.png)
A Practical Handbook On Using Artificial Intelligence In Day-To-Day Legal Practice
Created by Shubhank Suman, AI, Privacy & IP Counsel
The Supreme Court of India has released the draft Regulations for Use of Artificial Intelligence (AI) in Courts, 2026 for public consultation, inviting comments and suggestions until 20 June 2026. Although the draft Regulations are not yet legally enforceable, they provide an important indication of how the Indian judiciary intends to regulate and govern the use of AI in courts. As AI tools become increasingly common in legal research, drafting, and evidence management, legal professionals should consider adopting the principles reflected in the draft Regulations now, as they represent both emerging regulatory expectations and prudent professional practice.
What This Handbook Covers:
- Core principles governing AI in Indian courts (Regs. 4-18) - The draft establishes the foundational safeguards for judicial AI, including human oversight, transparency, accountability, fairness, data protection, and judicial independence.
- Permitted uses of AI in the justice system (Reg. 19) - The Regulations recognise AI as an assistive tool for legal research, document summarisation, translation, transcription, case management, and drafting support, while preserving human decision-making.
- Prohibited uses and judicial red lines (Reg. 20) - AI cannot decide cases, determine judicial outcomes, sentence individuals, profile litigants, predict behaviour, or perform functions that affect rights and liberties without meaningful human control.
- Duties of advocates and litigants using AI (Regs. 43–44) - Users of AI must disclose AI-assisted filings, verify AI-generated outputs, identify synthetic evidence, and remain personally accountable for any inaccuracies or hallucinations.
- AI-generated evidence, data protection and oversight (Regs. 44–48) - The draft introduces safeguards for AI-generated content, establishes verification and governance mechanisms, and imposes requirements relating to confidentiality, cybersecurity, and responsible data handling.
Handbook on AI in Day to Day Legal Practice · Shubhank Suman
The draft Regulations mainly govern how courts themselves use AI. But a small group of provisions apply directly to you the moment you use ChatGPT, Copilot, Gemini or any AI tool on a matter , in research, drafting, evidence or filing. This handbook pulls out only those provisions and turns them into plain rules you can follow every day.
One rule sits above everything else: AI assists; you remain fully and personally responsible for every word you file. If a pleading, document or piece of evidence turns out to be false because it was AI-generated, you carry the blame , and “the AI made a mistake” is expressly not a defence (Reg. 43(6)).
2. The Five Duties That Attach To You When You Use AI
Your duty | Source | What it means in practice |
|---|---|---|
Disclose AI-assisted work | Reg. 43(3) | If you used an AI tool to prepare or submit any document, pleading or evidence, tell the court at the time of filing, through a signed declaration in the prescribed format (Annexure I). |
Be ready to explain it | Reg. 43(4) | The court can ask which tool you used, how much it did, and what you did to check the output. You must be able to answer. |
Disclose AI / synthetic evidence | Reg. 20(h), 43(5), 3(1)(y) | AI-generated material cannot be put forward as evidence unless you openly say it is AI-generated. The same applies to synthetic images, audio or data. GenAI content also needs verification through the court’s designated mechanism. |
Verify before you rely | Reg. 8(3), 16 | AI output is only advisory. Check it against the original source before you use it in any document or argument. |
Own the outcome | Reg. 43(6) | You bear full responsibility for false or misleading AI content. You cannot blame the tool, and the court may pass any order it thinks fit against you. |
3. Using AI For Legal Research
The Regulation expressly recognises legal research, precedent retrieval, citation verification and document summarisation as legitimate uses of AI (Reg. 19(d)). So you may research with AI but “recognised” does not mean “unchecked.” Treat the AI as a junior who is fast, confident, and sometimes wrong, including about whether a case exists at all.
What you want to do | Allowed? — and what you must do | Quick example |
|---|---|---|
Get an overview of an area or a list of leading authorities | Yes, as a start. Use it to find leads, then read the actual judgments yourself. | Open each Section 138 NI Act case on SCC OnLine / Manupatra before relying on any. |
Cite a case the AI gave you | Only after you verify. Confirm the case exists; check parties, court, year, citation and the actual holding on a | Cannot find it on the official reporter or court site? Treat |
Handbook on AI in Day to Day Legal Practice · Shubhank Suman
What you want to do | Allowed? — and what you must do | Quick example |
|---|---|---|
primary database. | it as invented and drop it. | |
Use the AI’s summary of what a case held | With care. Read the ratio in the judgment itself; do not argue from the AI’s summary. | The AI says a case “held X”; you read it and find it held the opposite. |
Use a quotation the AI “extracted” from a judgment | Only after checking. Copy the quote from the judgment text yourself. | AI-supplied quotations are often reworded or wholly fabricated. |
Ask the AI for the latest legal position | Extra caution. Confirm the provision is in force and not amended, repealed or overruled. | AI knowledge is dated and weak on recent Indian law; confirm the current section text. |
4. When The AI Invents A Case - The Hallucination Trap
This is the single biggest risk in daily practice, and the Regulation names it directly. “Hallucination” is defined to include fabricated or wrong case precedents, statutory provisions, rules and legal facts (Reg. 3(1)(z)). Because the courts have formally recognised this risk, you cannot later claim you were taken by surprise.
