# AI judgment summarisation: 200-page verdicts to headnotes

**TL;DR:** A good AI summary saves you the hours it takes to read a 200-page verdict, but a bad one can quietly invert the holding or mis-state the ratio decidendi and you will not know until a judge corrects you. The safe pattern is simple. Use the summary as a map, not a substitute. Read the AI headnote, jump to the paragraph numbers it flags, and verify the rule in the original text before you rely on it. Extractive and grounded summaries that quote the judgment and cite paragraphs are far safer than free-form abstractive ones, which research shows still confuse parties, drop conditions, and hallucinate. Never outsource the actual legal conclusion. That stays with you.

---

## On this page

- [Why judgments are long and hard to read](#why-judgments-are-long-and-hard-to-read)
- [What a good headnote actually contains](#what-a-good-headnote-actually-contains)
- [How human editors write headnotes, and what AI can learn](#how-human-editors-write-headnotes-and-what-ai-can-learn)
- [Extractive vs abstractive summarisation, explained plainly](#extractive-vs-abstractive-summarisation-explained-plainly)
- [The real danger: an AI that inverts the holding](#the-real-danger-an-ai-that-inverts-the-holding)
- [A worked example: summarising a bail order](#a-worked-example-summarising-a-bail-order)
- [Grounding every sentence to paragraph numbers](#grounding-every-sentence-to-paragraph-numbers)
- [How to use an AI summary safely](#how-to-use-an-ai-summary-safely)
- [What you must never outsource](#what-you-must-never-outsource)
- [Quality checks before you rely on a summary](#quality-checks-before-you-rely-on-a-summary)
- [The access-to-justice upside](#the-access-to-justice-upside)
- [How Niyam approaches grounded summarisation](#how-niyam-approaches-grounded-summarisation)
- [Frequently asked questions](#frequently-asked-questions)
- [Start for ₹100](#start-for-100)

---

## Why judgments are long and hard to read

Indian judgments have a length problem, and everyone in practice knows it. The 2019 Ayodhya verdict ran to [1,045 pages](https://en.wikipedia.org/wiki/2019_Supreme_Court_verdict_on_Ayodhya_dispute). The 2015 NJAC judgment that struck down the National Judicial Appointments Commission ran past a thousand pages. The 2018 Aadhaar judgment in Puttaswamy crossed 1,400 pages. Even the foundational Kesavananda Bharati judgment from 1973 ran to roughly 700 pages and was famously hard to read, with eleven separate opinions that lawyers still argue about.

These are the headline-grabbing constitutional cases, but the length problem is not limited to them. A routine commercial appeal or a service matter can easily run to 60, 100, or 150 pages. There are good reasons for this. A judgment is not an essay. It is a public record that has to do several jobs at once. It sets out the facts, often in painful detail because the facts decide the case. It records the arguments of both sides, sometimes verbatim. It works through the statutes, the precedents cited, and the precedents distinguished. It reasons towards a conclusion, and under our system that reasoning is part of the law itself.

The trouble is that this density makes judgments slow to read and easy to misread. Legal commentators have long observed that the prose in many judgments is hard going even for trained lawyers, which is part of why the [Supreme Court Observer](https://www.scobserver.in/journal/judgments-in-plain-language/) and others keep pushing for plain-language writing. When the public reasoning function of a court depends on the judgment being intelligible, and the judgment is 200 pages of dense reasoning, something has to give.

What gives, in practice, is time. A junior associate handed a stack of long judgments to digest before a conference is doing one of the least glamorous and most error-prone jobs in litigation. Read fast and you miss the buried condition. Read slow and you blow the deadline. This is exactly the gap that AI summarisation is meant to fill, and exactly where it can do real damage if it is built carelessly.

## What a good headnote actually contains

Before we ask what an AI summary should look like, it helps to remember what a human headnote already does, because the headnote is centuries old and the editors who write them are very good at it.

