ChatGPT and Gemini for Indian Lawyers:
Prompt Engineering for Legal Drafting

Supreme Court Warnings on AI-Hallucinated Citations

Generative AI models are predictive text engines, not verified legal search engines. They predict the next word based on mathematical probability, which frequently leads to the creation of fictitious case citations (hallucinations) that look incredibly authentic—complete with fake bench names and SCC reporter volumes—but do not exist in reality.

Advocates MUST manually verify every single citation generated by an AI model. Presenting a hallucinated judgment in court constitutes misleading the bench and violates professional ethics codes, potentially leading to severe contempt warnings from presiding judges and disciplinary procedures by state bar councils.

These high-profile warnings highlight the non-negotiable necessity of human review. Junior associates must verify every ratio decidendi against physical reporter journals or official online portals before including them in final draft filings.

Prompt Structures for Contract Drafting

To draft corporate clauses effectively, you must define the AI's role and the governing jurisdiction. For instance, instruct the model: "Act as a Senior Indian Corporate Counsel. Draft a comprehensive Indemnity Clause governing liability under the Indian Contract Act, 1872."

Specify constraints like mutual obligations, territorial jurisdiction (e.g., Courts of Mumbai), or specific liability caps (e.g., limited to 12 months of contract value). Defining these parameters prevents vague, unusable outputs, ensuring the clause is structured to aggressively protect your corporate client's interests.

You should also define formatting styles, requesting section divisions, clear headings, and avoiding unnecessary legal jargon (legalese). This output layout ensures the generated text is easy to integrate into larger master service agreements (MSAs).

Prompt Engineering for Legal Notices

Legal notices require precise statutory timelines and accurate citations. When prompting for a breach of contract notice, explicitly specify the details of the breach, the exact contract execution dates, and the legally required remedy period (e.g., 15 days or 30 days depending on the statute).

Instruct the model to structure the notice in standard Indian legal format: (1) Background Facts of the transaction, (2) Specific details of the breach/default, (3) Statutory demands, and (4) Proposed remedies and consequences of non-compliance. This provides a clean, logical draft for advocate review.

Providing this highly structured framework prevents the AI from generating generic, overly emotional warning statements. It ensures the draft contains the necessary legal definitions and formatting to be an effective, court-admissible communication tool.

Prompting for Judgment Summaries

Reviewing 200-page Supreme Court judgments is incredibly time-consuming. You can use large context window models (like Gemini 1.5 Pro) to upload PDF case files and summarize them instantly. Instruct the model: "Read the attached PDF. Extract the core legal issues framed, summarize the ratio decidendi, and outline the final order."

Crucially, warn the model to "ignore obiter dicta and focus only on binding precedents and statutory interpretations." This allows you to evaluate case laws rapidly to see if they apply to your current brief before diving into deep reading.

These AI summaries serve as excellent quick reference guides for court preparation. They allow advocates to digest hundreds of pages of case records and prepare argument briefs or synopsis notes in significantly less time.

Prompting for Vernacular Legal Translation

In Indian District Courts, pleadings, FIRs, and witness statements are frequently recorded in local vernacular languages (Hindi, Marathi, Tamil, etc.), while High Court and Supreme Court proceedings require English translations.

AI models like ChatGPT and Gemini are exceptional at translating complex documents while retaining the legal context. Prompt structure: "Act as a professional legal translator. Translate the following Hindi FIR into formal Legal English. Maintain the exact chronological order of events and accurately translate penal code sections without altering the meaning."

Always review the translated output to ensure nuances (such as specific regional terms for land measurement or family relations) are not lost or mistranslated, as these details often form the crux of cross-examinations.

Prompt Analysis: Vague vs. Structured

Reviewing the differences in prompt structures clearly demonstrates how precise role definitions and constraints affect the quality and usability of AI outputs.

Preventing Client Data Leaks in Public LLMs

Pasting confidential client agreements, unredacted FIRs, or sensitive litigation files into public AI search systems represents a massive, career-ending breach of confidentiality under professional ethics. Public models store input data to retrain future algorithms, meaning highly sensitive corporate terms could become discoverable to opposing counsel or the public.

To protect client privacy, advocates MUST: (1) Use offline or enterprise-grade models (like ChatGPT Enterprise or Google Workspace Gemini) with strict zero-data-retention terms. (2) Replace all names, financial values, account numbers, and location identifiers with generic placeholder terms (e.g., Party A, City X, INR [Amount]) before pasting drafts. (3) Manually turn off history and data training settings in your personal AI profiles.

Protecting client details protects your practice from devastating ethical liability. Establish strict office data guidelines, ensuring all junior associates and paralegals know exactly how to anonymize documents before using AI platforms.

Ethical Privacy Setup and History Settings

Before integrating AI into your chamber workflow, heavily audit the privacy settings of the application. Most platforms (including ChatGPT and Claude) offer a dashboard where you can completely disable model training on your prompts.

Ensure this setting is configured and locked for all computers in your office. Training staff on basic anonymization protocols prevents accidental leaks of case details, protecting your practice from severe ethical liabilities and potential lawsuits from corporate clients.

Using offline, open-source models (like Llama) on local chamber servers is the absolute safest way to ensure data security for highly sensitive matters. Transitioning to enterprise AI software with zero-training service level agreements (SLAs) is highly recommended for firms handling M&A transactions or high-profile criminal defense.