AI Training Gets A ‘fair Dealing’ Boost In India — With A Canadian Assist

Key takeaways

• The Delhi High Court ruled that temporary storage of copyrighted content for AI training may qualify as fair dealing, but outputs must be evaluated separately for infringement.
• The decision highlights the need for AI developers to document content ingestion practices clearly.
• While the ruling is significant, it is only a preliminary injunction, leaving several legal questions about copyright and AI usage open for future trials.

AI developers have been waiting for courts to draw a workable line between learning from copyrighted content and reproducing it. The Delhi High Court has now offered a preliminary answer.

In ANI Media Pvt. Ltd. v. Open AI OPCO LLC, the Court declined to grant an interim injunction over the training and outputs of the large language models underlying ChatGPT. Its prima facie conclusion was that temporary storage of news articles for training may qualify as fair dealing for private or personal use, including research, while outputs must be assessed separately for substantial reproduction. Put less judicially, the model may read what’s in a library, but it may not hand out photocopies.

The decision in brief

ANI alleged that OpenAI infringed copyright by storing its news content for model training, by the process of training and by generating responses that reproduced its works. ANI also argued that the Delhi court should hear the case, despite OpenAI’s servers being located in the United States.

On the jurisdictional question, the Court agreed the proceeding could continue in Delhi for several reasons, including that ANI is based there, that OpenAI actively serves Indian users and that relevant prompts and outputs occur within the Court’s territorial jurisdiction.

On the merits, the Court found, at this preliminary stage, that training-time storage of copyrighted works may fall within the “fair dealing” p of India’s Copyright Act, 1957 and that the challenged outputs were not substantially similar to ANI’s articles.

Location, in other words, did not make the claim impossible — but neither did ingestion of copyrighted works alone constitute infringement.

Canada mattered

The Delhi High Court recited and relied upon the Supreme Court of Canada’s decisions in CCH Canadian Ltd. v. Law Society of Upper Canada and Society of Composers, Authors and Music Publishers of Canada v. Bell Canada, concluding that research should be interpreted broadly. The Court rejected the proposition that a commercial interest in the dealing alone defeats fair dealing as a user’s right. It distinguished cases involving commercial distribution of copyrighted works from internal model training that did not release the training corpus to the public.

The Delhi High Court assessed fairness by asking whether the use was confined to training, whether it caused competitive or economic harm and whether the technology served the public interest.

ANI had not demonstrated a loss of subscribers or advertising revenue. Drawing from U.S. jurisprudence, the Court further noted that the trained models had “multifarious functions”, whereas ANI’s business is news reporting. Additionally, the Court noted that ANI had offered OpenAI a US$7.5-million license — a fact the Court viewed as a concession that if there were damages, they are economically quantifiable.

Inputs and outputs are different questions

The most important part of the judgment for AI businesses and content owners is the Court’s refusal to collapse training and output into one infringement analysis. For outputs, the Court compared ANI’s works “as a whole” rather than assembling a claim from selected fragments.

Adversarial prompts asking ChatGPT to reproduce content “exactly” did not produce responses that the Court considered substantially similar to ANI’s articles. The Court distinguished the German GEMA v. Open AI decision, where ordinary prompts produced verbatim song lyrics, and also found that the cited ChatGPT responses relied on retrieval-augmented generation (RAG) using live web sources rather than memorized training data.

The latter distinction was legally relevant because ANI’s examples of alleged infringement could not have been memorized during training — the models had been trained before the articles existed. This illustrates how technical architecture may have a critical impact in copyright analysis. Outputs generated solely from a model’s learned parameters may raise questions regarding whether training materials were memorized and reproduced, while RAG-based reproduction, by contrast, may resemble an “AI-enabled search engine” that retrieves and summarizes publicly available content from external sources in real time.

The Court noted that RAG had not been pleaded in ANI’s complaint and that whether RAG outputs might independently infringe remains an open question. These architectural distinctions reinforce the Delhi High Court’s broader approach: courts should assess infringement claims in respect of the ingestion of works as part of model training under the applicable “fair dealing” exception and separately assess outputs for substantial reproduction. Reading and copying are not the same act.

What the ruling means — and what it does not

The decision is important persuasive authority from a fellow common law, “fair dealing” jurisdiction, particularly because the Delhi High Court adopted and applied the Canadian conception of fair dealing as a user’s right deserving a broad and liberal interpretation. This was, however, only an interlocutory ruling on an interim injunction, not a final determination after trial.

Further, the Delhi High Court did not address whether “vectorization” (converting data like words or images into lists of numbers — vectors — so a computer system can derive meaning) itself amounts to actionable reproduction or whether retrieval-augmented outputs may independently infringe copyright.

No Canadian court has yet applied the “fair dealing” analysis to LLM training, so this decision may be persuasive. However, while the Delhi High Court borrowed Canada’s interpretive method, Canadian courts may yet differ in their application.

For AI developers and deployers: get inputs right

Companies that acquire, scrape, license, index or otherwise ingest content for AI systems may wish to view this decision as a prompt to consider strengthening governance. It may be prudent to consider documenting the purpose and limits of ingestion, preserving evidence of safeguards, testing outputs for substantial reproduction, separating training data from live retrieval sources and integrating these steps into a cross-border strategy.

The decision also illustrates why system architecture and contracting can be legally consequential. Different participants may be responsible for different parts of the AI supply chain. A model developer may train the foundation model; an application developer may introduce additional content through fine-tuning or RAG; and an enterprise customer may be responsible for connecting its system to proprietary or third-party sources in its deployment process. Companies may wish to ensure that their commercial arrangements reflect those respective roles and allocate responsibility accordingly.

Allocation of responsibility, however, does not provide a full answer to whether rights to the underlying content should be obtained. A “fair dealing” position may be strengthened by the record supporting it, and it may not provide a complete solution to mitigating risk. Fair dealing is fact- and jurisdiction-specific, whereas obtaining a licence where practical may provide greater predictability across markets and broader use rights; both fair dealing and licensing may serve complementary purposes.