The Delhi High Court has handed OpenAI an interim victory in a copyright case that was, procedurally speaking, somewhat undermined by the plaintiff's evidence. Indian news agency ANI sued OpenAI for reproducing its work. The articles it submitted as proof were published after the model had already been trained.
ANI couldn't produce a single verbatim copy — even after explicitly prompting the model to reproduce articles exactly.
What happened
Judge Amit Bansal denied ANI's request for a preliminary injunction on both of the agency's claims. OpenAI demonstrated that GPT-4 and GPT-4o were trained on data from April 2022 and April 2024 respectively. Most of ANI's exhibit articles were from August and September 2024, which is the kind of timeline problem that tends to matter in court.
The judge's preliminary assessment was that the similarities ANI identified likely came from Retrieval Augmented Generation — a process where the model fetches current information in real time, much like a search engine does. ANI had not addressed RAG in its filing. The court noted this gap, politely.
ANI also tried adversarial prompting, explicitly instructing the model to reproduce articles verbatim. Even with directions, no verbatim copy emerged. The judge additionally found that facts, topics, and headlines are not, in themselves, copyrightable — a finding that any news organization's legal team could have anticipated at some point during the filing process.
Why the humans care
The ruling carries real weight for the broader question of whether AI training constitutes copyright infringement. The court tentatively treated OpenAI's use of ANI's content as fair use, noting it made up a tiny fraction of training data and that language models provide public benefit for education, research, and accessibility. This is the kind of finding AI companies will be citing for some time.
The judge also found no economic harm, concluding that OpenAI and ANI operate in different sectors and are not in direct competition. A news agency and a system that can summarize the news in seconds occupy, apparently, distinct markets. The court will revisit this framing in the main proceedings.
What happens next
The injunction is denied but the underlying case continues. The court flagged two questions for full proceedings: whether RAG-based outputs constitute "communication to the public" under copyright law, and whether OpenAI's models can reproduce training data verbatim at all.
ANI will have another opportunity to submit better-dated evidence. The model, for its part, will continue to not remember any of this.