# Legal research in Hindi and Indian languages: AI opens the law

# Legal research in Hindi and Indian languages: how AI is opening access to the law

**TL;DR:** For most of independent India's history, the law has spoken English to a country that mostly does not. Article 348 of the Constitution fixes English as the language of the Supreme Court and the High Courts, and only four High Courts are formally authorised to also use Hindi. A litigant who has lost a house, a job, or custody of a child often cannot read the judgment that decided it. That is now changing. The Supreme Court has translated tens of thousands of its judgments into Hindi and other scheduled languages using its own AI tool, SUVAS, and a new generation of India-built language models can read, search, and explain legal text across the 22 scheduled languages. This piece walks through the English barrier, what the Constitution actually permits, the official translation drive and its real figures, how Indic LLMs handle messy real-world legal language, where machine translation still goes wrong, and what all of this unlocks for ordinary people - and where human verification stays non-negotiable.

---

## On this page

- [The English wall around Indian law](#the-english-wall-around-indian-law)
- [Article 348: why the court speaks English](#article-348-why-the-court-speaks-english)
- [Which courts may use Hindi and regional languages](#which-courts-may-use-hindi-and-regional-languages)
- [The Supreme Court's translation drive: SUVAS and e-SCR](#the-supreme-courts-translation-drive-suvas-and-e-scr)
- [SUPACE and AI inside the courtroom](#supace-and-ai-inside-the-courtroom)
- [How Indic language models handle legal text](#how-indic-language-models-handle-legal-text)
- [Code-mixing, transliteration, and the Hinglish problem](#code-mixing-transliteration-and-the-hinglish-problem)
- [Where machine translation of legal text breaks](#where-machine-translation-of-legal-text-breaks)
- [What vernacular legal research actually unlocks](#what-vernacular-legal-research-actually-unlocks)
- [The limits: why a human still has to verify](#the-limits-why-a-human-still-has-to-verify)
- [How Niyam approaches Indian-language legal research](#how-niyam-approaches-indian-language-legal-research)
- [Frequently asked questions](#frequently-asked-questions)

---

## The English wall around Indian law

Walk into almost any district court in India and you will hear the case argued, in part, in the local language. The witness deposes in Bhojpuri or Tamil or Marathi. The lawyer cross-examines in a mix of that language and English. The judge follows all of it. Then the order comes out, and it is in English.

That gap is the quiet centre of India's access-to-justice problem. The person whose life the order changes - the tenant, the daily-wage worker, the widow chasing a pension - frequently cannot read it. They depend entirely on their advocate to tell them what happened, and on the advocate's reading being both correct and honestly relayed.

Former Chief Justice of India D.Y. Chandrachud put the scale of it bluntly. As [reported by LiveLaw](https://www.livelaw.in/news-updates/supreme-court-judgments-translated-four-regional-languages-english-citizens-cji-dy-chandrachud-219840), he said English in its "legal avatar" is not comprehensible to 99.9 per cent of citizens, and that access to justice cannot be meaningful unless people can understand proceedings in a language they actually speak. His successor as Chief Justice, Surya Kant, has framed it as a democratic right, telling audiences that citizens are entitled to read court judgments in their own language, as [covered by The Researchers](https://www.theresearchers.us/2025/12/21/cji-judgments-own-language/).

This is not a small or fringe issue. It sits at the join between two facts about India: the legal system inherited an English-language spine from the colonial period, and the population it serves speaks more than 1,300 mother tongues, organised constitutionally around 22 scheduled languages. For decades, the only bridge across that gap was the lawyer. AI is now building a second bridge - imperfect, supervised, but real.

A useful way to hold the whole picture is this table of where the language barrier bites.

| Stage of a legal matter | Language reality | Who is shut out |
| --- | --- | --- |
| Oral arguments in trial courts | Often the local language plus English | Few - this is the most accessible stage |
| Pleadings and written submissions | Mostly English | Litigants who cannot read English |
| The judgment or final order | English (Article 348) | The vast majority of citizens |
| Reported precedent and case law | Almost entirely English | Litigants, paralegals, law students from non-English backgrounds |
| Legal research and "find similar cases" | English-first databases | Anyone researching in a regional language |

The judgment line and the research line are where the wall is highest. They are also where AI is making the fastest progress.

