OpenAI vs ANI case: what it means for the future of information

News agency Asian News International (ANI) sued AI research firm OpenAI in November 2024 for using its content to train ChatGPT without permission or payment. The generative-AI model allegedly hallucinated quotes and reports, wrongly attributing to the agency. The ANI, in its plea in Delhi High Court, has sought ₹2 crore in damages and an order barring OpenAI from using its work. On July 24, 2026, Justice Amit Bansal ruled that storing ANI’s articles to train ChatGPT falls under Section 52 of India’s Copyright Act, which exempts private use and research from infringement. The judgment noted that ANI failed to show that ChatGPT reproduced or “memorised” its actual reporting, hence there was no infringement. The 135-page order is the first detailed attempt by an Indian court to place AI training within the Copyright Act. It is an interim order, though, and ANI can appeal. The sticking point At the heart of these disputes lies fair use (in the U.S.) or fair dealing (in India and the UK). Wire services such as PTI, ANI, Reuters, AP, and AFP license verified facts to other outlets, which rewrite them in their own style. The business depends on readers trusting that the wire got the facts right, and on the wire being paid for the work of verifying them. A search engine sends a reader back to the original story but an LLM takes the fact, drops the sourcing and verification behind it, and returns an answer with no byline. AI companies argument Copyright holders counter Copying materials for training models is research and analysis, not a substitute for the originals. Training involves massive, systematic copying of entire works, often scraped without consent. Models don’t store or reproduce works in full; they learn patterns and statistics, then produce new outputs. AI systems can and do reproduce substantial parts of those works, sometimes verbatim. This transformation should qualify as fair use. This undermines the market for the originals and hurt creators’ ability to earn a living. Does training a model with pirated material come under fair use in US? In August 2025, Anthropic agreed to pay $1.5 billion to settle a class action brought by authors after the company downloaded more than seven million pirated books from shadow libraries such as Library Genesis to train its Claude models. In 2026, a federal judge gave the settlement final approval which was incidentally the largest copyright class-action payout in American history, at roughly $3,000 per book across nearly 500,000 titles. Judge William Alsup, who oversaw the case before it settled, had already ruled that training AI on legally acquired books was fair use. It was the piracy (that is downloading books from illegal sites rather than buying them) that got Anthropic into trouble. Two days after Alsup’s ruling, a California judge, Vince Chhabria, found that Meta’s use of books, including pirated ones, to train its Llama models was fair use on the facts before him. But he told the losing authors how they should have argued the case: most unauthorised AI training, he wrote, will probably prove illegal, because it risks “market dilution”: flooding the market with AI-generated books, music, and articles that compete with the human-made originals the AI learned from. It doesn’t matter if the output looks nothing like the input; if the effect is a machine that can produce infinite competing content in seconds, the market for the original work is destroyed regardless. The New York Times’ case against OpenAI and Microsoft, filed in December 2023, still grinds through discovery in Manhattan, with no trial date set. However, the court had ordered OpenAI to preserve and eventually hand over 20 million anonymised ChatGPT conversation logs, which a magistrate judge upheld earlier this year over OpenAI’s privacy objections. Whatever the court eventually rules on the core question will likely shape how every subsequent US news-industry case is argued, ANI’s included. Training a model with copyright content: intent and consent matters Getty filed a plea against Stability AI in High Court of England and Wales in January 2023 over its use of millions of Getty photographs, some with visible watermarks intact, to train the Stable Diffusion image generator. By the time the case reached judgment in November 2025, Getty had abandoned its central claim that training itself infringed copyright because it couldn’t prove the training happened on the UK soil since the servers sit elsewhere. However, the High Court found Stability liable for trademark infringement. Some early versions of Stable Diffusion generated images had Getty’s watermark, which the court said could confuse customers about the image’s origin. Meanwhile, Japan settled the question by statute in 2018. Article 30-4 of its Copyright Act permits copyrighted works to be used for “information analysis,” including AI training, for commercial or non-commercial purposes as long as the use doesn’t involve a person actually “enjoying” the work’s expressive content, and doesn’t unreasonably damage the rights holder’s interests. The European Union took a different path. Its text-and-data-mining rules let commercial developers train on copyrighted material by default, but rights holders can opt out, and developers are expected to honour that. Personality rights The SAG-AFTRA strikes in Hollywood and the video game industry brought AI issues into sharp relief. The 2023 film and TV strike covered many issues, but a core demand was protection against the uncontrolled use of AI to replicate actors’ voices, likenesses, and performances. The strike demanded: Notice and consent before their voices or movements could be used to create digital replicas or AI-generated performances. Clear limits on how those replicas could be used, and fair compensation when they were. Protections covering not only voice actors but also stunt performers and motion-capture artists, whose work AI could mimic even without recognisable faces. The resulting contracts required studios to obtain consent and compensate performers when replicating their work with AI. Paying the news publishers Australia’s News Media Bargaining Code, passed in 2021, forced Google and Meta to negotiate payment with news publishers or face a government-appointed arbitrator. Both platforms hated it, and both eventually signed deals worth a reported 200 million Australian dollars combined, rather than risk formal “designation” under the code. Canada also tried something similar with its Online News Act in 2023; Meta responded by blocking all news content for Canadian users rather than pay. When Meta’s Australian deals began expiring in 2024, it announced it would not renew them, and would shut down its dedicated news product in the country. The Australian government responded in 2025 by proposing a “news bargaining incentive” — a charge on platforms that decline to strike deals with publishers. When forced to pay, tech platforms often prefer to withdraw the product entirely rather than share revenue. Invisible doctrine A small group of billionaire founders and investors control the most powerful models and the data pipelines that feed them. They benefit from publicly funded research, open-source tools, and the creative labour of millions, while pocketing most of the financial upside. AI has disrupted traditional search and news discovery, shifting traffic away from publishers’ sites and toward chat interfaces controlled by a few firms. These firms effectively control the information water table. In short, the fight over AI and copyright is also a fight over who owns the future of knowledge, culture and democracy and whether those futures will be shaped by narrow corporate interests or by rules that reflect broader public values.