OpenAI wants to help teach your child's mother tongue

The Sunday evening scene is familiar in a lot of Singapore homes. The English homework goes fast. The mother tongue book comes out last, stays out longest, and somebody at the table starts negotiating.

So it is worth knowing that this particular lesson, the one that causes the most grief at home, is now the worked example in a S$300 million partnership between Singapore and OpenAI.

What Singapore actually signed

In May, OpenAI announced a partnership backed by more than S$300 million to support Singapore’s national AI strategy, with a memorandum of understanding signed with the Ministry of Digital Development and Information. The company is putting its first Applied AI Lab outside the United States here, with more than 200 Singapore-based technical roles planned over the next few years.

The education part of that deal is more specific than the headline figure. The partnership names the Ministry of Education and GovTech among its partners, and the stated plan is work on AI-enabled learning use cases and tools, “including more interactive support for Mother Tongue language learning”. Not maths. Not essay marking in English. Mother tongue is the one example given.

The pattern kept going through the year. On 11 August, OpenAI and the National University of Singapore announced campus-wide ChatGPT Edu and Codex access for students, faculty and staff, with OpenAI’s managing director for international, Oliver Jay, saying Singapore “has an extraordinary opportunity to shape how AI contributes to education, research and innovation”.

Universities are one thing. Primary school is another, and that is where a frontier lab’s model quality stops being a tech story and becomes something your child sits an exam in.

Three mother tongues, three very different piles of data

Here is the part that does not make the press release.

SEA-HELM is a benchmark built by AI Singapore, the national programme funded by the National Research Foundation and hosted at NUS, to test how well language models actually handle this region. Its 2025 paper covers Filipino, Indonesian, Tamil, Thai and Vietnamese, and reports that “the performance of the LLMs for Tamil and Filipino were relatively poor”, which the authors put down to “the lack of support for these languages in many of these models”, itself a consequence of “a lack of both training and testing data on the internet”.

AI Singapore’s own write-up on benchmarking is blunter about how invisible these languages are in the standard evaluations. Even Global MMLU, a serious multilingual benchmark, “does not cover Burmese, Tamil, or Thai, all of which are official languages in various SEA countries”. And there is a detail in there that any Tamil-speaking parent will recognise instantly: “Tamil datasets, for example, heavily feature literary Tamil, leading to lower support for spoken varieties.”

Read that again with the oral exam in mind. A model trained mostly on formal written Tamil is being asked to help a nine-year-old rehearse the way people actually talk.

This is not an argument against the deal. It is an argument for asking a different question about it. Notice that the benchmark evidence is itself uneven: Tamil is measured and comes out weak, while Malay and Chinese are not in that paper’s list at all, so nobody outside the labs really knows how a model performs in all three of Singapore’s mother tongues. Our view, offered as a view: whoever ships a national mother tongue tool should publish results language by language, not an average.

MOE is not waiting, and it is building narrow

Worth noting that the ministry already has its own AI in schools, and its instincts are conservative in a way that is quietly reassuring.

At the Mother Tongue Languages Symposium on 1 August, Education Minister Desmond Lee announced AIMS, an AI essay marking system launching in secondary schools in September for Chinese language. It handles routine marking of grammar, spelling and sentence structure, on handwritten and typed essays, so that teachers can spend their attention on creativity and the quality of ideas. Lee’s framing was that it “will free up teachers from repetitive marking tasks, so they can spend more time guiding students”. He also said the plain thing out loud: “Raising a bilingual child in an English-dominated environment is not easy at all.”

Chinese first. Malay and Tamil, presumably, later.

MOE’s stated position on AI in schools, set out in a parliamentary reply on 6 May 2026, runs on what it calls the Four Learns: learn about AI, learn to use AI, learn with AI, and “most importantly, learn beyond AI”. Primary 1 to 3 pupils learn about AI but do not use it for assigned work. From Primary 4, they use approved tools under teacher supervision. Student data is anonymised and not used to train external AI models. National examinations remain AI-free. A longitudinal study, SG-LEADS, starts in 2027 to look at what all this is doing to children’s development.

That is a ministry that has thought about the difference between a tool and a teacher. We wrote about the same instinct in Singapore’s school AI that listens and in AI marks the spelling.

What to do with this at home

You are not going to audit a benchmark. But you can do the thirty-second version of one, and it is genuinely worth doing before you trust any AI tool with your child’s mother tongue practice.

Ask it something in your language that you know the answer to and a textbook would not. A local idiom. The word your family uses for a dish. A sentence in the spoken register rather than the written one. Then ask the same thing in English. If the English answer is crisp and the mother tongue answer is vague, formal or subtly wrong, you have just learned exactly what the benchmark says, in less time than it takes to boil water.

And if a school tool arrives in the bag this term, the useful question at the parent teacher meeting is not whether the school uses AI. It is which language it was tested in.

The homework negotiation on Sunday evening will continue either way. But a child who looks up and says “the computer got that wrong, nobody at home says it like that” has picked up something more durable than the week’s vocabulary list.

Disclosure: Mentus AI builds AI mentors for children, including language practice, so we have a direct commercial interest in this subject. It is also why we read the benchmarks before the press releases.