specific exception for AI training.
However, Chinese courts have introduced flexibility through a "reasonable use" interpretation. Judges assess purpose, nature, quantity, and market impact - similar to US fair use - but within a framework that prioritizes state oversight and compliance. The approach reflects China's emphasis on regulatory control alongside innovation.
India's Problem: No Clear Rule, Growing Pressure
India's Copyright Act does not address AI training. Section 52 lists "fair dealing" exceptions for research, criticism, reporting, and review. Whether large-scale commercial AI training falls within these exceptions remains unclear.
The ambiguity creates risk for developers. A startup training a model on Indian news content cannot confidently say whether it has legal permission. The ANI v. OpenAI case will partly resolve this, but a court judgment is not the same as a clear rule.
India's government has proposed a solution: a "One Nation, One License, One Payment" collective licensing model. Under this system, AI developers would be permitted to train models on lawfully accessed content without negotiating individual licenses. Royalties would be triggered upon commercialization, with rates set by a government authority and collected centrally for distribution to rights holders.
What India Should Do
India should clarify whether fair dealing applies to LLM training. Explicit policy guidance would prevent further litigation and give developers certainty.
Second, India should operationalize the collective licensing model. This would allow developers to access large datasets legally while ensuring copyright holders receive compensation. Safeguards should include fair pricing, equitable revenue distribution, and provisions for startups to prevent market concentration.
Third, transparency requirements - such as dataset disclosures and audit mechanisms - would protect both innovation and creators' rights.
No single global model has emerged as dominant. The US prioritizes judicial flexibility. The EU prefers statutory rules. China emphasizes state oversight. India has an opportunity to adopt a framework suited to its legal system and development priorities, drawing from all three approaches while remaining attentive to its own context.
For legal professionals, the stakes are clear: copyright compliance in AI development will depend increasingly on statutory clarity rather than litigation risk. India's next move will determine whether that clarity exists.
Learn more about AI for Legal Professionals and how these frameworks affect legal practice.
Your membership also unlocks:
