After the Pooja Ramesh Singh judgment debacle, the question is no longer whether lawyers should use generative AI. It is how to use it judiciously so every outcome survives a judge, a client and a bar council.

This traces back to July 2, 2026, when a Supreme Court bench of Justices PS Narasimha and Alok Aradhe set aside NCLT and NCLAT orders in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd because they rested on precedents that didn’t even exist in real life. The Court called this out, stating zero tolerance for unverified, hallucinated AI-generated authorities and characterising their citation as misconduct by any advocate. The Court then directed the Bar Council of India to constitute a committee to address the problem.

It wasn’t an isolated incident that led the Supreme Court to take a strong stand against AI-generated content. A month earlier, the Supreme Court’s AI committee released the draft Regulations for Use of Artificial Intelligence in Courts, 2026, which require lawyers to disclose AI assistance in pleadings and filings.

Together, these developments raise a practical question. This piece explains how these tools actually work, where they genuinely help with drafting and legal content, and a working method that keeps advocates (not the model) in control of the outcome.

You need to understand that a large language model doesn't look anything up. It predicts the most probable next word, given its context, based on patterns it learned from its training data. This is how it writes fluent prose and why it can also produce a perfectly formatted, entirely fictional citation.

Everyone needs to distinguish between the two sources of an AI model's responses. The first is parametric memory. This is what the model absorbed in training, which stops at a fixed cutoff and is thin on Indian case law compared with American and English material. The second is the context window. It’s the text you place in front of the model in that conversation, such as a judgment, a statute or even a term sheet draft, followed by a prompt.

There is also a real difference between a general-purpose chatbot - as we’re seeing with models like Claude, Perplexity, ChatGPT, Gemini, etc - and India-specific research products such as SCC Online’s AI Tools, Manupatra’s assistant or CaseMine’s AMICUS. The latter use retrieval-augmented generation, which first searches a curated corpus and then writes from what it finds. Retrieval reduces hallucination, but the onus is still on the lawyer to verify and cross-check.

The safest and the most productive uses share one feature: the lawyer supplies the source and the AI model simply transforms it. AI deployed this way replaces a fast junior who reads quickly and drafts tirelessly, but must never be left unsupervised.

In disputes, the gains lie in compression and structure. An AI model can turn a long order sheet into a dated chronology that expedites the work. It extracts the issues framed in a judgment that a lawyer has already pulled and reorganises rough notes into a skeleton of written submissions.

In transactional work, this means producing a first draft of a non-disclosure agreement, or a master services agreement from the firm’s own precedent, not from the model’s imagination. It compares the two versions of a share purchase agreement and lists every change to indemnity caps, or checks a draft against a client’s term sheet for missed commercial points.

The caveat here is that the model should not be asked to supply the law. Authorities, limitation periods, the current text of a provision and the effect of a recent amendment must come from an authorised or official database. The lawyer must verify the output, no matter how advanced the model is.

Lawyers now publish constantly: client alerts, LinkedIn posts, newsletters, blogs, opinion articles, etc. AI is very useful for exploring content structures, ideas, themes, topics, sub-topics, structuring an argument, tightening a content draft, converting a long article into a short one, or turning a judgment summary into plain, easy-to-understand content for business readers. The risk here is subtler than a hallucinated citation, but it is still a real one.

The first problem here is the accuracy in public. Every section, rule, citation and analysis published deserves the same scrutiny as one in a filing, because here, the reputational cost is often irreparable and quite embarrassing.

The second problem is the conduct. Under Rule 36 of the BCI Rules, advocates cannot solicit or advertise. A model that is asked to make the post “more engaging and compelling for readers” can readily add superlatives, outcome claims or guarantees that might be gripping for the reader, but the tone/language of the content could also cross the very thin line of “solicitation”, as prohibited by the rules.

The third problem is voice. Readers, including journalists and in-house counsel, can recognise generic AI prose that tries to say nothing, just by one read. Thought leadership works only when it carries the author’s own judgment. The best AI is, therefore, editorial: it helps a lawyer articulate more clearly, not generate what they should think.

The method described here is not elaborate. Its true value lies in being applied every time.

1. It begins with sourcing. Authorities should be pulled up by a lawyer from an authorised database and statutory text from the official source before involving the model to do its work. The model may help frame search terms, but its case supply shouldn’t be finalised without cross-checking.

2. Next comes grounding. The lawyer must give the model accurate, actual materials and instruct it to work only from the supplied text, to quote paragraph numbers and to say so when the answer isn't in it. A prompt that asks for a clause “based on Indian Law” invites invention and hallucination. A prompt that attaches uploaded precedents does not.

3. Confidential inputs need their own discipline. Anonymise client names, identifying facts and commercially sensitive figures before they reach any AI tool whose data retention and training terms the lawyer has not reviewed. For regular confidential work, an enterprise arrangement with contractual limits on retention and training is the defensible choice, rather than a free or basic account.

4. Verification should be done outside of the tool. Every citation must be checked; the quoted paragraph should be read in the original version. Asking the same model to confirm its own citation is like asking a witness to cross-examine themselves.

5. Finally, the lawyer must keep a record. A short note on what tool was used, who verified the output, what sources were used for it, etc. This comes in handy when the court/client asks (if at all) about how the document was prepared.

This method also works well for thought leadership, where you become the first source of information and use AI to sharpen your articulation, rather than letting AI become a thought leader.

AI is a powerful engine for transforming text a lawyer already trusts, but an unreliable witness to the law itself. Used judiciously, the discipline can turn AI from a professional risk into one of the quieter advantages known to a practitioner today.

Ultimately, AI can assist with drafting and writing. Judgment, however, still needs to come from a lawyer. Unfortunately, no prompt ensures professional accountability.

Bhavya Bhatt is a lawyer turned entrepreneur and a founder of Clause Play.

Ravi Kanth is a Business Development and Growth professional at Bar & Bench.

Disclaimer: The opinions expressed in this article are those of the author(s). The opinions presented do not necessarily reflect the views of Bar & Bench.