How to Use ChatGPT (and Claude) for Legal Research and Drafting in India: A Step-by-Step Guide

ChatGPT, built by OpenAI, and Claude, built by Anthropic, are large language models. They predict the most probable next words given your prompt. They are extraordinarily good at language tasks such as summarising, restructuring, explaining, drafting, comparing and translating. They are not legal databases, and they were never designed to be.

Both are, by mid-2026, remarkably capable. The current Claude Opus 4.x and GPT-5.x families can hold roughly a million tokens of context, enough to paste an entire contract, a full judgment, or a long statute and have the model reason across all of it at once. But two limitations govern everything that follows.

First, they have a knowledge cutoff. Each model’s training data stops at a point in early 2026. It does not know a notification, amendment, or judgment issued after that date, and it cannot browse your paid legal databases.

Second, and this is the limitation that ends careers, they do not reliably know Indian case law from memory. When you ask an LLM to cite Indian cases on a point without giving it the source material, it will often produce citations that look perfectly real: correct-sounding party names, plausible neutral citations, confident paragraph numbers, for judgments that do not exist. The Delhi High Court flagged exactly this as early as 2023 in Christian Louboutin SAS v. Shoe Boutique, warning of fictional case laws and imaginative data and holding that the responses from ChatGPT cannot be the basis of adjudication of legal or factual issues in a court of law. AI, the Court said, is at best a tool for preliminary understanding or preliminary research.

This is precisely where India’s own legal-AI tools differ. Products like Manupatra.ai, SCC Online’s AI research assistant, Indian Kanoon’s Prism, and CaseMine’s AMICUS run retrieval-augmented generation: they answer by searching a real, curated corpus of Indian judgments and then summarising what they actually found. That grounding sharply reduces, though it does not eliminate, the hallucination risk. General-purpose ChatGPT and Claude have no such corpus unless you give them one. Which brings us to the single most important idea in this guide.

The golden rule: give the model the source, don’t ask it for the source

If you remember nothing else, remember this. There are two ways to use an LLM for law, and they have opposite risk profiles.

The first is “reason over what I paste”: you supply the judgment, the contract, the statute, the facts, and ask the model to summarise, compare, extract, or draft from that material. This is safe, powerful, and where most of the real value lives.

The second is “tell me the law”: you ask the model, from its own memory, to state the current position, cite authorities, or confirm a section number. This is where hallucination lives.

Every workflow below is built to keep you on the first side of that line. Use the AI to think with material you control, and use authorised databases such as SCC Online, Manupatra, or the official reports to source and verify the law itself.

Step-by-step: the legal research workflow

Step 1: Frame the issue with the AI, not the authorities. Start by using the model to sharpen your question and surface sub-issues you might miss. You are not asking it for law here; you are asking it to structure your thinking.

Step 2: Generate search terms and a research plan. Ask the model for the vocabulary and query strings you will feed into your legal database.

Step 3: Pull the actual authorities yourself. Run those searches on SCC Online, Manupatra, or Indian Kanoon. Download the real judgments. This is the step no general-purpose AI can do for you, and the step that keeps you honest.

Step 4: Feed the real material back for analysis. Now paste the genuine judgments in and let the model do what it is genuinely excellent at.

Step 5: Verify every citation before it goes anywhere. Even when working from pasted text, confirm that each neutral citation, party name, and paragraph number matches the authorised report. Treat an unverified citation as if it does not exist. The rule is simple: never cite what you have not opened.

Used this way, the AI compresses hours of reading into minutes of structured analysis, while the sourcing and the citing stay firmly in the authorised, human-verified lane.

Step-by-step: the drafting workflow

Drafting is where these tools shine brightest, because drafting is a language task. The workflow mirrors research: you supply the raw material and the model shapes it.

Step 1: Give the model the facts and your precedent. Never start from a blank “draft me a contract.” Feed it the deal facts and, ideally, a clause from your own precedent bank so the output matches your house style and Indian practice.

Step 2: Ask for alternatives and explanations. Use the model to widen your options, not just produce one draft.

Step 3: Edit for Indian statutory accuracy. This is your job, not the model’s. The AI’s draft is a starting point. You verify governing-law and jurisdiction clauses, stamp duty and registration requirements, section references, and anything statute-specific. The model may confidently name a wrong provision; you catch it.

Step 4: Red-team your own draft. One of the most valuable uses of AI is turning it against your own work.

This adversarial pass routinely surfaces missing termination triggers, silent dispute-resolution clauses, and undefined terms that a tired drafter glides past.

Prompting well: the part most guides skip

The difference between a useless AI answer and a genuinely good one is usually the prompt. A reliable structure is role, context, constraints and format: tell the model who to be, give it the facts, set the boundaries, and specify the output shape. Always state that it is Indian law. Ask it to show its reasoning, and ask it to flag its own uncertainty rather than bluff.

Iterate. Your first prompt rarely produces your best output, so refine it, add constraints, and push back when the answer is thin.

The red lines: India-specific risk and compliance

This is the section that separates a lawyer who uses AI well from one who is one bad filing away from disaster.