It helps to be precise about how big the affected population is. India does not have one regional language; it has twenty-two scheduled languages and, by the most recent census tabulations, more than a thousand mother tongues spoken by meaningful numbers of people. Hindi in its various forms is the first language of a large plurality, but hundreds of millions of citizens think and live primarily in Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Odia, Punjabi, Assamese, and others. English is, for the overwhelming majority, a second or third language at best, and formal legal English is a specialised dialect that even fluent English speakers find heavy going. When a system speaks only legal English, the number of people it speaks to directly is a rounding error against the number of people it governs.

There is also a dignity dimension that is easy to overlook if you have always been able to read the documents that decide your life. A person who cannot read their own judgment is, in a real sense, locked out of a process that has authority over them. They cannot check whether their advocate described the order accurately. They cannot independently weigh whether to appeal. They experience the law as something that happens to them rather than something they participate in. Closing the language gap is not only about efficiency or convenience. It is about whether the legal system treats the citizen as a participant or a subject.

## Article 348: why the court speaks English

To understand why the wall exists, you have to read one constitutional provision. [Article 348](https://www.constitutionofindia.net/articles/article-348-language-to-be-used-in-the-supreme-court-and-in-the-high-courts-and-for-acts-bills-etc/) of the Constitution of India is the source of the rule.

Article 348(1)(a) says that until Parliament provides otherwise, all proceedings in the Supreme Court and in every High Court "shall be in the English language." Article 348(1)(b) carries the same rule across to legislation: the authoritative texts of all Bills, Acts, ordinances, orders, rules, and regulations are in English. The drafters wanted one language in which the higher judiciary and the statute book spoke, so that a precedent set in Chennai meant the same thing when cited in Chandigarh.

So English is not a habit or a convenience that the courts have drifted into. It is a constitutional default. That matters because it means you cannot simply decide, court by court, to start writing judgments in Hindi or Kannada. The Constitution provides one specific escape hatch, and it is narrow.

That escape hatch is Article 348(2). It allows the Governor of a State, with the prior consent of the President, to authorise the use of Hindi or any other official language of the State in proceedings in the High Court for that State. But the provision keeps a hard limit: even where this is authorised, the judgments, decrees, and orders themselves must still be passed in English. In other words, you can argue and conduct proceedings in Hindi in an authorised High Court, but the binding written output of the court stays in English unless a separate translation is also issued.

That single design choice - argue in your language, but the precedent is English - is exactly the gap that translation tools are now trying to close.

## Which courts may use Hindi and regional languages

The Constitution opened a door in 1950. Only a handful of States walked through it, and the window effectively closed not long after.

Four High Courts are authorised to use Hindi in their proceedings under Article 348(2). Per a [Press Information Bureau release on the use of regional languages in High Courts](https://www.pib.gov.in/PressReleasePage.aspx?PRID=1896029), these are Rajasthan (authorised in 1950), Uttar Pradesh (1969), Madhya Pradesh (1971), and Bihar (1972). After Bihar, no further authorisations were granted, and the reason is recent and instructive.

When other States asked to add their own languages, they were turned down. The Government of India informed Parliament, as [reported by The South First](https://thesouthfirst.com/tamilnadu/what-stops-tamil-being-made-the-official-language-of-the-madras-high-court/), that proposals to permit Tamil, Gujarati, Hindi, Bengali, and Kannada in the High Courts of Tamil Nadu, Gujarat, Chhattisgarh, West Bengal, and Karnataka were not accepted. The advice of the Chief Justice of India was sought, and the Full Court of the Supreme Court, after deliberation, declined the proposals. The concern, repeatedly stated, is that fragmenting the language of the higher judiciary could weaken the consistency of precedent and the mobility of judges across the country.

