After ANI v OpenAI, AI Copyright Risk Moves to Outputs: Check Your IP and Media Liability Cover

On 24 July 2026, Justice Amit Bansal of the Delhi High Court dismissed ANI Media's application for an interim injunction against OpenAI, in the suit that had been the reference point for Indian AI copyright risk since it was filed. The order was reported by the EU IP Helpdesk on 28 July 2026.

Two findings carry the weight. First, the court held that copyright protects original expression and not the underlying facts, which limits how far a claim over factual news reporting can reach. Second, on outputs, the court found that ChatGPT responses generated through retrieval-augmented generation were not substantially similar to ANI's reports, and that there was no evidence of memorisation (Mondaq, 2026).

That second finding is the one commercial buyers should read carefully, because of how it was reasoned. The output escaped liability on the facts of that deployment: what the model produced did not reproduce the protected expression. It is not a ruling that model outputs are categorically safe. It is a ruling that a specific output, from a specific retrieval pipeline, over specific source material, did not cross the substantial-similarity line.

So the question an Indian company deploying generative AI now has to answer is not whether its vendor trained lawfully. It is what its own deployment publishes, and whether that output crosses the same line the court just drew.

Read the procedural posture before you read the reasoning into your risk register. The order is interim, it is appealable, and it is not precedential. The substantive suit continues (Mondaq, 2026).

An interim order on an injunction application turns on a prima facie view, balance of convenience, and irreparable harm. It is not a final determination of infringement. A different bench at the appellate stage, or the same bench at trial with a fuller record, can reach a different conclusion. Nothing in the order binds any other court hearing any other AI copyright claim in India.

That matters for insurance in a specific way. Underwriters price the probability of a claim being brought and defended, not only the probability of a final loss. Defence costs are incurred whether or not the plaintiff eventually wins. An order that declines an injunction while leaving the suit alive is, from a defence-cost perspective, an expensive outcome for the defendant and an invitation to the next plaintiff to plead the facts differently.

If your generative AI deployment produces text, images, audio, or code that reaches customers, the exposure lands across four places in a typical Indian corporate programme, and none of them was designed with model output in mind.

1. Standalone IP infringement cover. Available in the Indian market largely through international carriers and reinsurance capacity, usually written on a claims-made basis, covering defence costs and damages for unintentional infringement of third-party intellectual property. This is the most direct fit for an output-similarity claim, and it is also the cover most Indian buyers outside media and technology do not hold.
2. Media and advertising liability. Written for publishers, broadcasters, agencies, and increasingly for any business that produces content at volume. Typical grants respond to copyright infringement, passing off, defamation, and breach of privacy in published material. If your AI writes marketing copy, product descriptions, or customer-facing articles, this is where a plagiarism-flavoured claim naturally goes.
3. The advertising injury grant inside the liability policy. Commercial general liability wordings carry a personal and advertising injury section that historically covers infringement of copyright, trade dress, or slogan in the insured's advertisement. Many Indian corporates hold this without knowing it, and it is the coverage most likely to be quietly removed at the next renewal (see the next section).
4. Technology errors and omissions or professional indemnity. Responds where the AI output is part of a professional service delivered to a client and the claim is framed as negligent performance rather than infringement.

The common failure is assuming one of these is doing work that another one has to do. An IP infringement policy that responds to patent and trade mark claims may not extend to copyright in generated text. A media liability policy scoped to "editorial content" may not reach an internal chatbot whose answers get pasted into a proposal.

The personal and advertising injury grant is the oldest place copyright risk sits in a liability policy, and it is the narrowest fit for AI.

Standard wordings tie the grant to infringement committed in your advertisement. That phrase is doing a lot of work. Output that appears in a marketing campaign is plausibly in an advertisement. Output that appears in a support chatbot reply, a generated user manual, a code suggestion, or an internal research summary that later gets published is much harder to fit inside the definition. A claimant does not need the grant to be a poor fit for their theory; the insured needs it to be a good fit for their defence.

The second problem is the knowledge and intent conditions. Advertising injury grants commonly exclude infringement committed by an insured who knew the act would violate rights, and material published with knowledge of its falsity. A generative deployment sits in an uncomfortable position here: the company did not know any specific output would infringe, but it did know, in a general way, that the model could produce material resembling its training sources. Whether that general awareness engages the exclusion is untested in India.