A headnote in a law report is a defining feature of the report. According to the [Incorporated Council of Law Reporting](https://www.iclr.co.uk/knowledge/glossary/ratio-decidendi/), a headnote should ideally contain an accurate and authoritative statement of the ratio decidendi, usually set out in one or more paragraphs beginning with the word "Held". As the editors at [Counsel Magazine](https://www.counselmagazine.co.uk/articles/the-art-craft-of-the-headnote) describe it, most modern headnotes have two parts. The first sets out the essential facts and issues. The second explains, as briefly as possible, the legal conclusion.

In Indian practice, the editors at SCC, AIR, and the other reporters write headnotes the same way, and the Supreme Court's own e-SCR project now publishes headnotes too. Those e-SCR headnotes are, importantly, [approved by the judges](https://www.millenniumpost.in/big-stories/cji-announces-launch-of-electronic-supreme-court-reports-e-scr-project-to-provide-access-to-judgements-504141) who delivered the judgment, which is about as authoritative as a summary gets.

Here is what a complete headnote captures, and what any AI summary should aim to reproduce.

| Element | What it captures | Why it matters |
|---|---|---|
| Facts | The material facts the court relied on | The ratio is tied to these facts; change them and the rule may not apply |
| Issues | The questions of law the court had to decide | Tells you whether the case is even on point for your matter |
| Holding | What the court decided on each issue | The actual outcome, appeal allowed or dismissed and on what terms |
| Ratio decidendi | The rule of law the court had to accept to reach the result | This is the binding part under Article 141; everything else is obiter |
| Obiter dicta | Observations not necessary to the decision | Persuasive, not binding; useful but must be labelled as such |
| Citations referred | Precedents and statutes relied on, distinguished, or overruled | Tells you the chain of authority and whether good law was disturbed |

The single hardest item on that list, for a human and for a machine, is the ratio decidendi. The ratio is not a sentence you can copy out. It is the rule the court was *required* to accept to reach its result, and isolating it takes legal judgment. We have written a full guide on [how to read and brief an Indian judgment](/blog/how-to-read-a-judgment) that walks through finding the ratio by hand. The short version is that the ratio lives in the relationship between the material facts and the conclusion, and a summary that gets the facts right but states the rule too broadly or too narrowly has failed at the one job that matters most.

This is the standard to hold an AI summary against. Not "is it readable" but "does it state the holding correctly and the ratio faithfully, and does it tell me which paragraphs to check".

## How human editors write headnotes, and what AI can learn

It is worth pausing on how a skilled law-report editor actually produces a headnote, because the craft tells you exactly where an AI is likely to go wrong.

An editor at SCC or AIR reads the whole judgment first. They do not start writing until they understand what the case decided, and crucially, what the case decided that is *new* or *clarifying* enough to be worth reporting. Many judgments simply apply settled law to fresh facts and never get a full headnote at all. The editor's first act of judgment is deciding what is reportable.

Then they separate the operative reasoning from everything around it. A judgment is full of material that is not the ratio. There are pages of facts, the arguments of counsel set out at length, summaries of the precedents cited, and observations the court makes in passing. The editor's skill is in walking through all of that and pulling out the rule the court was required to accept, then stating it tightly, usually beginning with "Held". As [Counsel Magazine](https://www.counselmagazine.co.uk/articles/the-art-craft-of-the-headnote) notes, a headnote sometimes states a bald proposition of law and sometimes pairs it with the circumstances in which it arises, but in either form the editor is making a deliberate choice about how widely or narrowly to state the rule.

That choice is the hardest part of the job, and it is exactly where an unguided AI struggles. State the rule too narrowly and you bury a principle that lawyers needed to find. State it too widely and you turn a fact-bound decision into a general law the court never intended. Human editors get this right because they have read tens of thousands of judgments and they understand the consequence of pitching a rule at the wrong level. They are also accountable. An editor who consistently mis-states ratios produces a law report nobody trusts.