The result is a frozen map.

| High Court | Hindi authorised? | Year | Notes |
| --- | --- | --- | --- |
| Rajasthan | Yes | 1950 | First to be authorised |
| Uttar Pradesh (Allahabad) | Yes | 1969 | |
| Madhya Pradesh | Yes | 1971 | |
| Bihar (Patna) | Yes | 1972 | Last authorisation granted |
| Madras, Karnataka, Gujarat, Calcutta, Chhattisgarh | No | - | Requests declined by the Full Court |
| All others | No | - | English default under Article 348 |

So the formal answer to "can I get my judgment in my language as a matter of right?" is still, for almost everyone, no. The authorisation route has been stuck for over fifty years. This is precisely why the action has moved from changing the language of the court to translating the output of the court - and why AI became the obvious instrument. You cannot manually translate the country's accumulated case law. You can ask software to.

## The Supreme Court's translation drive: SUVAS and e-SCR

The most concrete vernacular-access programme in Indian law today is not a policy paper. It is a running translation pipeline, and the engine has a name: SUVAS.

The Supreme Court Vidhik Anuvaad Software (SUVAS) is an AI-powered, neural-machine-translation tool built to translate judicial documents between English and Indian languages. It was launched in November 2019 by the then Chief Justice of India, S.A. Bobde, as [documented by Drishti Judiciary](https://www.drishtijudiciary.com/current-affairs/supreme-court-vidhik-anuvaad-software-suvas). The point of SUVAS was always volume. Human legal translators are scarce and slow; the corpus of reportable judgments runs into the tens of thousands. Only machine translation, with humans checking the output, could move at the required scale.

The numbers it has produced are the part of this story that is easy to verify and genuinely large. Per figures the government placed on record and which were [reported by news agencies in December 2024](https://www.newkerala.com/news/2024/80331.htm), as of that month around 36,300 Supreme Court judgments had been translated into Hindi, and more than 42,700 judgments had been translated into other regional languages, all uploaded to the electronic Supreme Court Reports (e-SCR) portal. More recent figures, attributed to the government and circulated in 2025, put the count higher still, with roughly 36,344 judgments in Hindi and over 47,000 in other scheduled languages, and the programme described as covering 18 vernacular languages. These counts are as reported in official communications and press coverage rather than something an outside reader can independently re-derive, so treat the exact figure as a moving target that keeps rising.

The oversight structure is worth knowing because it tells you how seriously the judiciary takes the accuracy risk. The Chief Justice of India constituted an AI-Assisted Legal Translation Advisory Committee, headed by a sitting Supreme Court judge, to supervise the translation of reportable judgments into vernacular languages using AI tools. Mirror committees, each headed by a judge of the respective court, were set up in the High Courts, which carry much of the actual translation and verification load and publish their own e-High Court Reports. The machine drafts; judges and their committees own the sign-off.

| Element | What it is | Why it matters |
| --- | --- | --- |
| SUVAS | The Supreme Court's neural machine translation tool | Lets translation happen at the scale of the case-law corpus |
| e-SCR portal | Free public access to reportable judgments, including translations | The delivery channel for vernacular judgments |
| AI Translation Advisory Committee | Judge-led oversight body | Human accountability over machine output |
| High Court AI committees | Per-court verification bodies | Distributes the verification load and adds local-language expertise |

A note of honesty: e-SCR holds reportable judgments, which are a fraction of everything the courts decide. Translation of the operative part of orders into an applicant's mother tongue has also been piloted, and was [praised by Prime Minister Modi on Independence Day in 2023](https://www.businesstoday.in/latest/in-focus/story/pm-modi-praises-scs-initiative-of-making-judgments-available-in-regional-languages-cji-responds-with-folded-hands-394200-2023-08-15), with the Chief Justice acknowledging the remark from the audience. But a litigant in a routine matter before a district court will not usually find a clean, official, vernacular version of their own order waiting for them. The translation drive has reached the apex of the system first. Coverage thins out fast as you go down.

## SUPACE and AI inside the courtroom

Translation is one half of the judiciary's AI story. The other is research assistance, and that has a separate tool.

SUPACE - the Supreme Court Portal for Assistance in Courts Efficiency - was launched in April 2021, also under Chief Justice Bobde. As [described in coverage of its launch](https://www.lexology.com/library/detail.aspx?g=4000e4f2-6f32-4616-ab0a-138d04c2c5a6), it is an AI-assisted tool meant to help judges and their researchers handle the volume of case material: previewing files, extracting facts, building chronologies, and surfacing relevant provisions and case law. The design principle was stated up front and has held: SUPACE is assistive, not decisional. It collects and organises information; it does not decide cases.

