Indic & Citizen Services
Legal Text in Indian Languages: Hard Problems
· 10 minute read
Legal language in India is bilingual, defined and proviso-heavy. A model that smooths a notification into readable Hindi can change a right. Treat that as a legal incident, not as a BLEU gain.
A compliance officer asked whether an agent could 'keep our Hindi and English notifications in sync'. The demo paraphrased a proviso into a shorter Hindi sentence that a journalist later quoted as the rule. The English authentic text still contained the proviso. Two authentic-looking pages, one missing condition.
That is the hard problem. Not rare words. Structure. Defined terms. Provisos. Commencement. The fact that India often has more than one official linguistic face for the same instrument, and that citizens will obey the face they can read.
This teardown is for DPOs, legal cells and the officers who will be asked why the bot said may when the Gazette said shall. It is not a product claim. It is not legal advice.
The legal object is not a paragraph
An instrument has operative sentences, definitions, provisos, explanations, schedules, commencement, and sometimes a savings clause. A model that is good at 'summarise this policy' will flatten those into a pamphlet. Pamphlets are useful. They are not the instrument.
Section 3 of the Official Languages Act requires both Hindi and English for listed Union instruments. That creates an alignment duty the Union already carries with humans. An agent does not inherit the power to discharge that duty unsupervised.
Failure modes that change rights
Shall becomes may, or the reverse. A proviso detaches and attaches to the next clause. A defined term is replaced by a synonym that is not defined. A date of commencement is dropped. A schedule is summarised and a row disappears. A non obstante clause is treated as colour. A Hindi official term is back-formed from English into a word the earlier authentic corpus never used, creating a second term of art.
| Failure | Why it is legal | Agent behaviour |
|---|---|---|
| Shall / may swap | Duty versus discretion | Hard fail; do not publish |
| Detached proviso | Condition silently removed | Surface both sentences; human pass |
| Defined-term synonym | A new undefined object | Glossary lock; fail if banned form |
| Dropped commencement | People obey on the wrong day | Date is an operative field |
| Picked one authentic version | The other version still binds | Show both or show the conflict |
What models cannot see unless you make them
They cannot see that two documents are the same instrument in two languages unless you pair them. They cannot see that a 2019 amendment silently rewrote a 2014 definition unless you retrieve both. They cannot see that a popular website paraphrase is not authentic unless you tag the source class.
Retrieval design is therefore a legal control. The corpus must distinguish authentic, unofficial, expired and draft. An agent that cites a blog explainer next to a Gazette is not 'being helpful'. It is mixing legal classes.
A defensible role for an agent
Retrieve the authentic text and the paired language version. Diff a draft against the last authentic. Highlight operative sentences for the legal cell. Answer a citizen with a labelled unofficial explanation plus a link to the instrument. Refuse to apply a rule to a named person without the officer path.
That role still processes personal data when the query names someone. The DPO's map and the legal cell's map have to be the same drawing.
Objections you will hear — and what to do with them
These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.
If we are this strict, we cannot offer Indic legal help at all.
You can offer retrieval, pairing, diffs and labelled explainers. You cannot offer unsupervised authentic text. That is already how a careful legal cell works without a model.
Judges and officers already use English as the real law.
Citizens and many officers do not. The Official Languages Act exists because the Union must speak two official faces. Design for the face people will obey.
Our model was trained on Indian Kanoon and is therefore safe.
A scrape is not a corpus card. It mixes authentic, commentary and error. It does not give you commencement hygiene. Sit your instruments.
Legal review will become the bottleneck.
Legal review is the product for Class A text. If that is the bottleneck, you have found the real staffing problem.
Pamphlets need expiry
An unofficial explainer that ranked last year will outlive the amendment if you do not date and expire it. Fluency makes the old page look current. That is the opposite of a Gazette.
Put an expiry on every unofficial Indic page the agent is allowed to emit. When the instrument changes, the page dies or it is rebuilt under a human pass. No third option.
Three weeks to a legal-language protocol
Sit the DPO and the legal cell in the same room. If they draw different pictures, stop the agent.
- Week 1: inventory authentic sources versus unofficial pages.
- Week 1: pair bilingual instruments you already issued.
- Week 2: lock shall/may and defined terms in the glossary.
- Week 2: tag the corpus by legal class.
- Week 2: write the refuse list — no unsupervised authentic, no speaking order from a pamphlet.
- Week 3: trial diffs on three old bilingual pairs.
