Delhi court rejects ANI's copyright case against OpenAI, backs AI training as fair use

The Delhi ruling arrives amid a patchwork of global court decisions on AI copyright. Some courts have sided with AI developers on fair-use grounds—the U.S. courts dismissed lawsuits from Raw Story, AlterNet, and GitHub Copilot's challengers, citing low odds of exact reproduction and lack of direct competition. A San Francisco federal court called Anthropic's AI training "spectacularly" transformative. Others have ruled against AI companies: the Munich Regional Court found that song lyrics are reproducible in model weights, and a court denied fair use in Ross Intelligence v. Thomson Reuters because the AI tool directly competed with Thomson Reuters' Westlaw database. The U.S. Copyright Office has rejected the AI industry's broad interpretation of fair use for training on "vast troves of copyrighted works," and Anthropic faced reputational and legal consequences for using pirated books, later paying $1.5 billion（約2400億円） to authors. Judge Bansal's decision to treat AI training as private research use is a notable first, grounding it in conditions: lawful sourcing and internal processing only.

However, the ruling does not resolve the question of retrieval-augmented generation (RAG)—the real-time retrieval of information that ChatGPT uses to supplement its base knowledge. ANI did not address RAG in its filing, leaving the court unable to make a final ruling. This gap may prove consequential. A recent Munich court decision held Google directly liable for false claims in its AI summaries, treating them as independent content rather than search results, which exposes AI providers to media-provider liability. If RAG outputs are deemed "communication to the public," OpenAI could face similar liability, and courts may reconsider the fairness analysis: if AI summaries replace the need to visit original news sites, the use ceases to be non-competitive.