For the purposes of this article, SUPACE matters for one reason. It shows that the Indian judiciary's own appetite for AI is real but deliberately bounded. The institution that translates 47,000 judgments with neural software is the same institution that keeps AI firmly on the research-support side of the line and away from the judging side. That instinct - use the machine to read, organise, and translate, never to decide - is exactly the right model for vernacular legal research too. AI can find the case, render it in Hindi, and summarise the holding. A human still confirms what the holding is.

## How Indic language models handle legal text

Outside the courts, the bigger shift is in the language models themselves. Until recently, a lawyer who wanted AI help in an Indian language was stuck routing everything through models tuned mostly on English and Western data. That produced fluent-sounding nonsense in Indian languages and a thin grasp of Indian legal concepts. India's sovereign-AI push is changing the supply side.

The clearest example is [Sarvam AI](https://en.wikipedia.org/wiki/Sarvam_AI), which in February 2026 announced two foundation models, Sarvam-30B and Sarvam-105B. Per [coverage of the launch](https://www.buildfastwithai.com/blogs/sarvam-105b-india-s-open-source-llm-for-22-indian-languages-2026), Sarvam-105B is built to support all 22 scheduled Indian languages and was trained not only on native scripts but also on romanised Latin script and code-mixed inputs for the most-spoken languages. A separate Sarvam-Translate offering targets all 22 scheduled languages and, importantly, structured documents - which is what legal text is. BharatGen and other initiatives are pushing in the same direction, with [BharatGen describing an Indian sovereign-AI stack moving from general models toward verticalised, domain-specific use](https://bharatgen.com/from-llms-to-verticalisation-india-sovereign-ai-stack-takes-shape/), of which law is an obvious vertical.

Why does Indic-first matter for legal work specifically? Three reasons.

First, tokenisation. Indian scripts are expensive for models built around English tokenisers; the same sentence costs many more tokens in Devanagari or Tamil than in Latin script, which degrades both quality and cost. Indic-first models retrain the tokeniser so that an Indian-language sentence is represented efficiently, which directly improves how well the model reads a Hindi pleading or a Marathi order.

Second, concept grounding. Indian law has its own vocabulary - vakalatnama, anticipatory bail, lok adalat, the structure of writ jurisdiction - that has no clean equivalent abroad. A model that has only ever seen these terms in passing will mishandle them. A model trained on Indian legal and linguistic data has a far better chance of treating "bail" and "anticipatory bail" as the distinct things they are.

Third, script and register. Legal language is a register of its own even within a language. Hindi legal Hindi is not conversational Hindi; it carries Persian-Arabic and Sanskrit-derived legal vocabulary that everyday models stumble on. Models trained with Indian legal text in the mix learn that register.

There is also a sovereignty argument that lines up neatly with the legal one. Legal data is among the most sensitive data a person owns - case files carry privileged communications, financial details, family disputes, and worse. An India-built model that can be run on India-hosted infrastructure lets that data stay in-country rather than being shipped to a foreign provider for processing. So the move to Indic-first models is not only about linguistic accuracy; it is about keeping a privileged corpus of Indian legal text under Indian control while making it more accessible to Indians. The two goals reinforce each other.

The upshot is that the raw ingredient for genuine vernacular legal research - a model that can actually read and write Indian-language legal text well - now exists in a way it did not two or three years ago. What was a hard ceiling on what vernacular legal tools could do has, fairly suddenly, become a solvable engineering problem.

## Code-mixing, transliteration, and the Hinglish problem

Here is the part that makes Indian legal AI genuinely hard, and that generic Western models fail at most visibly: real Indians do not write in one language at a time.

A junior advocate's WhatsApp note about a matter might read: "Sir, the anticipatory bail wali application ready hai, kal Patna HC mein file karenge." That is Hindi grammar, Roman script, English legal nouns, and a Hindi verb, all in one breath. This is code-mixing, and it is the default register of working Indian professional life, not an exception. Add transliteration - the same Hindi word spelled five different ways in Roman script by five different people - and you have an input that breaks any model that assumes clean, single-language text.

