The Real Fight Over AI and Copyright Isn’t About Training Anymore

By Shradhanjali Sarma and Shatakshi Shekhar

On 24 July 2026, the Delhi High Court refused to grant ANI an interim injunction against OpenAI. Within hours, the ruling was being celebrated as a decisive win for AI companies, the first time an Indian court had weighed in substantively on generative AI and copyright, and one that had come out in OpenAI’s favour. While the headline is not false, it flattens a narrow, fact-specific finding into something far broader than the Court decided.

Read closely, the order does not hold that training AI on copyrighted material is lawful as a general proposition. It holds that ANI, on the evidence before the Court, could not show that ChatGPT had memorised and reproduced its specific articles. That is a considerably smaller finding than the coverage suggests, and it happens to be nearly the same question that German courts have already answered, in the opposite direction.

Three findings emerge from the Delhi order. First, the Court held that OpenAI’s storage of ANI’s content for training its models falls within the “private or personal use, including research” exception under Section 52(1)(a)(i) of India’s Copyright Act, so the storage itself does not amount to infringement. Second, it found that ChatGPT’s actual outputs, generated through a retrieval-augmented architecture that pulls in live external facts before drafting a response, were not substantially similar to ANI’s original articles, which is the standard test for reproduction infringement in India. Third, and this is the finding getting lost in the coverage, the Court noted that ANI had simply failed to prove that any “memorisation” or “regurgitation” of its work was occurring in ChatGPT’s responses. That absence of proof was doing the real analytical work in the ruling. This was not a categorical shield for AI developers. It was a case where a specific evidentiary burden had not been discharged.

Notably, the Court also sided with ANI on the threshold question of jurisdiction, rejecting OpenAI’s argument that Indian courts had no business hearing a claim against a company whose servers sit in the United States. The underlying suit now proceeds to trial despite the refusal of interim relief. The Court further reasoned that an injunction at this stage would cause real harm, not just to OpenAI but to the public, given how embedded the tool already is in daily use across India.

There is something else worth noticing here, easy to miss beneath the “AI wins” framing. The Court did not reach for imported American doctrine to arrive at its conclusion. Indian copyright jurisprudence has a documented history of borrowing American concepts without a clear statutory basis for doing so. The Supreme Court did exactly this in EBC v. DB Modak, importing originality and derivative-work standards developed under a US Copyright Act that has no true Indian counterpart. India’s own statute contains no general derivative-work right of the kind found in Section 106 of the US Act. What it has instead is a narrower right of adaptation under Section 14, and a correspondingly narrower fair-use framework under Section 52. In the ANI order, the Court stayed inside that statutory framework: Section 52 for the training and storage question, ordinary substantial-similarity analysis under Section 51 for the output question. There is no visible attempt to construct an Indian analogue of “transformative use.” Whether that discipline survives once the detailed order is published, and whether it holds through trial, remains to be seen. But on what has been reported so far, the Court appears to have avoided a detour that commentators had been warning about for months.

Now set this beside Germany. In GEMA v. OpenAI, decided by the Regional Court of Munich I in November 2025, Germany’s music rights collecting society demonstrated that ChatGPT could reproduce nine well-known German songs almost verbatim on simple request. OpenAI raised the defence that has become standard in these disputes: the model does not store specific text, it only reflects statistical correlations learned across an entire training corpus. The Munich court rejected that framing outright. It held that where a model can be prompted to output a work in recognisable form, the work has been memorised within the model’s parameters, and that this memorisation is itself an act of reproduction under German law, independent of whatever happens later when a user extracts it. A few months later, Penguin Random House’s German publishing arm filed a similar suit, alleging that ChatGPT could reproduce Ingo Siegner’s children’s book series in text and illustrations “virtually indistinguishable” from the original, complete with a fabricated cover.

Placed side by side, the Delhi and Munich rulings are not really in tension. They are answers to the same question on different facts. Both courts used strikingly similar vocabulary, invoking memorisation and regurgitation directly, yet reached opposite outcomes because the underlying evidence differed. GEMA could produce verbatim lyrics on demand. ANI could produce nothing comparable, and ChatGPT’s retrieval-based architecture, which pulls in live content at the time of generation rather than reciting anything stored from training, worked structurally against a memorisation finding.

The practical implication is this: arguing that AI training is impermissible in principle is losing force as a legal strategy. What is replacing it is a narrower, evidence-driven question. Can a claimant show, through systematic prompting, that a model can reconstruct a specific work in recognisable form? For rights holders, that question is now the whole contest. For AI developers, the older defence, that a model merely learns statistical patterns rather than storing text, is no longer a complete answer wherever memorisation can be demonstrated. The Delhi ruling did not close that door in India. It simply found that, for now, no one had managed to open it.
Shradhanjali Sarma is the Founding Partner of Sakura Law Chambers and Shatakshi Shekhar is the Lead, Product Policy and Government Affairs at Sakura Law Chambers.