There are three things a good AI summary should borrow from this craft. First, read the whole judgment before deciding what matters, rather than summarising the first chunk that fits the context window. Second, distinguish operative reasoning from recited arguments, which is the discipline that prevents the most common attribution errors. Third, pitch the rule at the level the court actually pitched it, preserving conditions and qualifications rather than smoothing them away for readability. An AI cannot replace the editor's accumulated judgment. It can, if built carefully, imitate the editor's method, and it can do something the editor's printed headnote cannot, which is hand you the paragraph numbers so you can check its work yourself.

## Extractive vs abstractive summarisation, explained plainly

There are two broad ways a machine can summarise a judgment, and the difference between them is the single most important thing to understand before you trust any summary.

**Extractive summarisation** picks sentences directly out of the judgment and stitches them together. Nothing is rewritten. If the summary says the court held something, that exact sentence appears somewhere in the original. The strength is fidelity. Because every line is the court's own words, there is much lower risk of misinterpreting a key legal phrase, a statute, or a cited case. The weakness is that extractive summaries can read like a clipped highlight reel, and they cannot smooth two paragraphs into one clean statement.

**Abstractive summarisation** generates new sentences that paraphrase the judgment, the way a human editor writes a headnote in their own words. The strength is fluency. A good abstractive summary reads like prose and can compress an argument that is spread across ten paragraphs into two clean lines. The weakness is the dangerous one. Because the model is generating new text, it can generate text that is not actually true to the judgment. It can paraphrase a conditional holding as an unconditional one. It can attribute an argument to the wrong party. It can invent a qualification the court never made.

| | Extractive | Abstractive |
|---|---|---|
| How it works | Selects real sentences from the judgment | Generates new sentences that paraphrase |
| Reads like | A highlight reel of the court's own lines | A polished, human-written headnote |
| Main risk | Choppy, may miss connective reasoning | Hallucination, mis-statement, wrong attribution |
| Verbatim accuracy | High, words are the court's own | Lower, words are the model's |
| Best for | Legal text where exact phrasing is non-negotiable | Readability, when checked by a human |
| Verification | Easy, the source line is right there | Hard, you must compare against the original |

This is not just a theoretical preference. The most directly relevant research comes from a study of Indian court judgments specifically. In [How Ready are Pre-trained Abstractive Models and LLMs for Legal Case Judgement Summarization?](https://arxiv.org/abs/2306.01248), the researchers ran several abstractive models and general-purpose large language models over Indian judgments and checked the output for inconsistencies and hallucinations. Their finding was blunt. Abstractive models often scored slightly higher on automatic metrics like ROUGE, but the summaries contained inconsistent or hallucinated information, and some of the errors were subtle and hard to catch automatically. One example they flagged was the model confusing the name of an appellant with the name of the lawyer representing the appellant. In a legal document, that is not a typo. That is a factual error that can mislead anyone who relies on the summary.

Their conclusion was that these models are not ready for fully automatic deployment, and that a human-in-the-loop approach with manual checks is the appropriate posture today. That conclusion, from researchers working on Indian judgments, is the foundation of everything in this guide.

The practical takeaway is not "abstractive is bad and extractive is good". A well-built system uses abstractive fluency for readability and extractive grounding for trust, and it always shows you where each claim came from. The danger is a pure abstractive summary with no anchors, presented as if it were authoritative.

## The real danger: an AI that inverts the holding

Let us be specific about the failure mode, because "hallucination" is an overused word and people have stopped hearing it.

Imagine a bail matter. The High Court hears a petition under Section 482 of the BNSS, considers the gravity of the allegations and the stage of investigation, and grants bail subject to three conditions. An abstractive summary, trying to be concise, might produce: "The Court held that bail should be granted in such cases." That sentence is a disaster. It has dropped the conditions, dropped the fact-specific reasoning, and converted a narrow, conditional, fact-bound order into what reads like a general rule. A junior who cites that summary in a different matter is citing something the court never said.