This is why the design choices in models like Sarvam matter so much. Training explicitly on code-mixed inputs such as Hinglish and on romanised script, rather than only on pristine native-script text, is not a nice-to-have. It is the difference between a model that can follow how Indian lawyers actually communicate and one that chokes the moment a Hindi sentence carries an English legal term, which is to say, almost always. As [one practitioner analysis of Sarvam's handling of code-switching notes](https://medium.com/@bavalpreetsinghh/sarvam-ai-is-it-worth-for-code-switching-problems-57e5d3348b1a), code-switching remains a hard problem even for purpose-built Indic models, but they start from a far better baseline than English-centric ones.

For legal research, the practical payoff is specific. A user should be able to type a query the way they think it - half English, half Hindi, in Roman script - and have the system understand that they are asking about, say, the limitation period for a particular kind of suit, then return the relevant English judgments with a Hindi explanation. Forcing the user to first translate their own question into formal English defeats the entire purpose of vernacular access.

## Where machine translation of legal text breaks

Now the uncomfortable part, and the reason this article keeps insisting on human verification: machine translation of legal text fails in ways that are quiet, plausible-looking, and capable of inverting a legal conclusion.

The failure mode is not gibberish. Gibberish is safe, because you notice it. The dangerous output is a clean, confident, grammatical Hindi sentence that means the opposite of the English original. Research on translating legal text into Indian languages, including the [MILPaC benchmark from researchers studying exactly this problem](https://arxiv.org/abs/2310.09765), has found that general-purpose translation systems struggle badly in the legal domain, where vocabulary, syntax, and semantics diverge sharply from ordinary language and where small slips carry large consequences.

The mechanics of how legal translation goes wrong are worth spelling out, because they are not obvious.

| Failure type | What happens | Why it is dangerous |
| --- | --- | --- |
| Prepositional inversion | "held by the defendant" becomes "held for the defendant" | Flips who owns or controls something in a property ruling |
| Negation drift | A "not liable" finding loses or softens the negation | Reverses the outcome of the case |
| Term collapse | Two distinct legal terms get the same translation | Erases a distinction the whole judgment turns on |
| Register flattening | Formal legal vocabulary rendered in casual words | Loses the precise meaning a term of art carries |
| Citation mangling | Case names, statute sections, dates corrupted | Sends the reader to the wrong authority |

A reported analysis of court-order translation, [When Words Betray Justice in India Legal](https://indialegallive.com/magazine/supreme-court-order-translation-linguistic-diversity/), captured the property-ruling inversion example directly and argued that such errors are systemic, not occasional, and that without human review the reliability of raw machine output in this domain is low. The exact accuracy figures vary by study and language pair and should be read as illustrative rather than fixed, but the direction is consistent across the literature: legal machine translation without a human in the loop is not safe to rely on for anything consequential.

There is a second reason legal translation is uniquely unforgiving. In ordinary translation, a near-miss is usually fine - if a tourist guide renders "delicious" as "tasty," nobody is harmed. In law, near-misses are often catastrophic, because legal meaning lives in tiny distinctions. The difference between "shall" and "may" is the difference between an obligation and a discretion. The difference between "and" and "or" in a statute can decide who qualifies for a benefit. Indian legal drafting is dense with these load-bearing small words, and a translation system optimised for general fluency will happily smooth them over in the name of reading naturally. Fluency and fidelity pull in opposite directions here, and for legal text, fidelity has to win.

This is also why the question "which language pair?" matters. Translation quality is not uniform across India's languages. Models tend to perform better on high-resource pairs such as English-Hindi, where training data is abundant, and noticeably worse on lower-resource scheduled languages where far less parallel legal text exists. A user reading a Hindi translation and a user reading the same judgment in a smaller language are not getting the same reliability, even from the same tool. Honest products surface that difference rather than presenting every translation as equally trustworthy.

This is the single most important thing to internalise about vernacular legal AI. The technology is good enough to make the law dramatically more accessible. It is not good enough to be trusted blind. Those two statements are both true, and the entire design of any responsible tool follows from holding them together.

