Reordering an argument
Give it your own points and ask for a better order, or a tighter version of a paragraph you have already written. It is very good at this, and nothing here is an assertion of law, so nothing here can be a fabricated citation.
This is not a page about AI being dangerous. It is a page about one specific failure, why it happens, and the working method that removes it.
A general chatbot is good at structure, summarising and first drafts. It is unreliable on Indian case citations, because it produces text that looks like a citation rather than looking a judgment up. Courts take a serious view of citations that turn out not to exist. Use a general chatbot for shape, not for authority.
Refusing to admit this is why most warnings on this subject get ignored.
Give it your own points and ask for a better order, or a tighter version of a paragraph you have already written. It is very good at this, and nothing here is an assertion of law, so nothing here can be a fabricated citation.
Paste a clause you drafted and ask what it would mean to a lay reader. Useful for client letters, for spotting an ambiguity you had stopped seeing, and for explaining a position to somebody who is not a lawyer.
Procedural checklists, document lists for a filing, a list of questions for a first conference. It will not be complete and you will correct it, but a list you correct is faster than a list you write.
A rough Hindi or regional-language version of something you wrote, for you to correct. Useful for client communication. Not for anything that goes on record without a proper translation.
When you paste the document yourself, the material is in front of it and the failure mode is different — it may miss something, but it is not inventing a source. Subject to the confidentiality question further down this page.
Structure and headings for a document type you draft rarely. Fill in the law yourself. For Indian formats with the right recitals already in place, use the drafting formats library instead — 3,500+ across 61 categories.
An Indian citation has a very regular shape: two party names, a year, a reporter abbreviation, a page number. Producing something in that shape is easy. Producing something in that shape that is also a real judgment is a completely different operation — it requires looking one up. A general chatbot is doing the first and reading as though it did the second. The output is fluent, correctly formatted, and confident, and there is nothing in the text itself that tells you which citations are real. More on the mechanism in why AI invents case citations.
This one has become sharper since the criminal law renumbering. Ask about a familiar offence and you may get the old IPC section, the new BNS section, or a number that belongs to neither, with equal confidence in each case. The same applies across the BNSS and the BSA. Section numbers are exactly the kind of detail that is easy to produce plausibly and hard to produce correctly, and the consequences of getting one wrong in a filing are immediate. Our IPC to BNS mapping is the safer place to check.
Even where the citation is real and the section is right, a general tool has no view on whether the proposition still stands. It will cite a judgment that was overruled by a Constitution Bench without a flicker, because nothing in what it is doing involves checking subsequent history. This is the failure that costs cases rather than merely embarrassing, and it is the one good-law checking exists to solve.
Whatever produced a paragraph — a junior, a precedent from a previous matter, a chatbot, a research platform — the advocate who signs the filing is answerable for every assertion and every citation in it. That has always been the position and no tool changes it.
There have been reported instances, in India and elsewhere, of filings containing citations that could not be traced, and courts have taken a serious view of it. We are not going to summarise any particular order on this page, because the practical rule does not depend on one and because we are a software company, not your advisers. Lawgger is not a law firm and gives no legal advice.
The rule that survives every version of this problem is short: nothing goes into a filing that you have not opened and read. A platform can make that fast — a pin-cite that opens on the paragraph, a verdict that tells you the authority still stands — but it cannot do it for you. What our platform commits to on accuracy, confidentiality and audit is set out on our trust page.
Ask these of any tool, ours included, before privileged material goes into it.
A general consumer tool and an enterprise arrangement can have very different answers. Read the vendor's own terms rather than assuming — the default for a free consumer account is rarely the strictest option.
This is the question that matters most for privileged material. If the answer is yes, or unclear, or configurable-but-not-configured, privileged material should not go in.
Ask about access controls and whether there is a record of who looked at what. A chamber's files should not be readable by anybody who happens to work at a software company.