Now imagine a worse case. A summary that inverts the holding. The court dismisses an appeal, but a poorly grounded model, confused by a long passage where the court summarised the appellant's own arguments before rejecting them, reports that the appeal was allowed. This is not far-fetched. Judgments routinely spend pages setting out arguments the court is about to reject, and a model reading sequentially can latch onto the argument and miss the rejection. The appellant's case and the court's holding can look very similar on the page right up until the line where the court says "we are unable to accept this submission".

These risks are not unique to free chatbots. The most rigorous study on commercial legal AI to date, the Stanford RegLab paper [Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools](https://arxiv.org/abs/2405.20362), tested purpose-built, retrieval-augmented legal research products and still found hallucination rates in the range of roughly 17 to 33 percent depending on the tool. The researchers' conclusion was that providers' claims of being hallucination-free were overstated. If specialist tools with retrieval over curated databases hallucinate that often on research tasks, a raw summary of a long judgment with no grounding is not something to trust on its face.

The three failure modes to watch for, in order of how often they bite:

1. **Dropped conditions and qualifications.** The summary states a holding but omits the "subject to", "only where", or "in the facts of this case" that limits it.
2. **Wrong attribution.** An argument made by a party, or a position the court was describing in order to reject, is reported as the court's own holding.
3. **Inverted or overbroad ratio.** The narrow rule the court actually laid down is widened into a general proposition, or the outcome itself is flipped.

Every one of these is invisible if you only read the summary. Every one of them is obvious the moment you read the cited paragraph in the original. That is the whole case for grounding.

## A worked example: summarising a bail order

To make the failure modes concrete, walk through a single hypothetical and watch where a careless summary goes wrong.

Take a High Court order on an application for anticipatory bail. The facts: the applicant is accused in a cheating matter, the investigation is at an early stage, custodial interrogation is not shown to be necessary, and there is no allegation that the applicant is a flight risk or likely to tamper with evidence. The court, after considering the gravity of the allegations against these specific factors, grants anticipatory bail subject to the applicant joining the investigation when called and not leaving the country without permission. The operative conclusion sits at, say, paragraph 18. Paragraphs 6 to 12 set out the prosecution's arguments about the seriousness of the offence, which the court records faithfully before declining to give them decisive weight.

Now look at three summaries of the same order.

| Version | What it says | Verdict |
|---|---|---|
| Overbroad abstractive | "The Court held that anticipatory bail should be granted in cheating cases." | Dangerous. Drops every condition and the fact-specific reasoning; reads like a general rule the court never laid down. |
| Mis-attributed | "The Court emphasised the gravity of the offence and the seriousness of the allegations." | Wrong. That is the prosecution's argument from paras 6 to 12, not the court's holding at para 18. |
| Grounded | "Held (para 18): anticipatory bail granted, given the early stage of investigation, absence of flight risk, and no need for custodial interrogation, subject to the applicant joining the investigation and not leaving the country without leave." | Safe. States the conditional holding, ties it to the operative paragraph, preserves the qualifications. |

The first version is what you get when a model optimises for a clean, quotable sentence. The second is what you get when a model reads sequentially and latches onto the most forceful passage, which is often counsel's argument rather than the court's conclusion. The third is what grounding produces, and the reason it is safe is not that the prose is better. It is that it tells you to look at paragraph 18, where you can confirm every word in under a minute.

This is also why a single judgment can support several different summaries depending on what you need. A summary for a lawyer arguing a similar bail matter should foreground the conditions and the factors the court weighed. A summary for someone checking whether the case is even relevant can be shorter. What never changes is the anchor. Every version points back to the operative paragraph, so the reader is never asked to trust the summariser over the court.

## Grounding every sentence to paragraph numbers

Grounding is the discipline that separates a summary you can work with from a summary you have to fear. A grounded summary does not just tell you what the court held. It tells you *where* the court held it, by paragraph number, so you can check in seconds.

In the research literature this is called attribution or citation-enforced generation. The idea, described in work like [Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions](https://arxiv.org/abs/2409.11242), is that every factual claim in the output should be traceable to a specific passage in the source, and the system should be built so that claims without support are flagged or refused rather than invented. When a claim is supported by the cited passage, it is grounded. When it is not, it is hallucination, full stop.