What happens if you cite a fake case? If a document or piece of evidence is found false “by reason of its AI-generated character,” you bear full responsibility, you cannot rely on the AI as an excuse, and the court may pass such orders as it deems fit (Reg. 43(6)). In practice that can mean costs, adverse strictures on the record, an adverse inference, the document being withdrawn or disbelieved, contempt, or a reference to the Bar Council, quite apart from the damage to your client’s case and your own credibility. Courts have already acted on this: in the well-known US case Mata v. Avianca (2023), lawyers were sanctioned for citing ChatGPT-invented cases, and Indian courts are alert to the same problem.
What the AI produces | Why it is dangerous | What you do instead |
|---|---|---|
A confident, properly formatted citation that does not exist | Opposing counsel and the court cannot find it | Never cite anything you have not personally opened on a primary source |
A real case, but the wrong proposition | The judgment does not actually say what the AI claims | Read the ratio yourself; rely only on what you have read |
A real case with a fabricated quotation | The “quoted” lines are not in the judgment | Copy quotations directly from the judgment text |
A repealed or amended section quoted as current law | You argue on law that no longer exists | Confirm the bare provision from the current Act or an official source |
Handbook on AI in Day to Day Legal Practice · Shubhank Suman
What the AI produces | Why it is dangerous | What you do instead |
|---|---|---|
A mix of real and invented authorities in one list | The fakes hide among the genuine ones | Check every single citation, one unchecked entry is enough to sink the set |
5. Using AI For Drafting
AI is genuinely useful for structure, first drafts, standard clauses and routine formats, the Regulation itself lists drafting assistance and generation of prescribed formats among permitted uses. The danger is not in using it; it is in filing what it produces without reading it. AI will confidently insert facts, figures, clauses and citations that were never in your brief.
Drafting task | Allowed?-and the safeguard | Quick example |
|---|---|---|
Generate a first draft or structure of a pleading, notice or contract | Yes. Read it end to end and rewrite it to fit the matter. | Use the draft as a skeleton, never as a final product. |
Let the AI state facts, dates, figures or party details | No. Take all facts only from the record/brief; check every number. | The AI “fills in” a date or amount not in your file -delete it. |
Keep statutory references or citations the AI inserted | Only after verifying each. Confirm every section and case against the source. | Do not leave in a section number you have not checked. |
Use AI to tidy language, shorten or improve tone | Yes. Confirm the meaning has not changed. | Check that “simplifying” a clause has not changed your client’s obligation. |
Reuse an AI clause across matters | With care. Tailor it to the specific transaction and parties. | Generic boilerplate that does not fit the deal causes more problems than it saves. |
6. Translation, Transcription And Summarising Documents
The Regulation recognises transcription (with certification), translation (with human verification) and document summarisation as permitted uses (Reg. 19(b), (c), (d)). Each saves real time, but each has a failure mode that can quietly change meaning or hide the one fact that matters.
Everyday task | Allowed? -and what to do | Quick example |
|---|---|---|
Translate an order, pleading or document | Yes, with verification. Have a competent person check the translation against the original; one wrong word can change the meaning. | A mistranslated obligation flips who must pay. |
Handbook on AI in Day to Day Legal Practice · Shubhank Suman
Everyday task | Allowed? -and what to do | Quick example |
|---|---|---|
Transcribe a hearing, dictation or recording | Yes. Proofread and correct it before treating it as the record; certify its accuracy. | Auto-transcripts drop negatives - “not liable” becomes “liable.” |
Summarise a long judgment, brief or contract | Yes, to triage. Spot-check the summary against the source, especially the points that go against you. | The summary omits the one paragraph that hurts your case. |
Summarise a large set of documents or discovery | With care. Never let a summary replace reading any document that decides the matter. | AI condenses 500 pages and silently leaves out the key document. |
Rely on an AI summary to advise the client | Not on its own. Read the key documents yourself before you advise. | Advice built on a wrong summary is your error, not the AI’s. |
7. Confidentiality And Client Data
This is the mistake lawyers make most often and notice least: pasting privileged or identifiable client information into a free, public AI tool. Anything you type into a consumer chatbot may be stored, processed on outside servers, or used to train the model. The Regulation leans on the Digital Personal Data Protection Act, 2023 and the Information Technology Act, 2000, and treats data minimisation and protection of sensitive information as basic obligations (Reg. 10, 47, 48).