## What vernacular legal research actually unlocks

Set the risks aside for a moment and look at the upside, because it is large and it reaches people the legal system has historically failed.

**The litigant who can finally read their own case.** This is the headline benefit. A worker who lost a regularisation claim, a homebuyer fighting a builder, a woman seeking maintenance - if the judgment and the precedents behind it can be rendered into Hindi or Tamil or Bengali, the person stops being a passive spectator to their own matter. They can ask their lawyer a better question. They can sense when something is off. That alone changes the power balance between client and counsel.

**Paralegals and clerks.** A huge amount of Indian legal work runs through people who are competent, experienced, and not fluent in formal legal English. Vernacular search and explanation lets a paralegal in a district town do real first-pass research - find the relevant judgments, understand the holding, flag the right cases for the advocate - instead of being limited to clerical tasks.

**Law students and young lawyers from regional-medium backgrounds.** A bright student who studied in a Hindi-medium or Tamil-medium school has historically faced a brutal English tax on entry to the profession. Tools that let them read precedent with vernacular support lower that tax without lowering the standard, because the underlying authority is still the English judgment.

**Rural and small-town legal aid.** Legal-aid clinics and paralegal volunteers working with the most vulnerable populations gain a force multiplier. The same query that once required an English-fluent supervisor can now be started by a local worker in their own language.

**Plain comprehension of complex orders.** Even English-comfortable readers benefit. A long, dense judgment summarised in clear language - whether English or a regional language - is simply faster to absorb. If you want a sense of how much structure a judgment carries and why reading one well is a skill, our guide on [how to read a judgment](/blog/how-to-read-a-judgment) breaks down the anatomy.

The connecting thread is that vernacular AI does not replace the lawyer. It widens the circle of people who can meaningfully participate in legal research - and participation is what access to justice ultimately means.

## The limits: why a human still has to verify

Everything good about vernacular legal AI comes with a non-negotiable condition attached, and pretending otherwise would be dishonest.

**The translation is a draft, not the authority.** Under Article 348, the English judgment remains the legally operative text. A Hindi translation, however good, is an aid to understanding. If there is ever a discrepancy between the translation and the English original, the English original governs. Any serious step - filing, advising, relying in court - must be checked against the English source.

**Confident wrongness is the core risk.** As the translation failure modes above show, the dangerous output looks right. A negation that quietly flipped, a term of art rendered casually, a "by" that became a "for" - none of these announce themselves. Only a human who reads the English source and the translation together catches them.

**Hallucinated authority is a separate, worse danger.** Beyond translation, when a general-purpose chatbot is asked to do legal research, it can invent case names, citations, and holdings that do not exist - and it does this just as fluently in Hindi as in English. Indian courts have already encountered fabricated citations produced by AI. If you want the detail on how and why this happens, see our piece on [AI-hallucinated citations in India](/blog/ai-hallucinated-citations-india). Vernacular output does not reduce this risk; if anything it raises the stakes, because a non-English-reading user is less able to sanity-check the underlying English authority.

**Grounding beats raw generation.** The safe architecture is not "ask a model to write about Indian law from memory." It is to retrieve real, verifiable judgments from an authoritative corpus and have the AI summarise and translate only what is actually in those documents, with citations the user can open and check. This is the same reason [neutral citations and the e-SCR system](/blog/e-scr-neutral-citations) matter so much: they give every judgment a stable, verifiable identity that an AI answer can point back to.

**The professional duty does not transfer to the machine.** An advocate who relies on a vernacular AI summary without verifying the source remains fully responsible for the result. The tool is an assistant. The lawyer is accountable.

Hold these limits next to the benefits and you arrive at the only honest position: vernacular legal AI is a powerful accelerant for research and comprehension, used by a human who verifies, and a liability the moment it is trusted blind.

## How Niyam approaches Indian-language legal research

Niyam is built on the assumptions this article has argued for, rather than against them.

The starting point is grounding. Niyam does not ask a language model to recall Indian law from training memory. It searches a corpus of real Indian judgments and answers from what is actually in those documents, with citations the user can open and read. That single design choice removes most of the hallucinated-citation risk, because the AI is summarising retrieved text rather than inventing it. For why this approach is the right default for Indian conditions specifically, see our argument on [sovereign AI and Indian legal tech](/blog/sovereign-ai-india-legal-tech).