For judgment summarisation, grounding means a few concrete things:

- Every sentence in the summary points to the paragraph or paragraphs it came from.
- The holding sentence cites the exact paragraph where the court stated its conclusion, not a paragraph where it was summarising arguments.
- Citations referred are pulled from the judgment's own text, so the summary cannot invent a case the court did not mention.
- Where the summary paraphrases, you can click through to the verbatim source and check the paraphrase yourself.

This matters enormously in Indian practice now that the neutral citation system is in place. The Supreme Court and High Courts assign paragraph-numbered neutral citations like 2024 INSC 835, which means a paragraph reference is stable and verifiable across every platform, as [Bar and Bench](https://www.barandbench.com/news/ending-citation-chaos-neutral-citation-simplifies-legal-referencing-in-indian-courts) has explained. A grounded summary that says "Held, at para 47" is making a checkable promise. A summary that just says "the Court held" is asking for blind trust.

The difference shows up the moment something looks off. With a grounded summary, when a holding reads strangely, you jump to the flagged paragraph and either confirm it or catch the error in under a minute. With an ungrounded summary, you have to re-read the whole judgment to find out whether the summary lied to you, which defeats the entire purpose of summarising it.

## How to use an AI summary safely

Here is the workflow that turns AI summarisation from a liability into a genuine time-saver. It is built around the human-in-the-loop principle the Indian research recommends, and it takes discipline, not extra hours.

**Step one: read the summary as a map.** Treat the AI headnote as an index to the judgment, not as the judgment. Its job is to tell you what is in there and roughly where, so you can decide whether this case is even relevant to your matter and which parts you need to read closely.

**Step two: jump to the flagged paragraphs.** For every claim that matters to you, especially the holding and the ratio, go to the paragraph the summary cites and read the court's own words. If the summary is grounded, this is a click. If it is not grounded, that is your signal to be far more careful or to discard the summary.

**Step three: verify the ratio in the original.** This is the step you cannot skip. Read enough of the surrounding paragraphs to confirm what rule the court actually laid down, on what facts, and subject to what conditions. Check that what the summary called the holding is the court's conclusion and not a party's argument. Confirm the appeal outcome. Confirm any qualifications.

**Step four: check the citations referred.** If you are going to rely on this judgment, you also need to know whether it relied on good law. Use the summary's list of cited cases as a starting point and confirm none of them have since been overruled. Our guide on [good law checking](/blog/good-law-checking) covers how to do this properly, and our piece on [how to cite Indian judgments](/blog/how-to-cite-indian-judgments) covers getting the citation itself right.

**Step five: write your own note.** When you record the case for your file or your brief, write the holding and ratio in your own words, having read the original. Do not paste the AI summary into your work product. The summary got you to the right paragraphs faster. Your understanding, verified against the text, is what goes into the brief.

| If the summary... | Then you should... |
|---|---|
| Cites a paragraph for the holding | Open that paragraph and confirm the outcome and conditions |
| States a broad rule with no qualifier | Be suspicious; judgments are usually fact-bound, check the original |
| Names a party as winning or losing | Confirm against the final operative paragraph of the judgment |
| Lists cases the judgment relied on | Verify those cases are still good law before you cite them |
| Has no paragraph anchors at all | Use it only to orient yourself, never to support a proposition |

Used this way, a summary that takes you ten seconds to read and two minutes to verify replaces an hour of cold reading. That is the real gain, and it is large. The mistake is to stop after step one.

## What you must never outsource

There is a clean line here, and it is worth stating plainly. You can outsource reading speed. You cannot outsource legal judgment.

An AI can tell you what a paragraph says. It cannot tell you, reliably and on its own authority, what the binding rule of the case is, because isolating the ratio decidendi requires deciding which facts were material and which parts of the reasoning were necessary to the result. That is the core skill of a lawyer, and it is the part that carries professional responsibility. Courts and bar regulators have been increasingly clear that the duty to verify AI output sits with the advocate, a point we have covered in detail in [the lawyer's duty to verify AI output](/blog/lawyer-duty-verify-ai-output).