The third problem is that the grant carries no defence-cost sub-limit designed for an IP suit. Defending a substantial-similarity claim means expert evidence on model behaviour, retrieval logs, and output provenance. That is not the cost profile the grant was priced against.

Buyers should read the actual policy wording rather than the schedule. The difference between "infringement of copyright in your advertisement" and a broader media grant is the difference between a covered claim and an uncovered one.

While the Indian litigation was running, the wordings were changing. This is the part of the story most Indian buyers have not priced in, because the changes originated in the US forms market and arrive in Indian placements through reinsurance and international programme wordings.

Verisk ISO filed a set of generative AI exclusions that took effect for commercial general liability renewals from 1 January 2026 (Gallagher; Claims Journal, 20 July 2026):

• CG 40 47 01 26, the broad form, removes bodily injury, property damage, and personal and advertising injury arising out of generative AI.
• CG 40 48 is the narrower version, limited to Coverage B, which is the personal and advertising injury section.
• CG 35 08 applies the exclusion to products and completed operations.

The broad form is the one to look for. It does not carve out copyright, it does not distinguish training from output, and it does not turn on whether the AI was your own or your vendor's. If it is attached, the advertising injury route described in the previous section is closed.

The pattern is not confined to general liability. W.R. Berkley introduced an absolute AI exclusion across directors and officers, errors and omissions, and fiduciary liability lines, with AIG and Great American filing their own exclusions (Claims Journal, 20 July 2026).

Indian liability programmes rarely sit on purely domestic wordings once the limits get large. Excess layers, international programme structures, and reinsurance-driven forms all import language written elsewhere. An exclusion filed in the US forms market in January 2026 shows up in an Indian renewal quotation later in the year, often as an unremarked endorsement in a slip that otherwise looks identical to last year's.

The practical checks at renewal:

• Compare this year's endorsement schedule against last year's line by line. An AI exclusion is a one-page attachment that changes the answer to every question in this article.
• Establish which layer carries it. An exclusion on a primary layer with clean excess is a different problem from an exclusion that follows form all the way up.
• Ask whether a narrower version is available. CG 40 48 limited to Coverage B leaves the bodily injury and property damage grants intact; the broad CG 40 47 01 26 does not. If the carrier is attaching an exclusion regardless, the negotiation is about which one.
• Test how "arising out of generative AI" is defined. Broad causation language can reach a claim where AI played a peripheral role in producing the material, such as a human-edited draft that started as model output.

This is the same wording discipline that applies elsewhere in the AI programme, and it connects directly to the governance work covered in AI governance frameworks for insurers and brokers and to the sectoral rules discussed in India's AI governance guidelines. Underwriters increasingly ask for the same artefacts a governance framework produces: system inventory, output review controls, and retention of generation logs.

The work is narrow and specific. It is a review of your own deployment, not of your vendor's training corpus.

Inventory what your AI publishes. List every generative system whose output reaches a third party: marketing copy, product listings, chatbot replies, generated documentation, code shipped to clients, images and audio. For each, record whether the output is reviewed by a human before publication, and whether that review is documented.

Identify the source-similarity risk. The ANI order turned on substantial similarity and the absence of memorisation. Deployments that retrieve from a narrow corpus of third-party material and paraphrase it closely carry more of that risk than deployments generating from broad general knowledge. Retrieval pipelines pointed at a small set of external publications are the highest-risk shape.

Keep the logs. Prompt, retrieved context, model version, and output, retained for the limitation period of a copyright claim. In a substantial-similarity dispute, the ability to show what the system actually did is the defence. A deployment that cannot reconstruct how a given output was produced defends on assertion alone.

Read three documents at renewal. The exclusion schedule on the liability policy, the definition of advertisement and of covered media in any media liability section, and the vendor contract's IP indemnity, including its cap and its carve-outs. Vendor indemnities for output claims are common in enterprise AI contracts and are frequently capped at a level well below the cost of defending a suit in the Delhi High Court.

Decide whether standalone cover is warranted. For a business whose AI output is customer-facing and content-heavy, a media and advertising liability policy or a standalone IP infringement policy buys certainty that the advertising injury grant no longer provides. For a business using AI internally with human review before anything is published, the honest answer may be that existing cover plus tightened controls is proportionate.

The ANI order made the vendor-side argument less frightening and the deployment-side argument more important. The insurance market had already moved, seven months earlier and in the other direction. Between the two, the company that can describe exactly what its AI publishes and how that gets checked is the one that will find cover at a sensible price.