- Week 3: write the expiry rule for unofficial explainers.
- Week 3: file the joint DPO–legal map.
How this shows up in the file
The note should say: the agent does not issue authentic legal text. It retrieves and diffs. Unofficial Indic explainers are labelled, dated and linked. Shall/may and defined terms are glossary-locked. Conflicts between authentic language versions are surfaced, not resolved by fluency. This is not a legal opinion.
Attach the class tags and the refuse list.
This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation, linguistics or engineering advice. Confirm against the current Gazette, Official Languages Act and Rules, state official-language law, GIGW, RPwD Act, DPDP text and Rules, CERT-In directions, departmental manual and your counsel before you file it.
How to test this with real speech, not staff English
“Legal Text in Indian Languages: Hard Problems” fails in the field if you only tested officers. A P6 Compliance/DPO should hear a first-generation student, a rural caller, or a Hinglish grievance before claiming “legal text Indic NLP”.
Legal language in India is bilingual, defined and proviso-heavy. A model that smooths a notification into readable Hindi can change a right. Treat that as a legal incident, not as a BLEU gain. Twenty-two scheduled languages is a Constitution fact, not a model fact. Script support is not language support. Official language rules may require bilingual output even when the model prefers one script.
- Name the languages and scripts in the eval set.
- Include code-mix and scheme-name tests.
- Measure comprehension, not BLEU alone.
- Design a human fallback when language fails.
Close this loop before the next CAB
Put “Legal Text in Indian Languages: Hard Problems” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P6 Compliance/DPO, not “the vendor.”
Revisit the item when the model, the GeM term, the region, or the SI changes. “legal text Indic NLP” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.
What must be true before you file this
If “Legal Text in Indian Languages: Hard Problems” is only a heading, it will not survive a file inspection. A P6 Compliance/DPO should be able to attach one artefact that proves “legal text Indic NLP”: a log export, a clause, a scored row, a dated notice, or a refusal rule.
Write three dated sentences: what was decided, who owns it, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.
- Name the owner of “legal text Indic NLP” inside the institution.
- Attach one artefact a stranger can open next year.
- Revisit when the model, the notice, or the SI changes.
- Do not treat a vendor slide as evidence.
What the next file must contain
“Legal Text in Indian Languages: Hard Problems” earns a line in the noting only if a P6 Compliance/DPO can attach proof of “legal text Indic NLP.” A heading is not proof. A vendor slide is not proof. A workshop photograph is not proof.
Write three dated sentences: what was decided, who owns it after the next posting order, and when it will be re-checked. If you cannot write the three sentences, you are not ready to buy, to sell, or to go live.
Leave unsourced percentages out of the note. DPDP is not a blanket localisation statute. The November 2025 AI governance text is guidance, not an Act. CERT-In’s 28 April 2022 directions still set specified incident and log clocks. A PAC, when lawful, lives in GFR Rule 166.
- Name the designation that owns “legal text Indic NLP.”
- Attach one artefact a stranger can open next year.
- Record the instrument you are actually using.
- Revisit when the model, the SI, the notice or the posting changes.
Questions this usually raises
- If Hindi and English versions both go out, which does the agent trust?
- The authentic text as your legal cell specifies. Where both are authentic, a conflict is a legal fact. The agent must surface the conflict, not pick the prettier paragraph.
- Can we use a model to keep the two authentic versions aligned?
- As a drafter and a diff assistant, yes. As the signer, no. Alignment is a human legal pass.
- Is this only a Union Official Languages Act issue?
- No. States have official-language laws. High courts and tribunals have language practice. Universities have statutes. Read the instrument in front of you. The Act is the Union floor, not the whole country.
- Does DPDP apply to legal text?
- A notification can be public. A draft that names a person, or a file noting that applies a rule to a citizen, is personal data. The legal-language problem and the personal-data problem often sit in the same document.
- Are we giving legal advice in this article?
- No. This is field guidance on where models fail. Your counsel owns the authentic text.
Sources
- Official Languages Act, 1963 — Department of Official Language
- Section 3, Official Languages Act, 1963 (India Kanoon)
- Constitution of India — Eighth Schedule (languages)
- Department of Administrative Reforms — Central Secretariat Manual of Office Procedure
- Digital Personal Data Protection Act, 2023
- Prcept AI — on-prem / air-gapped agents
- MeitY — India AI Governance Guidelines (5 November 2025, PIB PDF)