So the things to never hand over without verification are:

- **The legal conclusion.** What the case holds, as a matter of binding law, is your call, made after reading the court's own words.
- **The ratio.** What rule the court was required to accept. A summary can point you to it; it cannot decide it for you.
- **Whether the case is good law.** Subsequent treatment, overruling, and per incuriam status are matters of judgment over a chain of authority.
- **Whether the case applies to your facts.** A judgment is binding only to the extent its material facts match yours. Only you know your facts.

The summary is a tool that gets you to the place where you exercise judgment faster. It is not a substitute for the judgment itself, and it is certainly not a substitute for yours.

## Quality checks before you rely on a summary

Whether you are evaluating a summarisation tool or sanity-checking a single summary, run it against these checks. They are quick, and they catch the failure modes that matter.

| Check | Question to ask | Red flag |
|---|---|---|
| Outcome | Does the summary state who won and on what terms? | Vague "the Court considered" with no clear result |
| Holding anchor | Is the holding tied to a specific paragraph? | A holding with no paragraph reference |
| Conditions | Are qualifications and conditions preserved? | A clean general rule where the judgment was fact-bound |
| Attribution | Are arguments labelled by who made them? | The court's view and a party's argument blurred together |
| Citations | Do cited cases actually appear in the judgment? | A neat list you cannot find in the original text |
| Verbatim access | Can you click through to the court's own words? | No way to see the source line behind a claim |
| Hedging | Does the tool flag low-confidence claims? | Uniform confidence across everything, including hard calls |

A summarisation tool worth using will pass most of these by design, because grounding and attribution are architectural choices, not afterthoughts. A tool that fails several of them is a tool that is asking you to trust paraphrase over text, which the research says you should not do.

One more practical test. Take a judgment you already know well, ideally one with a conditional or narrow holding, and summarise it. If the tool preserves the conditions and points you to the right paragraphs, that tells you a lot. If it produces a confident, clean, overbroad rule, that tells you more.

## The access-to-justice upside

It would be easy to read all of this as a case against AI summarisation. It is not. It is a case for doing it properly, because done properly the upside is real and it reaches further than the big law firms.

Indian judgments are now more accessible than they have ever been. The Supreme Court's e-SCR project put over [34,000 judgments online for free](https://www.millenniumpost.in/big-stories/cji-announces-launch-of-electronic-supreme-court-reports-e-scr-project-to-provide-access-to-judgements-504141), with search, and the Court has gone further by translating thousands of judgments into more than a dozen Indian languages. Free portals and the National Judicial Data Grid have opened the door wider still. Access to the raw text is no longer the bottleneck.

The bottleneck now is time and comprehension. A sole practitioner in a district court does not have a team of associates to digest a 150-page judgment overnight. A law student trying to understand a constitutional bench decision is staring at hundreds of pages of dense prose across multiple opinions. A litigant in person has even less to work with. For all of them, a grounded summary that turns a 200-page verdict into a checkable two-page map, with paragraph anchors so they can read the parts that matter, is a genuine leveller. It does not replace the reading. It makes the reading possible within the time and resources people actually have.

That is the promise worth protecting. And the only way to protect it is to keep the summaries honest, which means grounded, checkable, and never presented as a final answer. A summary that quietly misleads a well-resourced firm is a problem. A summary that quietly misleads a litigant in person who has no way to catch the error is a much bigger one. The bar for trust has to be set by the user with the least ability to verify.

## How Niyam approaches grounded summarisation

Niyam is built around the principle this whole guide argues for. A summary is only useful if you can check it, so every summary is built to be checked.

When Niyam summarises a judgment, it works over the verbatim text of the judgment and anchors its claims to paragraph numbers, so the holding and the ratio point you straight to where the court said it. The summary is a map into the judgment, not a replacement for it. You read the headnote, you click through to the court's own words for anything you intend to rely on, and you verify the ratio in the original, which is exactly the safe workflow described above. Niyam works over India's own corpus of Supreme Court and High Court judgments with neutral citations, so paragraph references are stable and verifiable, and the system is designed to surface the cases a judgment relied on rather than invent ones it did not.

