Delhi High Court: OpenAI Training Does Not Infringe Copyright — The Global AI Copyright Patchwork Just Got Wider

What the Delhi High Court Actually Ruled

On July 24, 2026, Justice Amit Bansal of the Delhi High Court rejected ANI Media's application for an interim injunction against OpenAI. The ruling is preliminary — it decides the injunction question, not the full case — but the reasoning is what matters.

Justice Bansal held that training LLMs on copyrighted content falls within Section 52(1)(a) of India's Copyright Act, the "fair dealing" exception for research. His core logic: training a neural network on text is fundamentally about statistical pattern analysis, not reproduction. The model does not store or copy the original works — it learns relationships between tokens. That, he concluded, is research, not piracy.

ANI, one of India's largest wire services, sued OpenAI in November 2024. Its complaint had two prongs: first, that ChatGPT was trained on ANI's articles without permission or payment; second, that the chatbot hallucinated quotes and reports, wrongly attributing them to the agency. The court addressed the training question now. The hallucination claim — a distinct legal issue — remains live.

The Global Scoreboard: Every Court Is Writing Its Own Rulebook

This ruling matters beyond India because the global AI copyright landscape is now a full-blown patchwork. The same technology, trained on the same type of content, is producing opposite first-instance verdicts depending on which courtroom you walk into. Here is where the major jurisdictions stand as of July 2026:

Sources: Axis Intelligence AI Copyright Lawsuits Tracker, CourtListener, Reuters, Norton Rose Fulbright. Compiled by ai-jarvis.eu.

The scoreboard tells a clear story: four rulings in the "yes" column, two in the "no" column — and the one that matters most for European readers is the last one.

Why GEMA v. OpenAI Haunts This Conversation

In November 2025, the Munich Regional Court handed German collecting society GEMA a landmark victory against OpenAI. The court held that storing lyrics in model parameters ("memorization") and reproducing them in ChatGPT outputs ("regurgitation") both constitute copyright infringement under German law. Crucially, the court rejected the text-and-data-mining exception in Section 44b UrhG — Germany's implementation of the EU DSM Directive — because memorization exceeds what transient analysis requires.

The contrast with Delhi is stark. Justice Bansal in India drew the line at the training stage and said it was research. The Munich court drew the line at the output stage — if the model can spit back verbatim training data, the training itself was infringement. Two courts, two legal traditions, two opposite answers to the same question.

OpenAI is appealing in Germany. A CJEU referral remains possible, which could produce EU-wide harmonized standards. Until that happens, every AI provider operating in the EU faces real legal uncertainty: if your model can reproduce any scrap of its training data on demand, Munich says you are liable for each reproduction.

What the EU AI Act Adds to the Equation

Even before courts settle the copyright question, the EU AI Act — fully in force since August 2024 with obligations phasing in through 2027 — imposes its own transparency requirements. Article 53 mandates that providers of general-purpose AI models publish a "sufficiently detailed summary" of the content used for training. The AI Office is expected to publish a template for this summary in late 2026.

This is not a copyright licence. It is a disclosure obligation. But the practical consequence is significant: if you trained on copyrighted material and the rights holder can now prove it because you had to disclose a training-data summary, the litigation path becomes much shorter. The Munich ruling gave rights holders a legal theory; the AI Act gives them the evidence-gathering tool.

For European companies building or fine-tuning models, the safe path is increasingly clear: use only licensed or demonstrably public-domain training data, document every source, and test your model for verbatim reproduction before deployment. The days of "scrape first, ask questions later" are not legally over — but the cost of getting it wrong just keeps rising.

The Money Question: What Is This Actually Worth?

The numbers behind these cases are staggering. The Axis Intelligence AI Copyright Litigation Exposure Index pegs cumulative claimed financial exposure across all active AI copyright cases at over $50 billion (approximately €46 billion). The largest confirmed settlement — Anthropic's $1.5 billion (€1.38 billion) deal in Bartz v. Anthropic — established a per-work benchmark of roughly $3,113 (€2,870) per book for pirated training data.

OpenAI itself faces an estimated $10 billion (€9.2 billion) in claimed exposure across 16 consolidated US cases plus international suits. The consolidated MDL in the Southern District of New York — where the New York Times and 15 other plaintiffs are suing — is the heavyweight. Judge Sidney Stein ordered OpenAI to produce 20 million anonymized ChatGPT conversation logs in January 2026, and the outcome of that analysis will likely dictate settlement terms across the entire consolidated litigation.

For context, OpenAI's valuation sat at $852 billion (€786 billion) as of May 2026, and the company went into the Musk trial — which it won unanimously — with that figure intact. A $10 billion exposure is about 1.2% of that valuation. Serious money, but not existential.

What the Delhi Ruling Does — and Does Not — Settle

Three things to keep in perspective:

1. This is an interim order. Justice Bansal was deciding whether to grant ANI an injunction while the case proceeds — not delivering a final judgment. ANI can (and likely will) argue the full case on its merits. But the reasoning on fair dealing sets a strong precedent for Indian courts.

2. The hallucination claim is separate. ANI's argument that ChatGPT fabricated quotes attributed to the agency is a defamation or false-attribution claim, not a copyright claim. The court did not rule on it. This matters because hallucination liability is an emerging legal front globally — if a model invents damaging information and attributes it to a real person or company, is the AI developer liable? No court has given a clear answer yet.

3. India is not Europe. A Delhi ruling has zero binding effect in Munich, London, or New York. But in a world where AI companies are being sued in multiple jurisdictions simultaneously, every ruling becomes ammunition in every other courtroom. OpenAI's lawyers in Germany will cite Justice Bansal. GEMA's lawyers in Munich will cite the Munich ruling. The global conversation is one conversation.

Practical Takeaway: What This Means If You Build or Use AI

If you are a developer or company working with LLMs in the EU, here is what actually changes: nothing, yet. The European legal position remains what it was on Thursday morning — the Munich ruling is the operative precedent within Germany, and the EU AI Act's transparency obligations are what they are.

But the direction of travel is worth watching. The Delhi ruling adds to a growing body of first-instance decisions that treat AI training as legally permissible under existing copyright frameworks — Bartz, Kadrey, Getty UK, and now ANI. The counter-current is Munich, and potentially the Third Circuit in Thomson Reuters v. Ross if the appeals court upholds Judge Bibas's finding against fair use.

The real test case for Europe is the GEMA appeal. If the Munich Court of Appeal affirms the lower court — or if the CJEU takes it up — European AI developers face a fundamentally different legal landscape than their American or Indian counterparts. The cost of training a frontier model in Europe, with European-licensed data only, would be multiples higher than training it with broadly scraped data. That is either a competitive disadvantage or a competitive advantage for European rights holders, depending on which side of the table you sit on.