ANI vs OpenAI: Delhi HC Ruling on AI Training Data in India

Quick answer: On 24 July 2026, the Delhi High Court refused ANI an interim injunction against OpenAI, holding that storing news articles to train the large language models behind ChatGPT is prima facie “fair dealing” under Section 52(1)(a) of the Copyright Act, 1957. It is an interim order in an ongoing suit, not a final verdict on AI training data in India.

Key takeaways

• On 24 July 2026, Justice Amit Bansal of the Delhi High Court declined ANI’s plea for an interim injunction against OpenAI (LiveLaw, ThePrint).

• Storing ANI’s articles to train the models behind ChatGPT was held to fall prima facie under Section 52(1)(a) of the Copyright Act, 1957, as private use including research (Verdictum).

• ANI also failed to show that ChatGPT’s outputs were substantially similar to its articles, or that memorisation or market harm occurred (Outlook Business).

• These are prima facie findings on an interim application. The suit, filed in November 2024, continues.

• Separately, a DPIIT committee has proposed a statutory licence with royalties, so the law could change regardless of how the litigation ends.

What the Delhi High Court actually decided

ANI Media Pvt Ltd sued OpenAI in November 2024, the first Indian media organisation to bring a copyright claim against the company (EU IP Helpdesk). The suit, CS(COMM) 1028/2024, raised two complaints: that OpenAI stored ANI’s articles to train the models behind ChatGPT, and that ChatGPT reproduced that material in its answers. Reporting across Business Standard, ThePrint and LiveLaw agrees on the substance of the 24 July order.

On storing data for training

The reasoning turned on two ideas. First, the data absorbed by a large language model is accessible to the model itself and is not published to any human reader, which the court treated as private use. Second, it characterised training — screening and organising stored text, extracting from it, converting it into machine-readable inputs — as a form of research that generates new knowledge.

This matters because India has no dedicated text and data mining exception of the kind found in the EU Copyright Directive or Japanese copyright law. Indian AI developers have had to argue under a fair dealing provision written long before machine learning existed, and this is the first substantial judicial signal on that argument.

On what ChatGPT produces

The second limb failed on evidence rather than principle. The court was not persuaded that ChatGPT’s outputs were substantially similar to ANI’s articles, and ANI did not establish memorisation or harm to the market for its journalism. The burden sat with the publisher to show the model reproduces its work, not merely that it was ingested.

What the ruling does not decide

The headline version of this story overstates it. This is an interim order, and the court framed its conclusions as prima facie — a provisional view formed only to decide whether to restrain a party while the case runs. The suit has not been tried.

Nor does it settle anything for other claimants. The Federation of Indian Publishers, the Digital News Publishers Association and the Indian Music Industry intervened in the proceedings. A different record, particularly one with solid evidence of memorisation, could produce a different result.

How this sits alongside the rest of the world

India now fits a global pattern in which courts separate the act of training from the act of reproducing. In Germany, the Munich Regional Court held in GEMA v OpenAI in November 2025 that the EU mining exception generally covers training, but that a model which actually reproduces complete training data falls outside it (Norton Rose Fulbright). The Court of Justice of the European Union heard a Grand Chamber case on generative AI and copyright in March 2026, with the ruling still awaited (Stephenson Harwood). In the United States, the first fair use rulings on AI training arrived in 2025, broadly treating training as transformative while splitting sharply on market harm (Norton Rose Fulbright).

The policy question India has not settled

Litigation is only half the picture. A committee constituted by the Department for Promotion of Industry and Internal Trade published a working paper on generative AI and copyright in December 2025 (DPIIT working paper). Rather than a broad mining exception, its majority view proposed a mandatory blanket licence for AI training with statutory royalties routed through a new collective body. NASSCOM dissented from within the committee, arguing for a purpose-neutral mining exception with a machine-readable opt-out (Ikigai Law). A favourable reading of Section 52 therefore does not close the question: Parliament can amend the Act.

What this means for data and AI work in India

Data provenance becomes a documented artefact. If legal exposure sits at the output layer, teams need to know what went into a system and be able to show it. Dataset cards, licence records and ingestion logs move from good practice to expected practice.

Evaluation work gains weight. Testing whether a model regurgitates source text, and building guardrails when it does, is now a business need rather than a research curiosity — ordinary applied machine learning work, and adjacent ground to our note on Generative AI skills every data analyst needs.

Licensing becomes a data pipeline problem. If India moves towards royalty collection, someone has to track usage, attribute it and reconcile payments — data engineering and analytics work. None of it changes the answer to the question students ask most often, which we cover in Will AI replace data analysts?

The Nagpur and Vidarbha angle

To be straightforward: we found no verifiable Nagpur-level news item connected to this ruling, and we are not going to manufacture one. What follows is analysis, not reported local news.

Nagpur’s relevance is structural. The city hosts development centres for large IT services firms in the MIHAN area, and services firms are precisely the organisations clients will ask to demonstrate clean data practices in AI projects. At the state level, the Maharashtra AI Policy announced in 2026 includes an agreement under which AI centres are to be set up in Mumbai, Pune and Nagpur, with training for government employees and students (Akashvani News).

The useful read for a learner is narrow: compliance-aware AI work reaches Tier-2 cities through services delivery, not only through product companies in Bengaluru. Being able to explain in an interview where your project’s data came from, and why you were entitled to use it, reads as professional maturity. See also our note on why Nagpur students are choosing data and AI careers and our programme list.

The bottom line

The Delhi High Court has given Indian AI developers their first meaningful judicial comfort on training data, by reading a decades-old fair dealing provision generously. In the same breath it told publishers what evidence would actually win: proof that a model reproduces their work or damages their market. Neither side has a final answer, and the legislative track may overtake the judicial one.

FAQ

Frequently Asked Questions

Did the Delhi High Court say AI companies can freely use any copyrighted content in India?

No. The order decides one interim application in one suit. On the material before it, the court held that OpenAI storing ANI’s articles to train its models was prima facie covered by fair dealing — a provisional finding, not a blanket permission.

What is Section 52(1)(a) of the Copyright Act, 1957?

It is a fair dealing exception allowing certain uses of a work without the owner’s permission, including private or personal use and research. The court read training data held inside a model, never published to a human reader, as fitting that description.

Can ANI still challenge this outcome?

The main suit continues, and the court framed its findings as prima facie, so they do not settle the dispute. An interim order of a single judge is also ordinarily open to challenge within the High Court.

Is Indian law on AI training data now settled?

No. A DPIIT committee working paper published in December 2025 proposed a statutory licensing model with royalties for rights holders rather than a broad mining exception, and that track runs separately from this litigation. Parliament can change the position by amending the Copyright Act.

Does this ruling change what a data science student in India should learn?

It does not change the core syllabus, but it raises the value of skills around data provenance, licensing and output evaluation alongside modelling. Teams building on large language models need people who can document where data came from and test what the system reproduces.