Hallucinated citations

This is the cardinal risk, and India now has a wall of cautionary precedent. Beyond the Supreme Court’s 2 July 2026 zero-tolerance ruling, consider the earlier incidents. In the Buckeye Trust matter (December 2024), an order of the Bengaluru bench of the Income Tax Appellate Tribunal, in a roughly Rs 669-crore trust-taxation dispute, cited three Supreme Court judgments and one Madras High Court ruling that simply did not exist, and the order had to be recalled under Section 254(2) of the Income-tax Act. In October 2025, the Bombay High Court quashed a Rs 27.91-crore tax assessment built on three non-existent, AI-generated precedents. And globally, the case that started the reckoning, Mata v. Avianca (June 2023), saw New York lawyers sanctioned for a brief full of fake ChatGPT cases. The rule that would have saved every one of them is the same: never cite what you have not opened and verified in an authorised report.

Confidentiality

Pasting privileged client information into a consumer AI tool is a real risk, not a theoretical one. On free and consumer tiers, your prompts may be retained and used to improve the model. Client information leaving your control can breach an advocate’s duty of confidentiality under the Bar Council of India Rules and the professional-communication privilege recognised in Section 132 of the Bharatiya Sakshya Adhiniyam, 2023. Both the Kerala High Court’s AI policy (July 2025) and the Gujarat High Court’s policy (April 2026) specifically warn against feeding official case data into public, cloud-based AI tools such as ChatGPT. The practical rule: anonymise facts, strip identifiers, and use enterprise tiers with data-retention controls for anything sensitive. Once the core obligations of the Digital Personal Data Protection Act, 2023 come fully into force, expected around 2027, this discipline will move from best practice to legal duty.

Disclosure to the court

The direction of travel is clear. On 3 June 2026, the Supreme Court’s AI Committee released the Draft Regulations for the Use of Artificial Intelligence in Courts, 2026 for public consultation, with comments open into July 2026. These are not yet law, but they signal the future: AI may assist but never adjudicate, a human must stay in the loop, and, critically for practitioners, the use of AI in preparing court documents may have to be disclosed to the court. Several High Courts have already restricted AI internally. Kerala (July 2025), Gujarat (April 2026) and Punjab and Haryana (April 2026) have barred judicial officers from using tools like ChatGPT to write judgments or conduct research. Those restrictions bind judges and court staff rather than lawyers, but they tell you exactly how seriously the judiciary now takes the issue.

Professional accountability

Under the Advocates Act, 1961 and the Bar Council of India Rules, the duties not to mislead the court, to remain competent, and to keep client confidences apply to AI-assisted work exactly as they apply to everything else. The competence duty requires you to verify every authority you place before a court. AI has no legal recognition; the advocate alone is accountable for whatever the AI produced. As of July 2026, the BCI has no AI-specific rules, and the Supreme Court has asked it to frame them. Until it does, the safe assumption is that you are fully responsible for every word.

A safe checklist before you rely on any AI output

Run every AI-assisted work product through the checklist below before it leaves your desk. Verify every citation in an authorised database, keep privileged client data off consumer tiers, keep a human in the loop on every legal conclusion, check the position for currency against the latest law, be ready to disclose AI assistance where a court asks, and remember that you, not the model, own the output.

Conclusion

The most useful way to think about ChatGPT and Claude in Indian legal practice is as a brilliant, fast, tireless junior associate: one who drafts a first cut in seconds and reads a hundred pages without complaint, but who occasionally invents a case with total confidence and must never be left unsupervised. The lawyers who win with these tools are not the ones with the best model; they are the ones with the best workflow. Give the model the source, use it to think and to draft, and keep the sourcing, the verifying and the accountability firmly human. Do that, and AI becomes the biggest quiet advantage in your practice. Skip it, and you are one confident hallucination away from the wrong kind of Supreme Court mention.

Frequently asked questions

Can lawyers in India use ChatGPT and Claude for legal work?

Yes. No law or Bar Council of India rule prohibits an advocate from using general-purpose AI tools for research or drafting. What the law does require is that the advocate independently verify everything the tool produces and take full responsibility for it. The recent High Court policies restricting AI apply to judges and court staff, not to practising lawyers.

Is it safe to use ChatGPT for legal research in India?

It is safe only if you use it correctly. The danger is asking the model to recall Indian case law from memory, which produces hallucinated citations. The safe method is to pull real judgments from an authorised database yourself and paste them in for the model to analyse, and then verify every citation before you rely on it.

Can I cite a case that ChatGPT or Claude gives me?

Never without independently verifying it. In several Indian matters, including the Buckeye Trust ITAT order and a 2 July 2026 Supreme Court ruling, filings collapsed because AI-generated citations turned out to be fabricated. Always confirm that the case exists, and that the cited paragraph says what the AI claimed, in an authorised report before citing it.

Does the Bar Council of India allow AI use?

As of July 2026, the BCI has no specific rules on AI, either permitting or prohibiting it. In its 2 July 2026 judgment, the Supreme Court asked the BCI to frame guiding principles and consider disciplinary norms for lawyers who place fake AI-generated material before courts. Until those norms exist, an advocate’s existing duties of competence, candour and confidentiality govern AI use.

Can I put confidential client information into ChatGPT or Claude?

Be very careful. On free and consumer tiers, prompts may be retained and used to train the model, which can breach your duty of confidentiality. Anonymise facts, remove identifying details, and use enterprise tiers with data-retention controls for anything sensitive. Both the Kerala and Gujarat High Courts have warned against feeding case data into public cloud AI tools.

Which is better for legal work, ChatGPT or Claude?

Both are highly capable, and the differences matter less than your workflow. The current Claude and ChatGPT families both offer very large context windows that let you paste long documents for analysis. Many lawyers keep both and cross-check important outputs. Whichever you use, the golden rule is identical: supply the source text and verify every citation.

References