Delhi High Court refuses ANI’s emergency bid against OpenAI, a narrow win for AI, a warning for newsrooms
On July 27, 2026, Justice Amit Bansal of the Delhi High Court denied Asian News International’s (ANI) request for a preliminary injunction that would have forced OpenAI to delete or stop using certain news content, according to reporting in The Hindu. The order does not resolve the core copyright dispute, but it shows how an Indian court handled memorization, training exceptions and interim relief in an AI copyright case.
What the court decided, and why
The court refused ANI’s interim relief because, at this provisional stage, the plaintiff had not made a persuasive showing of (a) verbatim reproduction and (b) irreparable harm. The Hindu reports the judge found ANI had not produced material proving that OpenAI’s systems caused ANI to lose subscribers or syndication revenue.
- Prima facie showing (basic likelihood): The bench concluded ANI had not produced persuasive verbatim examples at the interim stage and therefore failed the threshold needed for an injunction (The Hindu).
- Statutory exception cited: The judge relied on Section 52(1)(a) of the Indian Copyright Act (the “private or personal use, including research” provision), treating certain training uses as potentially falling within that exception for the purposes of interim relief (The Hindu).
- Timing and provenance weakened the claim: The court accepted OpenAI’s timeline and filings enough to undermine ANI’s contention that the models had memorized those specific articles (The Hindu).
- Alternative explanations and expert commentary: Commentators and some technical submissions have pointed to retrieval and summarization behavior (commonly called Retrieval Augmented Generation, or RAG) as a non memorization explanation for similarities between model outputs and later published articles; the court noted that resemblance alone is not proof of memorization (MediaCopilot; The Hindu).
- No deletion order: The court declined ANI’s request to order deletion, citing preservation obligations and public interest and practical concerns that an interim deletion would cause disproportionate prejudice (The Hindu).
“There is no averment in the plaint nor has any material been placed on record to show that the activities of Open AI have resulted in ANI losing subscribers or suffering a loss from its news syndication business.”
That line, reported by The Hindu as the court’s finding, explains why the balance of convenience weighed against interim relief for ANI.
Why businesses should read this ruling early and closely
For product leaders, in house counsel and publishers, the order signals three practical themes worth attention:
- Evidence matters more than headlines: Courts deciding interim remedies focus on concrete, attributable evidence, verbatim reproductions, ingestion timestamps, supplier contracts, not broad allegations.
- “Research” exceptions are fragile and conditional: India’s Section 52(1)(a) can be invoked at the interim stage, but the judge framed it as conditional; lawful sourcing and limits on internal use will be central to holding that defense (The Hindu).
- RAG changes the compliance calculus: Systems that explicitly retrieve third party text create different legal pathways, with clearer attribution versus hidden memorization. Technical and product owners should treat RAG enabled features as a distinct compliance vector (MediaCopilot commentary).
Practical steps for AI teams and publishers
These are immediate, implementable controls that reduce litigation risk and strengthen commercial positions:
- Document provenance thoroughly: Keep supplier contracts, ingestion timestamps, source URLs, checksums or hashes of ingested files, and chain of custody records. Record who authorized each dataset and when.
- Segment research vs. production datasets: Enforce strict access controls and labeling for datasets used in internal model research so you can credibly claim internal, non public use where needed.
- Log and preserve retrieval activity: If you deploy RAG, keep logs that tie retrieved passages to source documents and timestamps. Those records help show whether output came from live retrieval or from internalized model weights.
- Design for attribution: When a model surfaces third party content, include clear source links or short excerpts rather than long verbatim passages. That limits substitution risk and supports downstream licensing conversations.
- Treat commercial substitution as a product risk: Monitor whether your agent’s summaries materially reduce visits or subscriptions to publishers and be prepared to negotiate licensing if your service substitutes for publisher traffic.
- Prepare an evidentiary playbook: Expect forensic analysis, n gram comparisons, similarity metrics, metadata audits and expert testimony, to be central in any merits hearing. Preserve raw data and chain of custody documents defenders and plaintiffs will seek.
Evidence, precedent and the international landscape
This interim order is one stop on an uneven global road. Some U.S. federal judges, on the facts presented in their dockets, have found training can be lawful and transformative (see analysis of Bartz v. Anthropic and Kadrey v. Meta by law firm commentators at JW). Other courts in different jurisdictions have reached different outcomes or left key questions open. The common thread is that outcomes hinge on factual detail, such as how data was acquired, what outputs the model actually generates, and whether there is measurable market harm.
Because the Delhi bench limited its analysis to the interim stage, expect the main trial and any appeals to focus on three contested issues: (1) whether models can be shown to reproduce copyrighted text from their parameters, (2) whether any training used unlawfully obtained material, and (3) whether the model’s outputs amount to a “communication to the public” that displaces publishers economically. Commentators note these are precisely the points courts elsewhere have grappled with, so the factual record ANI and OpenAI assemble will matter as much as legal theory (MediaCopilot; JW).
What publishers and legal teams should do now
- Audit and catalogue what publishers own and how those assets are distributed; know where your audience and revenue come from.
- For publishers considering license offers, document offers and responses carefully, the order notes ANI made a licensing proposal to OpenAI (reported at USD 7.5 million in The Hindu), which the court used to illuminate the commercial backdrop.
- For AI vendors, build a provenance first compliance program: contracts with data suppliers, ingestion logs, and a defensible chain of custody reduce risk and preserve commercial optionality.
- Engage empirically: if you claim market harm, quantify it. Courts will want concrete metrics tying model behavior to lost traffic or subscriptions.
Key questions readers are asking, answered
- Did ANI get an interim injunction?
No. Justice Amit Bansal denied ANI’s request for preliminary relief on July 27, 2026, finding ANI had not produced sufficiently persuasive evidence of verbatim reproduction or irreparable harm (The Hindu).
- Did the court say training is automatically lawful?
No. The judge tentatively applied Section 52(1)(a) (the “private or personal use, including research” exception) for the interim stage, but made that finding conditional on lawful sourcing and internal, non public use (The Hindu).
- Was RAG treated as the court’s explanation for similarities?
The court noted resemblance between outputs and later publications is not, by itself, proof of memorization; independent commentators and technical submissions have pointed to retrieval and summarization mechanisms (RAG) as a non memorization explanation (MediaCopilot; The Hindu reports the court’s concern, commentators supply the RAG framing).
- What happens next?
The merits hearing will probe provenance, acquisition practices and economic impact; expect forensic similarity testing, metadata audits and expert testimony. The Delhi order is interim and fact dependent, final outcomes can diverge once the full record is in.
- Should companies stop building AI agents or RAG features?
No, but they should adjust product risk management: document sourcing, log retrievals, design for attribution, and keep licensing conversations open where your product substitutes for publisher revenue.
The Delhi High Court’s interim order is a measured, evidence driven checkpoint. It resists sweeping conclusions at this early stage, but it signals what will matter in a full trial. For publishers, the crucial battleground will be proving economic displacement and unlawful acquisition. For AI vendors, the immediate priorities are provenance, logging and thoughtful RAG design, because the merits fight will be decided on paperwork and data as much as on doctrine.