Lawgger's position, plainly: chamber data sits in private per-chamber containers, encrypted with TLS 1.2 or above in transit and AES-256 at rest, with a hash-chained tamper-evident audit trail. It is never used as AI training data. Our practices are aligned to ISO/IEC 27001 — aligned, not certified, and we will not claim otherwise — and built for DPDP Act 2023 readiness. The detail is on our trust page and in our privacy policy.
| A general chatbot | Lawgger | |
|---|---|---|
| Where the answer comes from | Written from patterns in text | Generated only from judgments in the collection |
| Pin-cite to the paragraph | Not offered | Every answer |
| Good-law verdict | Not offered | Every authority |
| Indian court coverage | Not a defined collection | 2.07 crore+ judgments, SC and all 25 High Courts |
| Behaviour when it does not know | Usually answers anyway | Says so rather than filling the gap |
| Cause lists and deadlines | Not offered | Cause lists, display boards, 21 limitation clocks |
| Confidentiality posture | Depends entirely on the account and plan — read the terms | Private per-chamber containers; never used as AI training data |
| Price | Free to modest monthly tiers | ₹1,999 / month · ₹19,999 / year |
You do not have to give up the tool that helps you write. You have to be strict about which step it is allowed to touch.
Order your points, tighten a paragraph, produce an outline. Nothing at this stage is an assertion of law, so nothing at this stage can be a fabricated citation.
Move to a platform that answers from actual judgments and names the judgment and the paragraph. If a tool cannot show you where an answer came from, treat it as a hypothesis. See how research works.
The step that cannot be delegated. A proposition that survives your own reading of the paragraph is safe to use; one that does not, is not.
Confirm the authority still stands. Every authority carries a verdict, and Authority Check runs the same test across every citation in a brief you paste in.
Applied consistently, this one rule removes the fabricated-citation risk entirely, whatever produced the first draft of the paragraph.
Nothing stops an advocate using a general chatbot as a drafting aid, and many do. The duty that does not move is verification: the advocate who signs the filing is answerable for every assertion and every citation in it, regardless of what produced the words. Use it for shape and structure, not as a source of authority.
No, not as a source of citations. A general chatbot produces text that resembles a citation rather than looking a judgment up. Case names, reporter references, years and paragraph numbers can all be plausible and all be wrong, and the answer will read exactly as confident either way. Treat every citation from a general tool as unverified.
Because it is completing a pattern, not consulting a library. Indian citations have a very regular shape — party names, a year, a reporter abbreviation, a page. Producing something in that shape is easy. Producing something in that shape that is also a real judgment requires actually looking one up, which is a different operation. More in why AI invents citations.
There have been reported instances of filings containing citations that could not be traced, in India and in other jurisdictions, and courts have taken a serious view of it. We do not summarise any particular order here, because the practical rule does not depend on one: do not put an authority in a filing that you have not opened and read yourself.
Ask three questions before you paste anything: is the material retained, is it used to improve the service, and who inside the vendor can see it. If you cannot answer all three from the vendor's own terms, do not paste privileged material. Lawgger keeps chamber data in private per-chamber containers, encrypted in transit and at rest, and never uses it as AI training data — see our trust page.
Where the answer comes from. A general chatbot writes an answer. A legal research platform answers from judgments it can show you, names the judgment and the paragraph, and tells you whether that authority still stands. The difference only becomes visible at the moment you try to check a citation, which is usually the worst moment to discover it.
Lawgger answers only from judgments in its collection and pin-cites every answer to the page it came from, so you can open the paragraph and read it. Where it cannot support an answer from a judgment, it says so rather than filling the gap. We still expect you to read the paragraph before you cite it — that duty is yours and always will be.
Practice varies and some courts have begun issuing directions on the point, so check the practice directions of the court you are before. Whatever the position on disclosure, it does not reduce the verification duty. An accurate, checked citation raises no issue; an unchecked one raises the same issue whether or not it was disclosed.
Yes, and that is the sensible arrangement. Use a general tool for structure, plain-English explanation and first drafts of non-authoritative text. Use Lawgger for anything that becomes an assertion of law: the authority, the pin-cite, the good-law verdict, and Authority Check across the citations in an opponent's brief.
Fourteen days free, no card — enough to check your own authorities before you file. Then ₹1,999 a month or ₹19,999 a year.