It pairs naturally with the rest of the workflow. When a summary points you to a relevant case, you can use [AI to find similar judgments](/blog/ai-find-similar-judgments) to widen the net, and you check that anything you cite is still good law before you rely on it. The summary saves you the cold-reading time. The judgment, and your reading of it, remains the authority.

The honest framing matters to us. We do not claim a summary removes the need to read the judgment. We claim it gets you to the paragraphs that matter far faster, and that it never asks you to trust a paraphrase you cannot trace. That is the difference between a tool that helps and a tool that quietly creates risk.

## Frequently asked questions

**Can I cite an AI-generated headnote in court?**

No. You cite the judgment, using the court's own words and the correct neutral citation, after reading the relevant paragraphs yourself. A headnote, even a human-written law-report headnote, has no authoritative value because it is not written by the court. The Supreme Court's e-SCR headnotes are an exception because they are approved by the judges who delivered the judgment, but an AI summary is not. Use the summary to find the paragraph, then cite the paragraph.

**Is extractive summarisation always safer than abstractive?**

Extractive summaries carry lower hallucination risk because every line is the court's own words, which is why they are preferred where verbatim accuracy is non-negotiable. But extractive summaries can be choppy and can miss connective reasoning. The best systems combine abstractive readability with extractive grounding and paragraph anchors, so you get a readable summary that you can still trace back to the exact source. Safety comes from grounding and verification, not from the technique alone.

**How often do legal AI tools actually get summaries wrong?**

Often enough that you cannot rely on them blind. Research on Indian judgment summarisation found abstractive models producing inconsistent and hallucinated information, including confusing an appellant with their lawyer. The Stanford RegLab study of leading commercial legal research tools found hallucination rates of roughly 17 to 33 percent on research tasks despite retrieval grounding. The right response is not to avoid the tools but to verify their output against the original text every time.

**What is the single most dangerous summarisation error?**

Mis-stating the holding or the ratio, especially by dropping a condition or inverting the outcome. A narrow, fact-bound, conditional order summarised as a clean general rule is the error that causes lawyers to cite cases for propositions the court never laid down. It is also invisible if you only read the summary, which is why jumping to the cited paragraph and verifying the ratio is non-negotiable.

**Why do paragraph numbers matter so much?**

Because they make the summary checkable. A grounded summary that ties its holding to "para 47" is making a promise you can confirm in seconds. The neutral citation system used by Indian courts assigns stable paragraph numbers, so a paragraph reference is verifiable across every platform. A summary with no paragraph anchors is asking for blind trust, which is exactly what you should not give it.

**Does a good AI summary mean I do not have to read the judgment?**

No. It means you do not have to cold-read the whole judgment to find the parts that matter. You read the summary to orient yourself, then you read the paragraphs that carry the holding, the ratio, and anything you intend to rely on. The summary saves the reading you do not need. It does not remove the reading you do.

**Who is responsible if an AI summary is wrong and I rely on it?**

You are. The professional duty to verify AI output sits with the advocate, and courts have been increasingly firm about this. The tool is an aid. The legal conclusion, and the responsibility for it, stays with the lawyer who signs the brief. This is precisely why grounded, checkable summaries matter, because they make the verification fast enough to actually do every time.

## Start for ₹100

If you want judgment summaries you can actually trust, the test is simple. Can you check them? Niyam summarises Indian Supreme Court and High Court judgments with paragraph anchors and verbatim source access, so every headnote is a map you can verify against the court's own words, not a paraphrase you have to take on faith.

Try it on a judgment you already know. See whether it preserves the conditions, points you to the right paragraphs, and keeps the ratio intact.

[Start for ₹100](https://app.niyam.ai/register) and turn your next 200-page verdict into a headnote you can stand behind.