Delhi High Court's OpenAI Ruling Forces Copyright Law to Grapple With How AI Actually Works
A recent interim order from the Delhi High Court has pushed Indian copyright jurisprudence into unfamiliar territory, requiring judges to parse the technical architecture of Large Language Models before they can even begin applying century-old legal concepts like copying and reproduction.
The case, ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, 2026 SCC OnLine Del 5291, arose from ANI's plea for an interim injunction against OpenAI. But what makes the ruling notable is not merely its outcome at the interim stage — it is the extent to which the Court felt compelled to explain, in some technical depth, how an LLM actually functions before it could assess whether copyright had been infringed.
When the Copyright Act, 1957 was drafted, lawmakers were dealing with a world of physical copying and straightforward reproduction. They could scarcely have anticipated a scenario where courts would need to distinguish between training a model, retrieving data, storing information, and generating text — four processes that, while they may look similar from the outside, are technically distinct operations.
Breaking Down the Black Box
Drawing on submissions from the parties, inputs from Amici Curiae, and what it termed common general knowledge, the Delhi High Court attempted to walk through the lifecycle of an LLM in stages — from the assembly of training data to the generation of a final response.
The judgment starts with the raw material: training data. An LLM, the Court explained, is fundamentally a predictive system — one that guesses the next word in a sequence based on the words that came before it. To make these predictions, the model absorbs a massive corpus of text, drawn from publicly scraped internet material as well as licensed content, which is then filtered for quality and safety before being used.
From there, the Court moved to data preparation and training itself. Raw text gets chopped into smaller units called tokens, each assigned a numerical Token ID. These tokens are then converted into embeddings — numerical representations that the model can mathematically process. During training, the Court noted, portions of text are masked, and the model is tasked with predicting what comes next; its guesses are checked against the actual text, and the model is refined accordingly.
Post-training, the model undergoes further refinement to follow instructions and generate useful responses. When a user types a prompt, the model produces its answer by generating one token at a time, each conditioned on what came before.
Finally, the judgment addressed Retrieval-Augmented Generation (RAG) — a method by which an LLM can pull information from sources outside its original training data before crafting a response. Here, an information-retrieval component uses the user's prompt to fetch relevant material from an external database, which is then blended with the model's trained knowledge to produce the final output.
This technical groundwork, the Court suggested, was essential to its subsequent legal reasoning — making it worth examining more closely.
Tokens, Embeddings and the Limits of Analogy
Understanding an LLM begins with how it converts text into something it can actually process. Text is first broken into tokens, each tagged with a numerical ID. Importantly, a token does not always equal a full word — depending on the tokenizer, it might represent a whole word, a fragment, or even punctuation.
Take the word "unbelievable." Depending on how the tokenizer slices it, this could split into fragments resembling "un," "believ," and "able," rather than being treated as a single indivisible unit. Each token is then mapped onto an embedding — essentially a vector that allows mathematical manipulation. The key takeaway: a token, its Token ID, and its embedding are three separate stages of processing, not interchangeable concepts.
But tokenisation and embedding merely prepare text for the model — the real learning happens during training. In GPT-style architectures, the model is fed a sequence of tokens and learns to predict probabilities for what comes next. If given "The Court dismissed the," the model might assign probabilities to words like "appeal," "petition," or "application." Its guess is compared against what actually appears in the training text, and the resulting error adjusts the model's internal parameters. Multiply this across enormous volumes of text, and the model gradually develops a statistical sense of language patterns.
This is distinct from a different technique known as masked-language modelling, where a token is deliberately hidden within a sentence and the model must infer it from context — for instance, filling in the blank in "The Court ___ the appeal." Autoregressive models, by contrast, only look forward, predicting what logically follows a given sequence. GPT-type systems primarily rely on the autoregressive, next-token-prediction approach rather than masked-language modelling, though both represent legitimate but different training objectives.
Crucially, an embedding is not the same thing as the underlying text, and the model's parameters should not be mistaken for a searchable database from which individual documents can be pulled out. What actually happens during training is a continuous adjustment of parameters as the model absorbs patterns — not a simple storage-and-retrieval process.
This raises an obvious follow-up question: if the model isn't storing text in any conventional sense, what exactly does "learning" mean here — and how does memorisation fit into the picture?
Learning Is Not the Same as Memorising
Storage, learning, and memorisation, the analysis makes clear, are three technically distinct phenomena — a distinction with significant legal implications.
A copyrighted work might initially exist within a training dataset, stored in the conventional sense as raw data. But during training, the model typically doesn't retain documents as discrete, retrievable records. Instead, exposure to the data causes adjustments across a vast number of numerical parameters, through which the model absorbs statistical relationships spread across the entire dataset. When generating a response later, these learned parameters help assign probabilities to the next token, building the output piece by piece.
This means learning from a work doesn't automatically translate into memorising it. A model could theoretically absorb patterns from millions of documents without ever being able to spit out any specific one verbatim. Memorisation, in the technical sense, refers to situations where particular sequences from the training data can actually be reproduced under the right conditions. And even memorisation is conceptually separate from regurgitation — the actual appearance of that memorised material within a generated response.
Where Does the Information Actually Come From?
Not every response an LLM produces stems purely from what it absorbed during training. Systems can be designed to fetch information from external sources in real time, with RAG being one common method for doing so.
In a typical RAG setup, a user's query triggers a search of an external knowledge base. Whatever is retrieved gets fed back into the LLM as supplementary context before it crafts its response. This involves two separate components working together — retrieval, which locates relevant information, and generation, where the LLM combines that retrieved material with the user's original prompt.
