Has “private use” become an AI safe harbour?

The Delhi High Court’s decision in ANI Media Private Limited v. OpenAI OpCo LLC may prove to be one of the most consequential interim judgments related to Generative AI era in India. This is not because it has settled the legality of AI training, but it has, at the interim stage, placed commercial Large Language Model (LLM) training within the architecture of Section 52(1)(a) of the Copyright Act, 1957 fair dealing exception.

In dismissing the application for an interim injunction against OpenAI, the Court prima facie held that training the LLM falls under the fair dealing exception of ‘personal and private use, including research’ and, therefore, does not amount to infringement under Section 51 of the Act.

The judgment is expressly interim, but interim copyright rulings have a tendency to transform into settled law. Unlike other interim judgements, this one potentially expands the interpretation of Section 52(1)(a) in a way that strains the text, inverts the legislative history, sidelines Indian precedents and quietly imports the American fair use doctrine, which the same judgment concedes has no ‘persuasive value in Indian context’.

The judgment expands the definition of ‘private’ in ‘private and personal use’ to include corporations because dictionaries define the term ‘private’ as restricted to ‘a person, group or class’. Since OpenAI holds ANI works ‘in a closed space without access to the public’, the Court deemed it private.

This turns the accepted rule on its head. The amicus opined that the words “private” and “personal” are overlapping and they draw meaning from each other. Therefore, the interpretative rule of noscitur a sociis would require that the meaning of “private” and “personal” be interpreted in an analogous and cognate sense. The Court’s response, that a cognate reading renders the term ‘private’ otiose, is open to question.

“Personal use" simply means using something for yourself, while "private use" means using it in your own private space. The law makes sense without having to stretch the word "private" to justify massive, industrial-scale data copying by corporations.

The history of law proves this point. In 1994, lawmakers changed the phrase ‘private study’ to ‘private use’ to protect everyday people from being harassed over casual copying. In 2012, ‘personal’ was added to double down on protecting individuals. The legislative intent was to strengthen the individual use dimension rather than establishing a safe harbour for industrial scale copying.

Making an unauthorised copy, including digital storage, is illegal under copyright law, regardless of whether it is ever shared with the public. As per the current ruling, as long as corporations are keeping the copies, albeit obtained without consent, private/internal, it is deemed kosher and well within the realm of the law. Effectively, the Court has erased the rule against copying, and companies can now copy whatever they want, provided they are careful not to share the data with anyone else.

The Court ruled that the process of training an LLM counts as research. It applied the ‘doctrine of updating construction’ that updates old definitions for modern times and further reasoned that confining research to humans would be a regressive view.

However, updating the definition only applies to new ways of doing the same thing, like moving from handwriting to typing. It is not meant to cover a completely different activity with different beneficiary. As per copyright law, ‘research’ means a person studying a text to learn something. Even precedents cited by the Court define ‘research’ as human being studying to add to human knowledge.

Furthermore, calling AI training process ‘research’ defeats the Court’s observation about ‘private use’. As per the Court’s accepted definition, the purpose of research is to share the results with the public, whereas ‘private use’ means keeping things restricted. OpenAI’s process cannot claim both things at once. It cannot claim that data is used for ‘private use’ to qualify as ‘private use’ while simultaneously claiming it is research meant to benefit the world.

The Court also held that since lawmakers explicitly wrote the term ‘non-commercial’ elsewhere in section 52 (1) but not in clause (a), commercial use is no bar. While it may sound like a legal technicality, the phrase ‘private or personal use’ naturally excludes the corporation. A restaurant serving dinner to 10,000 customers wouldn't be called a ‘private dinner.’ The words themselves set the limit.

In allowing this, the Court deviated from settled precedents like Jiwan Publishing House and TIPS v. Wynk, wherein the courts have ruled that any attempt to make a copy of books for profit cannot be termed as ‘research’ and any attempt to copy songs by a music app to make a profit cannot be permitted because its users are listening to the music privately. The ratio is that commercial intermediaries cannot put forward the end users’ research defence.

The Court ruled against these settled precedents citing that OpenAI is making copies for internal purposes. It is imperative to discern that OpenAI is doing the same thing as the companies in aforementioned cases were doing - they were making unauthorised copies of the protected work for a large user base without consent to make monetary gains. The copies made by OpenAI are in digital AI data instead of the conventional copying of books or streaming of songs. A mere change of technology does not change the basic rules of copyright law.

The Court held that since ANI’s content was freely available on its website and not protected by any paywall, copies made by OpenAI were not violative of the Act. The conflation of accessibility with authorisation runs through the entire judgment. Publication of an article on the web is an invitation to read the article and an exercise of communication rights; by no stretch is it a license to reproduce. The protection under Section 14 of the Act does not hinge on the erection of any technical barrier. Section 51 of the Act does not allow ‘it was there for taking’ as a defence.

Further, the onus was shifted onto ANI, with the Court stating that ANI should have used code to block OpenAI’s web crawler if they didn’t want their data taken. The Court has imported the European concept, that creators must actively ‘opt out’ of being copied, a concept that does not exist in India. Indian copyright jurisprudence has remained simple - ‘opt in’, meaning a company must ask your permission before taking your work. By flipping this rule backward, the Court is forcing millions of creators to build a digital wall to stop tech companies from taking their content.

The Court observed that OpenAI taking this data for free did not financially harm ANI. However, the Court ignored that OpenAI is already paying millions of dollars to other news publishers for the exact same right to train its AI. Under standard copyright law, if a market exists where people buy licenses, taking the work for free is the exact definition of financial harm. Instead of recognising this, the Court weaponised ANI’s license offer against them and refused to stop the copying because ANI was negotiating a deal with OpenAI in exchange for the usage of news content. Effectively, if a creator tries to negotiate a price with a tech company, that attempt can be used against them to deny protection.

The consequences compound. Doctrinally, Section 52(1)(a) of the Act was drafted as a protection for the individual reader and researcher. Now, it functions as a general corporate exemption with no limit on commerciality, provenance, or quantum. Institutionally, the Court has assumed the role of Parliament and DPIIT’s own expert on AI and copyright by applying ‘doctrine of updating construction’, a policy question that the European Union, Japan and Singapore resolved through legislation. Economically, it creates a two-tier world in which OpenAI pays western news publishers under license while taking Indian content for free. It results in devaluing the data assets of Indian news publishers and the AI ecosystem which the judgment claims to protect. It also accelerates the substitution of AI-based answer engines for the traffic on which Indian journalism survives.

To be fair, the Court did make some valid points. ANI’s examples of ChatGPT copying their articles weren’t perfect and there is a genuine public interest in developing artificial intelligence. Though this is only a temporary and early-stage ruling, the Court could have reached the exact same temporary decision without upending the foundation of copyright law to get there.

The question for the appellate court and ultimately perhaps Parliament is whether this is interpretation or legislation by another name.

Chaitanya Kumar Verma is an in-house counsel to a US-headquartered venture capital fund investing in Indian consumer businesses.