On top of that grounding, Indian-language capability is treated as a first-class feature, not a translation afterthought. A user should be able to ask a question the way they naturally would - in Hindi, in code-mixed Hinglish, in Roman script - and have the system understand it, find the relevant English judgments, and explain the holding in language they can read. The authoritative English source stays one click away, exactly because the English judgment remains the operative text under Article 348.

The "find similar cases" workflow is where this compounds. A paralegal or junior advocate who can describe a fact pattern in their own language and get back genuinely relevant precedent - rather than keyword soup - is doing real legal research, regardless of which language they think in. Our explanation of [how AI finds similar judgments](/blog/ai-find-similar-judgments) covers the mechanics of that retrieval.

And the verification discipline is built into the experience rather than bolted on. Every answer points back to the source judgment. The product's job is to get a verifying human to the right document faster and in a language they understand. It is not to be the final word. That is the difference between an AI that widens access to justice and one that quietly manufactures new ways to get the law wrong.

## Frequently asked questions

**Can I get a Supreme Court judgment officially in Hindi or my regional language?**
For many reportable Supreme Court judgments, yes - the Supreme Court has translated tens of thousands of judgments into Hindi and other scheduled languages using its SUVAS tool, and these are available free on the e-SCR portal. But coverage is concentrated on reportable apex-court judgments. A routine order from your local district court usually will not have an official, verified vernacular version, and even where a translation exists, the English text remains the legally operative one.

**Which High Courts are allowed to use Hindi?**
Four High Courts are formally authorised under Article 348(2) to use Hindi in their proceedings: Rajasthan, Uttar Pradesh (Allahabad), Madhya Pradesh, and Bihar (Patna). Requests by States such as Tamil Nadu, Karnataka, Gujarat, West Bengal, and Chhattisgarh to use their own languages were declined by the Full Court of the Supreme Court, so the list has not grown since 1972.

**Why is the Supreme Court in English at all?**
Because Article 348 of the Constitution makes English the default language of the Supreme Court and the High Courts, and of the authoritative texts of legislation. The drafters wanted a single language for the higher judiciary and the statute book so that precedent meant the same thing across the whole country. Changing that requires the narrow authorisation route in Article 348(2), which only four States ever used.

**Is AI translation of judgments accurate enough to rely on?**
For understanding, often. For relying on without checking, no. Machine translation of legal text is prone to quiet, dangerous errors - flipped negations, inverted prepositions, collapsed terms of art - that look perfectly fluent. That is exactly why the Supreme Court runs its translations through judge-led AI translation committees, and why any consequential use must be verified against the English source.

**What makes Indian-language legal AI harder than ordinary translation?**
Three things stacked together: Indian legal vocabulary that has no clean foreign equivalent, the heavy code-mixing of how Indians actually write (Hindi grammar, English legal nouns, Roman script, all at once), and the specialised legal register within each language. Models built mainly on English data fail at all three. India-built models like those from Sarvam AI are trained on the 22 scheduled languages, code-mixed input, and romanised script specifically to handle this.

**Does vernacular AI replace the need for a lawyer?**
No. It widens who can participate in legal research - litigants, paralegals, regional-medium students - but it does not transfer professional responsibility to the machine. The English judgment remains the authority, AI output must be verified against the source, and an advocate who relies on an unverified summary is still fully accountable for the outcome.

**Can I research Indian case law by typing my question in Hindi or Hinglish?**
With a tool designed for Indian conditions, yes. The goal is to let you ask in the language you think in, including code-mixed Hinglish in Roman script, and get back relevant, real, citable judgments with an explanation you can read - while keeping the authoritative English source one click away for verification.

---

## Start researching in the language you think in

The English wall around Indian law is real, and it is constitutional. But the tools to read around it now exist - if they are built the right way: grounded in real judgments, honest about translation limits, and designed to get a verifying human to the right source fast. That is what Niyam is for.

[Start for ₹100](https://app.niyam.ai/register) and search Indian case law the way you actually think about it. Real judgments. Real citations you can open. Indian-language support that treats verification as the point, not an afterthought.