Digital news faces existential issues due to LLMs like ChatGPT

In a recent development in the publishers versus tech tussle, the Digital News Publishers Association (DNPA) appeared before the Delhi High Court on August 5 to argue that OpenAI was infringing the copyright rights of Indian media outlets through unauthorised use of news content in its AI models.

During the court hearing, Senior Advocate Raj Shekhar Rao, who was representing DNPA, made the following submissions:

Intent of Copyright Act: Rao commenced his submissions by highlighting that the Copyright Act is a comprehensive legislation intended to consolidate and codify the law on copyright in India. The statutory framework sets out the complete scheme governing the rights of copyright owners and the permissible uses of protected works. Accordingly, any limitations or exceptions to copyright protection, such as those set out in Section 52, must be construed strictly.

Distinction between public availability and public domain: Rao submitted that there is a distinction between publicly accessible content and content that is in the public domain. Merely because something is accessible does not mean it is not protected by copyright. He submitted that the act of scraping, storing, and reproducing the copyrighted content of news publishers for commercial purposes, without authorisation, constitutes infringement under Indian copyright law. He submitted that digital news faces existential issues due to LLMs like ChatGPT.

Distinction between ‘infringing copy’ and ‘lawfully acquired copy’: Rao submitted that the term ‘infringing copy’ is different from ‘lawfully acquired’ and ‘lawful access’, which are used in copyright laws abroad. He highlighted that the Explanation to Section 52(1)(a) of the Copyright Act excludes the storage of “infringing copies” from the purview of fair dealing. . Rao cited various provisions of the Copyright Act to demonstrate that the legislative intent is further reflected in the deliberate use of distinct terms throughout the Copyright Act, underscoring that any use falling outside the purposes expressly recognised is to be treated as an infringement of the copyright owner’s rights.

Scraping would amount to reproduction and therefore infringement: He also submitted that the act of scraping or ingesting content from a website results in the reproduction of the work under Section 14 of the Copyright Act, i.e the creation of a distinct “copy” of the original. In the absence of a licence permitting such scraping, this reproduction would constitute the creation of an “infringing copy” under copyright law.Reproduction does not require exact copying, and partial copying is sufficient to attract liability.

Regurgitation is an admitted fact and is a result of memorisation: Rao highlighted that LLM models regurgitate from their training data. This is an admitted fact (albeit presented as a glitch) is an acknowledgement that reproduction/memorisation occurs. Rao also referred to Infopaq International A/S v. Danske Dagblades Forening, which held that storing an extract of a protected work comprising just 11 words and printing out that extract would be considered reproduction. The admitted fact of regurgitation and the form and format of the output itself is proof of the fact that the Defendant does store a copy of work it is trained on.

Tokenisation and Vectorisation as Reproduction: Rao submitted that technical processes such as tokenisation and vectorisation retain the essential expression of the work. Tokenisation breaks content into syntactic units, and vectorisation maps them semantically. He referred to OpenAI’s own disclosures and stated that some level of human review/intervention is essential for model outputs of such LLMs, thereby establishing a link to the original input, because reproduction/regurgitation warrants the fact that there is memorisation. The model stores a statistical representation of protected expression.

Reproduction/Storing the copyrighted works, even at the intermediate step, amounts to infringement: Rao referred to the case MySpace Inc. v. Super Cassettes Industries Ltd., to submit that transient or incidental storage amounts to infringement, unless it is covered by the fair use exception.

Nature of output: Rao further submitted that since the output of ChatGPT relies on the works of the plaintiff and the intervenors, such as DNPA, the output amounts to derivative work, which is an exclusive right of the copyright owner. The output generated using the expression of news is not transformative, and there is no modicum of creativity. He relied on The Chancellor, Masters & Scholars of the University of Oxford & Ors. vs. Rameshwari Photocopy Services & Ors to submit that since ChatGPT is used for commercial purposes, it does not fit in the exceptions under section 52 of the Act. Therefore, use of the works for training of the said software cannot be considered fair use. He also relied on Disney Enterprises, Inc. LLC v. VidAngel, Inc to highlight that Copyright infringement would still arise where the use is for commercial purposes. To conclude otherwise would amount to endorsing a loophole in the law.

ANI’s copyright suit against OpenAI, filed in Nov 2024 in the Delhi High Court, has drawn major media and music industry bodies as co-interveners. The case, a first-of-its-kind in India, will decide if AI training on copyrighted news content amounts to infringement or fair use.