The Delhi High Court, presided over by Justice Amit Bansal, heard extensive arguments in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, a closely watched case raising foundational questions on copyright law in the context of artificial intelligence and large language models (LLMs).
Appearing for OpenAI OpCo LLC, Senior Advocate Amit Sibal advanced submissions on the core issues framed by the Court particularly the “training claim” and the “output claim” asserted by ANI Media Pvt. Ltd.. On behalf of ANI, arguments were initially addressed by Advocate Sidhant Kumar and Senior Advocate Chander Lall.
During the hearing, OpenAI contested the allegation that training LLMs involves copyright infringement. Responding to submissions made by interveners, OpenAI argued that the conversion of text into tokens, vectors, and model parameters does not amount to reproduction of copyrighted works. It was submitted that vectors merely encode semantic and statistical relationships between tokens reflecting rules of grammar, syntax, and contextual usage and do not reproduce the expressive elements of any individual work. According to OpenAI, training data is used to extract broad linguistic patterns rather than copying or storing protected expression.
On this basis, OpenAI contended that neither vectors nor model parameters can be characterised as “reproduction” under the Copyright Act, and therefore the training process does not infringe copyright at the threshold stage. It was further argued that even if the Court were to hold otherwise, the issue must be examined together with the doctrine of fair dealing, as the statute read in light of judicial precedents on the idea–expression dichotomy protects expression, not ideas such as grammatical rules or statistical patterns.
Addressing the output claim, OpenAI submitted that the instances cited by ANI do not disclose any material or substantial reproduction of copyrighted works. Reliance was placed on comparative jurisprudence, including foreign decisions on news aggregation and clipping services, to argue that ChatGPT’s responses when viewed as a whole add original commentary and do not amount to verbatim or cumulative reproduction. It was emphasised that facts and news events enjoy narrow copyright protection, and similarities arising from reporting the same event cannot, by themselves, establish infringement.
On the other hand, ANI argued that OpenAI’s defence improperly narrows the infringement inquiry to vectors, tokens and parameters, while ignoring the initial act of copying. According to ANI, the training process necessarily begins with the wholesale copying and storage of ANI’s copyrighted works from its website onto OpenAI’s systems without licence or authorisation. This act, ANI submitted, by itself completes infringement under Section 14(1)(a) read with Section 51 of the Act of the Copyright Act.
ANI contended that copyright infringement is a strict liability offence and that the purpose for which copying is undertaken whether for linguistic analysis, pattern recognition, or prediction has no relevance at the threshold stage. It was submitted that once unauthorised copying and storage in a material form is established, infringement stands, and any claim of fair dealing can arise only thereafter as a limited statutory defence.
ANI further argued that OpenAI’s characterisation of copying as “scraping” or “accessing” content masks the reality of unauthorised reproduction. Merely because ANI’s website is publicly accessible, it was submitted, does not imply consent to copy, store, or commercially exploit its content for training AI models.
After hearing detailed submissions, the Court indicated that the questions surrounding reproduction, fair dealing, and the application of copyright law to emerging AI technologies would require careful consideration. The matter has been listed for further hearing, with parties directed to place brief notes on specific statutory issues raised during the course of arguments.