Do | Don’t |
|---|---|
Redact or anonymise names and identifiers before you type anything in | Paste privileged advice, draft pleadings or deal terms into free/public chatbots |
Use a paid/enterprise tool with a “no-training, no-retention” setting | Upload whole case files or client documents to consumer tools |
Read the tool’s data and privacy policy before first use | Assume “delete chat” means the data is gone or unused |
Treat anything you input as potentially disclosed to a third party | Enter sensitive personal data (health, financial, identity) casually |
8. Disclosure - When And How To Tell The Court
When disclosure is triggered | What you must disclose, and how | Provision |
|---|---|---|
You used AI to prepare or submit any document, pleading or evidence | That the material is AI-assisted - by a signed declaration in the prescribed format (Annexure I), at the time of filing. | Reg. 43(3) |
The court asks about your AI use | The tool used, the nature and extent of assistance, and how you verified it -by direct answer, on affidavit if directed. | Reg. 43(4) |
Handbook on AI in Day to Day Legal Practice · Shubhank Suman
When disclosure is triggered | What you must disclose, and how | Provision |
|---|---|---|
You tender AI-generated output as evidence | That it is AI-generated, fully and openly - as part of the tender; no silent filing. | Reg. 20(h) |
You use synthetic / AI-generated images, audio or data | That synthetic content is being used - in the form the court or authority prescribes. | Reg. 43(5) |
You file GenAI-generated content | Its origin, plus verification through the designated mechanism (AI Content Verification Authority). | Reg. 3(1)(y), 44 |
9. The Forbidden Zone - What AI Must Never Do
These prohibitions are aimed at courts and are absolute and non-derogable (Reg. 20). They matter to you for two reasons: you cannot ask a court to use AI in these ways, and AI output of this kind will not be accepted. Do not build a strategy around any of them.
AI must never be used to… | What it means for you |
|---|---|
Decide a case or reach any judicial outcome on its own | Only a judge decides; AI analysis is never the authority (Reg. 20(b)) |
Pass a sentence or adjudicate without a human in control | Sentencing and adjudication stay with the judge; AI is advisory only (Reg. 20(c)) |
Score a person’s risk -bail, flight risk, reoffending, credibility | You cannot seek or rely on AI “risk scores” for bail or to attack a witness (Reg. 20(d)) |
Predict or profile the future conduct of parties, witnesses or lawyers | No predictive profiling of the people in your case (Reg. 20(f)) |
Run surveillance or monitoring of judges, advocates or litigants | No AI surveillance in or around court (Reg. 20(g)) |
Use opaque “black-box” systems where rights or liberty are at stake | If it cannot be explained, it cannot be used in such matters (Reg. 20(e)) |
10. Do’s And Don’Ts - The Basic Mistakes, Fixed
Keep this one-minute card on your desk. It covers the slips that cause the most trouble in everyday work.
Do | Don’t |
|---|---|
Open every case, section and quote on a primary source before relying on it | Cite anything just because the AI sounded confident |
Read every AI draft fully before it leaves your desk | File or send AI text you have not read line by line |
Take all facts and figures only from the record | Let the AI supply facts, dates or numbers |
Verify each citation and statutory reference yourself | Trust the AI to “check” its own citations |
Handbook on AI in Day to Day Legal Practice · Shubhank Suman
Do | Don’t |
|---|---|
Confirm a provision is current and not amended or overruled | Assume the AI knows the latest law |
Redact client data before using public tools | Paste privileged or personal client information into free chatbots |
Disclose AI use in your filing and note the tool and how you checked it | Hide AI assistance or appear unable to explain it |
Disclose AI-generated or synthetic evidence openly | Pass off AI output as independent evidence |
Treat AI as an assistant whose work you supervise | Outsource analysis, strategy or the final call to the AI |
Slow down on matters affecting liberty or important rights | Rely on AI in high-stakes matters without close human review |
11. Your 60-Second Check Before You File
1. Have I personally verified every case, citation and section in this document on a primary source? 2. Are all facts and figures taken from the record, not from the AI? 3. Have I read the whole document myself and made it fit this matter? 4. Have I kept the client’s confidential data out of public tools? 5. Have I disclosed any AI assistance and noted which tool I used and how I checked it? 6. Am I prepared to stand behind every word as my own?
Handbook on AI in Day to Day Legal Practice · Shubhank Suman