The Delhi High Court rejected a copyright injunction sought by Indian news agency ANI against OpenAI, ruling that AI training constitutes private use. This marks the first time a court has classified AI training as private use, dealing a blow to ANI's case. The main trial remains pending.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
The Delhi High Court rejected Asian News International's (ANI) preliminary injunction request against OpenAI, ruling that AI training may fall under the 'private use' exception and that ANI failed to prove ChatGPT reproduced its articles verbatim. However, the court reserved final judgment on core issues such as RAG output and permanent storage.
ANI submitted multiple ChatGPT outputs to the court, claiming they extensively copied ANI's articles. However, OpenAI proved that the models used, GPT-4 and GPT-4o, had training data cutoffs of April 2022 and April 2024, respectively, while the articles cited by ANI were mostly published in August and September 2024, making it impossible for them to be included in the training data.
The judge preliminarily attributed the output similarity to RAG (retrieval-augmented generation), where the model retrieves information from the internet in real time, similar to a search engine. ANI did not address the RAG issue in its complaint, so the court could not make a final ruling on it, but noted that RAG output may constitute 'communication to the public,' a matter to be decided at trial.
ANI used adversarial prompts explicitly asking the model to 'exactly' reproduce articles, yet even then, no verbatim copies were produced. The judge noted that facts in news are generally not copyrightable, and reproducing themes and headlines did not constitute direct competition with ANI in this case.
Both parties acknowledged that OpenAI used ANI content in training, but OpenAI argued that the material constituted a negligible portion of the overall dataset and that the model only extracted non-expressive elements (e.g., grammar, syntax, and language patterns).
The judge relied on the 'private or personal use, including research' provision in Indian copyright law, broadly interpreting 'research' to encompass AI training. This is the first time a court has explicitly included AI training under the private use exception. However, the exception is conditional: training copies must come from lawful sources (not shadow libraries or paywalled sites), and OpenAI never publicly disclosed training copies, only processing them internally.
The court applied a three-factor fairness test, all favoring OpenAI: OpenAI's use of ANI works was limited to training (no proven memorization or reproduction); ANI failed to prove economic loss (the parties operate in different industries); even when ChatGPT was asked about ANI headlines, the model only returned themes and at most a few article titles.
The judge cited U.S. cases (e.g., Bartz v. Anthropic, Kadrey v. Meta) and the Google Books case, finding that language model outputs are transformative.
The Delhi ruling joins a series of conflicting international precedents. In the U.S., Raw Story and AlterNet v. OpenAI was dismissed for lack of harm; GitHub Copilot case failed because plaintiffs could not provide identical code examples; The Intercept achieved partial victory via a DMCA complaint.
Conversely, in Ross Intelligence v. Thomson Reuters, the court rejected fair use because the AI research tool directly competed with Westlaw. In the Anthropic case, the court called AI training 'strikingly' transformative but analogized it to Napster due to the use of pirated books as training data, ultimately ordering Anthropic to pay authors $1.5 billion.
The U.S. Copyright Office rejected the AI industry's argument that training on 'massive amounts of copyrighted works' broadly constitutes fair use. In Europe, the Munich Regional Court in the GEMA case found that lyrics could be reproduced in model weights, constituting copyright-relevant reproduction; while the London High Court dismissed Getty Images v. Stability AI, ruling that AI models are not 'infringing copies.'
In all cases, core issues remain unresolved: Do AI models permanently store training data? Can training constitute fair use? What is the boundary between lawful and unlawful data acquisition? Do copies generated via adversarial prompts reflect normal use?
Even if courts rule AI training lawful, AI search products may still undermine the news market. A recent Pew Research Center study shows that when using Google AI Overviews, external website click-through rates drop to just 8%, compared to 15% without AI summaries. Users often stop searching after receiving an AI answer, without consulting other sources.
For news organizations like ANI, the trend of AI systems summarizing news and making it unnecessary to click on original sources could erode their business models.
This article is based on THE DECODER's report on the Delhi High Court ruling, which cited the court's decision and comments from AI copyright law expert Andres Guadamuz. The global case law section lists known cases mentioned in the report, and the Pew Research Center data is third-party research cited in the report. All facts come from a single source and have not been cross-verified.
The Delhi High Court's ruling gives OpenAI significant breathing room in India but does not resolve the fundamental divisions in AI copyright law. ANI's injunction was denied primarily due to evidentiary flaws (article dates after training data) and failure to prove economic loss. However, key issues such as the characterization of RAG output and permanent storage of training data remain for trial. Globally, courts remain divided on whether AI training constitutes fair use, and the practical impact of AI search on news click-through rates may pose a more urgent threat than copyright itself.
Primary report
Primary source