AI and legal sources in Sinhala, Tamil and Hindi: how reliable?
By the ChamberOne team 5 minute read
Ask a general AI tool to translate a paragraph of a statute from English into Sinhala, Tamil or Hindi, and the result will usually read smoothly. That is the danger. Smooth is not the same as correct, and legal language is full of words whose everyday meaning differs from their legal one. Lawyers who work across languages, in Sri Lanka, India and Singapore among other places, have good reasons to want these tools to work. They save time on first readings and help with client communication. They also need a testing habit, because the quality is uneven and changes without notice.
Key points
- Fluent output is not the same as accurate legal translation, especially for terms with a legal and an everyday sense.
- Treat the official text in the official language as the reference and check which version prevails.
- Build a small test set from your own documents and have a fluent lawyer grade the results.
- Check names, numbers and dates by hand after any machine translation.
- Use human translators for anything that will be filed, signed or relied upon.
Why legal text is harder than ordinary text
General translation tools learn from large quantities of everyday writing. Legal writing is a small and odd corner of that material. It uses long sentences, defined terms, archaic phrases and cross-references. Words such as 'consideration', 'order', 'discharge' or 'party' carry specific meanings that a general tool may flatten.
Languages with less published legal text pose a further problem. Where there are fewer examples of how a term has been translated in judgments, statutes and contracts, a tool has less to learn from, and it may invent a plausible rendering. Quality also differs by language, by subject and by the kind of document. A tool that handles a simple notice in one language can stumble on a contract in another.
Official texts remain the reference
In multilingual jurisdictions, statutes and judgments are issued in particular languages, and the rules on which version prevails differ between places and between documents. Check the guidance from your regulator or bar, and the text of the instrument itself, rather than assuming that any translation is equal to the original.
In practice, treat the official text as the reference and every machine translation as a reading aid. If a passage will be quoted, filed or advised upon, go back to the official version in the official language and have a person fluent in that language read it.
- The reference, in its official language
- Go back to it before quoting, filing or advising
- Have a person fluent in that language read it
- A reading aid only
- Useful for first reads and client conversations
- Never assumed equal to the original
Building a small test set
Do not trust a vendor demonstration or a colleague's impression. Build your own test and keep it. It need not be large. Twenty carefully chosen passages will tell you more than a general claim of accuracy.
- Pick short passages from documents your firm actually handles: a lease clause, a notice, a witness statement, a section of a familiar statute.
- Include terms that are known to be tricky in your language pair, such as words that have both an everyday and a legal sense.
- Include names, addresses, dates and amounts, which are often mangled.
- Ask a fluent lawyer to translate or approve a reference version of each passage.
- Run each passage through the tool, then have the fluent lawyer mark errors as harmless, misleading or serious.
- Repeat the test when the tool or its model changes, and record the date and result.
- 1Pick passages from documents your firm handles
- 2Get a reference version from a fluent lawyer
- 3Run each passage through the tool
- 4Mark errors as harmless, misleading or serious
- 5Repeat when the tool or its model changes
- Then back to step 1.
Search in local languages
Translation is only half the story. Many lawyers also want to search in the language of the document, or to ask a question in one language and find sources in another. That depends on how well the search system handles spelling variants, word endings, transliteration and mixed-script text.
Test it the same way. Take five passages you know are in your document set, search for them using ordinary phrasing, and see whether they appear. Then search for something that is not there, and see whether the tool admits that or produces something plausible. A search tool that never says 'nothing found' deserves suspicion.
A workflow for a client who reads another language
Suppose a small Colombo firm has a client who reads Tamil and needs to understand an English contract. The lawyer uses a tool to produce a first Tamil rendering of a few key clauses for discussion. A bilingual colleague reads it and marks any phrase that distorts the meaning. The lawyer then walks the client through the clauses, uses the Tamil version only as a conversation aid, and notes on file that the English text is the operative document and the Tamil was informal.
The same pattern works for Hindi and Sinhala. The tool speeds up the first pass. People supply the accuracy and carry the responsibility.
Names, numbers and transliteration
The errors that do the most damage are often small. A personal name transliterated two ways can make one person look like two, which matters for conflict checks and for identity in a pleading. A figure written in local numerals can be misread. A date in one calendar convention can be converted wrongly.
Ask a person to read every name, number and date against the source after any machine translation. Where a name has more than one accepted spelling, record the variants on the file so that later searches find them all.
Data protection and tool choice
Sending a client document to a translation service means sending it to a third party. Before you do, check what the provider does with the text, where it is processed and whether it is kept. For confidential documents, use a tool your firm has approved, and consider whether the sensitivity of the document justifies using a human translator instead.
Keep a short list of approved tools for translation and tell staff that free consumer tools are not on it unless someone has checked them.
Questions to ask a vendor
- Which languages and scripts are supported for legal text, and how was quality tested?
- Does the tool say when it is unsure, or does it always produce an answer?
- Can we see the source passage beside each translated passage?
- Does the model change without notice, and will we be told?
- What happens to our documents after translation?
When to use a human translator
Machine translation is a sensible tool for a first read, for internal understanding and for conversation aids. It is not the right route for documents that will be filed, signed or relied upon as a statement of the parties' agreement. For those, use a qualified translator and follow any requirements the forum has for translated documents. The cost of a human translator is small beside the cost of a dispute over what a document meant.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
How to check AI legal research before you rely on it
A routine for tracing every AI-suggested authority to its source, confirming it is still good law and testing the proposition it is cited for.
5 minute readGeneral chatbots versus legal research tools: what differs and why it matters
How general chatbots and legal research tools differ on sourcing, links, jurisdiction coverage and currency, and when a general model will do.
5 minute readRetrieval-augmented generation explained for lawyers
A plain explanation of how AI answers from a document set, why it reduces but does not remove fabricated content, and what to ask vendors.
5 minute read