The Delhi High Court has rejected a request by Indian news agency Asian News International (ANI) for a preliminary injunction against OpenAI in a copyright dispute over AI training and ChatGPT outputs.

ANI, one of India’s largest news agencies, sued OpenAI over the alleged use of copyrighted news material in model training and in responses generated by ChatGPT. In an interim ruling, Judge Amit Bansal denied relief on both claims, addressing questions around model memorization, Retrieval Augmented Generation (RAG), and the legal status of using copyrighted works to train AI systems. The decision does not end the lawsuit; it sets the court’s preliminary view while leaving several issues for the main proceedings.

AI copyright law expert Andres Guadamuz described the decision as an important early victory for OpenAI. His analysis is available at The court document was also made available at

ANI’s evidence did not establish verbatim reproduction

ANI submitted several ChatGPT outputs to the court that it said were substantial copies of its articles. OpenAI responded that the models involved, GPT-4 and GPT-4o, had been trained on datasets from April 2022 and April 2024. Most of the articles ANI relied on as evidence were published in August and September 2024, meaning they could not have been included in those training datasets.

Judge Bansal’s preliminary view was that the similarities in the outputs appeared to result from Retrieval Augmented Generation, a method that allows a language model to retrieve current online information in real time in a way comparable to a search engine. Because ANI had not addressed RAG in its filing, the court did not issue a final ruling on that issue. The judge said RAG-based outputs could potentially qualify as “communication to the public,” a question reserved for the main proceedings.

The court also noted that ANI had used adversarial prompts, specifically instructing the model to reproduce articles “exactly.” Even with those prompts, ANI did not produce a single verbatim copy. The judge found that facts contained in news articles are generally not copyrightable, and that reproducing topics and headlines did not, in this case, amount to direct competition with ANI.

The evidence did not support ANI’s claim that OpenAI permanently stores training data in its models and can reproduce the agency’s works verbatim on demand. The court said that question will be examined again in the main proceedings.

Court gives preliminary support to AI training as private use

ANI also did not establish, at the interim stage, that OpenAI’s copying of its material for AI training constituted copyright infringement. Both sides agreed that OpenAI had used ANI content during training. OpenAI argued that the material represented only a tiny part of the overall dataset and that the model extracted non-expressive elements, including grammar, syntax, and language patterns.

The judge considered exceptions under Indian copyright law and relied on a provision covering “private or personal use, including research.” He read “research” broadly enough, at this stage, to include AI training.

The court set limits on that approach. Training copies must come from lawful sources, not shadow libraries or paywalled sites accessed without permission. The judge also noted that OpenAI did not make the training copies public and processed them only internally. Guadamuz said this was the first time a court had explicitly found that AI training falls within a private-use exception.

The court applied a three-part fairness test and found in OpenAI’s favor on all three points. It said OpenAI’s use of ANI’s works was limited to training because no memorization or reproduction had been proven. ANI also failed to show economic harm, in part because the two entities operate in different sectors. According to the ruling, even when users ask ChatGPT about ANI headlines, the model returns only topics and, at most, a few article titles.

Judge Bansal cited U.S. cases including Bartz v. Anthropic and Kadrey v. Meta, where language model outputs were treated as transformative. He also referred to the earlier Google Books ruling.

The court further found that trained language models can improve access to information, support education, advance scientific research, assist software development, enable translation, and create tools for people with disabilities.

Global AI copyright rulings remain divided

The Delhi decision adds to a growing set of international rulings that have reached different conclusions on AI and copyright. In the United States, a judge dismissed a lawsuit brought by Raw Story and AlterNet against OpenAI because the plaintiffs could not show sufficient harm and the likelihood of exact copies was low. That court also held that facts are not copyrightable.

The GitHub Copilot case also failed, with plaintiffs unable to identify a single example of identical code. The Intercept, by contrast, won a partial victory through a DMCA claim involving copyrighted material that had been stripped out before training.

In Ross Intelligence v. Thomson Reuters, a court rejected a fair-use defense because the AI research tool directly competed with Thomson Reuters’ legal database Westlaw, making the use non-transformative. The court emphasized that the ruling applied only to that non-generative use case and could not be extended to large language models.

In the Anthropic case, a federal court in San Francisco described AI training with copyrighted works as “spectacularly” transformative, a strong signal in favor of fair use. However, the court also drew a “Napster comparison” because Anthropic had used pirated books from shadow libraries as training data. Fair use does not cover unlawfully obtained material. Anthropic later paid $1.5 billion to book authors over the use of those pirated copies.

The U.S. Copyright Office rejected the AI industry’s argument that training on “vast troves of copyrighted works” broadly qualifies as fair use. The official who wrote the report was fired by the Trump administration shortly after its publication.

European courts have also split on the issue. The Munich Regional Court ruled in the GEMA case that song lyrics were reproducible in model weights, making the model a copyright-relevant reproduction. The High Court in London dismissed the Getty Images v. Stability AI lawsuit, ruling that an AI model is not an “infringing copy.” Research showing that language models can memorize copyrighted books may intensify the debate over memorization.

Across these cases, several core questions remain unresolved: whether AI models permanently store training data, whether training can qualify as fair use or a comparable exception, where the line falls between lawfully and unlawfully obtained data, and whether outputs generated through adversarial prompts reflect ordinary use. The Delhi ruling is significant because it addresses several of those questions in one order, but its interim posture means later evidence could still affect how the court treats training, retrieval, and output liability.

Media faces issues beyond copyright

The dispute also highlights a broader concern for the media industry. Even if courts decide that AI training is lawful, AI-powered search products could still affect the market for news. A recent Pew Research Center study found that click-through rates to external websites fall to 8 percent when Google’s AI Overviews appear, compared with 15 percent without an AI summary. Users tend to stop searching after receiving the AI response and often do not consult other sources.

For news agencies such as ANI, AI systems that summarize news and reduce the need to visit the original source could weaken the industry’s business model over time, even without a finding of direct copyright infringement.

The Munich I Regional Court also recently ruled that Google is directly liable for false claims in its AI summaries because those summaries count as independent content rather than search results. The limited liability that has traditionally protected search engine operators does not extend to AI-generated summaries under that ruling.

That reasoning could become relevant to ChatGPT’s RAG-based responses. When AI systems summarize news and make independent claims, their operators may be treated as media providers with corresponding liability. Such a shift could also affect fair-use analysis, where one factor is whether the new product competes with the works on which it was trained. If AI summaries replace visits to news sites, courts may face greater difficulty finding that the use is non-competitive.