However, RAG shouldn't be conflated with every possible way an AI system might access outside information. Models can also pull current data through web search, browsing tools, application programming interfaces (APIs), or various other connected tools. These might work alongside an LLM, but simply because information originated outside the model's original training data doesn't automatically mean RAG architecture was involved.
This distinction carries particular weight when a response includes details about events occurring after a model's training cutoff date. Such content might indicate the information wasn't sourced from the original training dataset — but that alone doesn't reveal the actual mechanism by which it reached the system. Whether it came through RAG, a web search, an API, some other connected tool, or was simply generated without any reliable external anchor remains a separate technical inquiry altogether.
Why This Technical Detail Actually Matters for Copyright Law
These distinctions aren't just academic — they go to the heart of what copyright law actually protects against, because legal consequences hinge on what specifically happened to a protected work.
Consider a hypothetical copyrighted legal commentary. It could form part of an LLM's training dataset; specific passages from it might get memorised during training; it could be retrieved from an external database in response to a user query; or fragments of its text might surface in a generated response. Even though it's the same underlying work in each scenario, the technical process involved differs substantially — and so, potentially, does the legal analysis.
The distinction sharpens further when a commentary was published after a model's training data cutoff. If a user asks the model about that work and receives a response, the answer itself won't reveal where that information originated. The system could have accessed the commentary via RAG, a web search, an API, or some other external tool. Alternatively, it might generate a response without ever having actually accessed the material — producing something that sounds plausible but is factually unsupported or simply wrong, a phenomenon widely known as "hallucination."
In other words, a model's apparent ability to discuss a work published after its training concluded doesn't by itself tell you how that information was obtained or generated.
Technology, in this sense, doesn't dictate the legal outcome — but it does help pinpoint exactly which act copyright law is being asked to evaluate. The inquiry, ultimately, has to start with a fairly basic question: what has the AI system actually done with the protected work in question?
A New Chapter for Indian Copyright Law
ANI v. OpenAI spotlights a growing challenge in applying traditional copyright principles to generative AI systems. Concepts like copying, storage, and reproduction — long settled in legal doctrine — now run up against technological processes that resist being neatly categorized using that old vocabulary.
Perhaps the judgment's most significant contribution lies in how it folds technical understanding directly into copyright analysis. Terms like training data, tokenisation, embeddings, model training, memorisation, and RAG are no longer confined to computer science textbooks — they've entered the language of copyright adjudication in India. This ruling stands as an early marker of that shift.
As generative AI technology continues to develop, courts will likely confront these same concepts across increasingly diverse factual scenarios. ANI v. OpenAI represents a moment where Indian copyright jurisprudence has begun directly engaging with the underlying architecture of these systems — not just their outputs.
Beyond resolving the specific interim dispute before it, the judgment offers an early framework for understanding how questions about generative AI's technical architecture might surface within copyright litigation going forward. In many respects, this case marks just the opening move in what promises to be a much longer dialogue between the technical realities of generative AI and the foundational principles of copyright law.
Tags
- News
- by-opportunities-desk
- Opportunities
- All High Courts
- Opportunities for Students
- Supreme Court of India
- by-nishtha
- Top Stories
- Job Openings
- Internships
- Criminal Law
- Delhi High Court
- by-tarun
- Constitutional Law
- Bail Matters
- Judiciary
- Governance
- Political
- Call for Papers
- Allahabad High Court
- Civil Procedure
- by-deepanshu
- by-neerav
- Kerala High Court
- Bombay High Court
- by-vivek
- Education
- by-utkarsh
- by-yash
- Advocacy
- Business
- Karnataka High Court
- Issues
- Matrimonial
- Media & Films
- Employment
- Moot Courts
- by-jasveer
- District Courts
- Jammu & Kashmir High Court
- Madras High Court
- Service Law
- by-anubhav
- Medical
- Public Interest / PIL
- by-ayush
- Transportation
- Tribunals & Authorities
- Banking & Finance
- Calcutta High Court
- Competitions
- Consumer
- Arbitration
- Environment
- Taxation
- Trademarks & Patents
- Advertising
- Money Laundering
- Electricity
- Gujarat High Court
- Land Acquisition
- Punjab & Haryana High Court
- Rajasthan High Court
- Aviation
- Madhya Pradesh High Court
- Meghalaya High Court
- Patna High Court
- Sports
- Articles
- Cyber Law & Data Protection
- Enviornment
- Events
- Gaming
- Himachal Pradesh High Court
- Insolvency & Bankruptcy
- National Company Law Appellate Tribunal
- Andhra Pradesh High Court
- Civil Law
- Jurisdiction
- Tripura High Court
- Uttarakhand High Court
- Central Consumer Protection Authority
- Chhattisgarh High Court
- Child Custody
- Delhi State Consumer Disputes Redressal Commission
- Finance
- Gauhati High Court
- Guahati High Court
- Insurance
- Jharkhand
- Jharkhand High Court
- Kerala
- Loans & Defaults
- Motor Accident Compensation
- NGOs
- National Green Tribunal
- Right To Information
- Sikkim High Court
- State Consumer Commissions
- Telangana High Court
Follow Nyaysha
Fresh legal updates across our platforms — follow wherever you already scroll.
LinkedInJudgments, appointments & career moves — for professionals Follow Instagram@nyaysha.in — visual explainers, reels & court-room snapshots Follow FacebookCommunity discussions & week-in-review roundups Follow TelegramLegal jobs, internships, moot courts, competitions, events & call for papers — instant alerts, first to know Join