YIP Note
Download
Two courts have now given early answers to the question every technology client asks: may an AI model be trained on someone else's copyright work? The answers point in different directions. The facts are different too, and that is the useful part.
Delhi: ANI v. OpenAI (interim stage)
As reported, in July 2026 a single judge of the Delhi High Court refused ANI an interim injunction. The prima facie view was that using news articles to train a language model is fair dealing under Section 52 of the Copyright Act, 1957, and that ANI had not shown the model reproducing its articles in its answers. The court also reportedly held it could hear the case although training happened on servers outside India. ANI has appealed to a Division Bench.
United States: Thomson Reuters v. ROSS Intelligence (appeal)
As reported, at the end of September 2026 the US Court of Appeals for the Third Circuit upheld a 2025 ruling that ROSS infringed Westlaw's headnotes, and could not rely on fair use, when it used material derived from them to train a competing legal-research tool. Reports describe it as the first US appellate ruling on fair use in AI training, and say the court stressed that ROSS was building a product to compete directly with the source.
Reading them together
- Competition with the source matters. The US court's refusal turned on a tool built to replace the very product it learnt from. The Delhi order dealt with a general-purpose model and found no copied output on the record.
- Evidence of copying in the output. In Delhi, the absence of proof that the model regurgitates the plaintiff's works weighed heavily at the interim stage. Build that evidence early.
- Stage of the case. The Delhi view is prima facie and under appeal. The US decision is an appellate ruling on summary judgment, and it is persuasive only in India.
- Jurisdiction. The Delhi court's reported view, that access from India is enough even when training happens abroad, will matter in every cross-border AI dispute.