Singapore's Polyglot: Using AI to translate government web pages into all four official languages

The problem

Singapore has four official languages, English, Chinese, Malay and Tamil, but most government websites are published only in English. Residents who are more confident in another language, many of them elderly, struggle to access official information and may have to rely on family members to interpret it.

Translating government content by hand is time-consuming and expensive: it needs dedicated professionals for each language, so most material does not get translated. Machine translation is faster, but standard tools work word-by-word and miss local meaning. The district known in English as Chinatown has a distinct local name, and the Malay term for "government" in Singapore differs from the one used in neighbouring countries, the kind of error a generic translator makes.

The solution

A five-person team at Open Government Products built Polyglot, a tool that lets a reader view a government web page in whichever of the four official languages they prefer. It went live on 26 June 2026 as a pilot on the Ministry of Digital Development and Information (MDDI) website, accessible via a drop-down menu on each page, with a banner directing visitors to the translated beta version. The underlying translation engine was developed jointly by MDDI and GovTech.

What distinguishes Polyglot from a standard translator is how it handles meaning. Rather than substituting one word at a time, it uses large language models, a form of AI trained on large volumes of text that can interpret whole sentences and passages at once, so the output reflects the sense of a passage rather than its literal words. The team pairs this with a reference of how official terms are rendered in Singapore, so the wording follows local usage instead of a generic equivalent. The live tool describes its purpose in those terms: recognising local terms such as CPF, HDB and MediShield rather than translating them literally, so a resident does not need someone else to explain an official letter or form.

Two features keep people in control of the output. Officials at the ministry can revise the translations by hand, a step the team built in deliberately to keep people in the loop and one it considers necessary with this kind of technology. Every translated page also carries a notice that AI produced the text and that it may not be fully accurate. Through the pilot, the team is inviting the public to report errors and suggest improvements.

The impact

1. Government information can now be read on demand in four languages on the pilot site. On the MDDI website, a visitor can switch the page to English, Chinese, Malay, or Tamil from a drop-down menu, whereas previously the material was effectively English-only. The intent is to bring government information within reach of more residents, including those who are more comfortable speaking their mother tongue.

2. The pilot is framed as a test, with any wider rollout left to depend on what it shows. The tool is confined to one ministry's website until October 2026. The team has described its longer aim as making reliable translation a standard part of every government website and digital service, while setting no date and saying that reaching that scale will have to be earned.

Lessons learned

• The problem with automated translation is local meaning, not the words. Polyglot's central design choice is to translate for context rather than word-for-word, using language models to read whole passages and a database of local term usage to keep the output faithful to how things are said in Singapore. For another government with multiple official languages, the transferable insight is that the hard part is not converting words but preserving meaning, and that a generic translation tool can introduce errors precisely in the local terms citizens most need to understand.
• Keeping a person able to correct the output was built in, not added later. The tool lets officers edit translations by hand, and every translated page indicates that AI was used and that the translation may not be fully accurate. The team frames human involvement as necessary when working with AI. For governments deploying AI on public-facing information, this is a practical pattern: let the system do the volume, keep a person able to override it, and be transparent with the public that a translation is machine-generated.
• Scaling is treated as something to earn. Rather than committing to a government-wide rollout at launch, the team has tied any expansion to what the pilot and public feedback show and has deliberately set no timeline. For other governments, this measured approach, prove it on one site, invite error reports, expand only on evidence, is a way to deploy a public-facing AI tool without over-promising before the results are in.

This case study was written with assistance from artificial intelligence.