Singapore wants regional language and health models. Give communities a data contract

THE SIGNAL IN ONE SENTENCE

A language dataset can make a region more visible to machines. It can also turn a community into a very large folder that nobody thought to ask. Nanyang Technological University Singapore and Google signed a memorandum of understanding on October 30 that expands their work across university teaching, Southeast Asian language technology, scientific research and AI safety. The language work is the sharpest part of the signal. NTU says it will work with AI Singapore and Google on regional initiatives that collect, catalogue and open-source datasets reflecting Southeast Asian linguistic nuances. The announcement names Projects ATLAS and Amplify and says AI Singapore will continue building independent evaluations and benchmarks tailored to Southeast Asian cultural contexts. The partners also say they want to develop sovereign AI models for education and healthcare that use Southeast Asian languages and Google's open-weight Gemma and MedGemma families. That sentence contains several different projects. Language collection. Dataset curation. Open release. Model development. Evaluation. Education deployment. Healthcare deployment. Each one needs its own evidence and permission boundary. The memorandum is not a finished model or a clinical result. It does not publish a final dataset inventory, license, consent process, benchmark score, budget or delivery date. It establishes a direction and a set of relationships. That makes this the right moment to write the data contract. Start with a plain question: whose language is being collected? Southeast Asia is not one linguistic market. People move among national languages, community languages, dialects, scripts, professional vocabularies and code-switched conversation. The same phrase can change by country, generation, setting and relationship. A health worker, teenager, teacher and market trader may all use the same named language differently. A dataset that counts languages but ignores those conditions can look broad while remaining thin. The contract should name the communities represented, not merely the language label attached to a file. For every source, record country or territory, locality when appropriate, language variety, script, domain, time period, speaker or author context, collection method and the people or institutions that reviewed the description. Do not turn that into a demand for invasive personal profiles. The goal is to document the social and technical context that affects meaning while minimizing personal data. A public dataset card can describe the collection population and known gaps without publishing the identity of an individual speaker, patient or student. Next comes authority. Some material may be created for the project. Some may come from archives, public websites, licensed collections, educational resources, community organizations or earlier research. Publicly reachable is not the same as freely reusable. Open-source release is a distribution decision. It is not evidence that collection was fair, that copyright was cleared or that the people represented expected their language to become training material for education or healthcare models. The data contract should identify the legal basis, consent process or license for every source category. It should record the intended uses disclosed at collection, the uses now proposed, restrictions on sensitive domains and the review required before a new use is added. When consent is the basis, people need a form they can understand in the language and setting where collection happens. It should say what is recorded, what metadata travels with it, whether it may be released openly, whether commercial systems may use it and whether withdrawal is possible after distribution. Withdrawal becomes complicated once a dataset is copied. That is not an excuse to offer no path. The contract can require maintainers to remove the material from future official releases, publish a revocation record, notify known downstream users and document which trained models cannot practically forget it. That is less satisfying than a magic delete button. It is also more honest. Sensitive material needs a separate lane. Healthcare is not simply another vocabulary domain. A clinical note, patient question or conversation about symptoms can reveal identity, disability, pregnancy, mental health, migration status, family relationships and other information that can harm a person when mishandled. Even material with names removed can be re-identified through rare conditions, places, dates or combinations of detail. MedGemma being open-weight does not make patient data open. Any healthcare dataset should begin with a purpose, a minimum necessary data rule, an ethics and privacy review, a de-identification test, access controls and a statement of whether the material may ever enter a public release. Some useful work should happen inside a secure environment and leave only aggregate findings or carefully reviewed model updates. The NTU agreement includes a secure research sandbox on Google Cloud for pre-approved environments, vetted datasets and controlled collaboration. That can help separate sensitive research from casual tool use. But a sandbox is not a substitute for the contract. It can control who opens a file. It cannot decide whether the file should exist, whether the label is culturally sound or whether the people represented share in the benefit. Education carries a different set of risks. Student writing, classroom discussion and assessment records can reveal age, performance, beliefs and learning needs. A dataset built from coursework can quietly turn participation in education into participation in model development. The contract should state whether students can decline without academic penalty, how teacher authority affects consent, how minors are protected and whether generated assistants are allowed to infer ability, behavior or risk from ordinary classwork. The model also needs a job description. The phrase sovereign AI model sounds sturdy. It can mean different things depending on who is speaking. It may refer to local language capability, local control over deployment, local hosting, domestic governance, regional data, open weights or an institution's ability to change providers. Those are not interchangeable. A model built from Gemma can support local adaptation while still depending on outside architecture, cloud services, update schedules or tooling. A model hosted in Singapore can still be weak in Khmer, Javanese or a mixed-language clinical conversation. A regionally branded model can still have unclear data rights. The project should publish which part is sovereign. Who controls the training data? Who can inspect the weights? Who can run the system without a particular vendor? Who sets the safety rules? Who can correct an error? Which countries and communities have a formal voice? What happens when a partner leaves? That last question matters because memoranda create cooperation, not permanence. The model should survive a change in vendor, grant, university leader or cloud contract without taking the public evidence with it. Evaluation needs the same regional precision. AI Singapore has already developed open regional models, datasets and benchmarks through efforts including SEA-LION and Project SEALD. Its public materials describe work across Indonesian, Malay, Tamil, Burmese, Filipino, Vietnamese, Thai, Lao and Khmer, with collaborators from universities, industry and government. That history is useful. It also shows why one regional score is not enough. Language performance should be reported by variety, script, task, domain and user group where the data permits responsible analysis. A model can translate a news sentence cleanly and fail on a clinic instruction. It can answer in formal Indonesian and misunderstand everyday mixed speech. It can summarize a school text while inventing a citation. It can produce fluent Malay and miss the point of a local metaphor. The benchmark should record those failures in categories people can act on. For education, test evidence tracing, hallucinated sources, age appropriateness, instruction following, accessibility and whether the system exposes uncertainty. For healthcare, test clinical terminology, dosage and unit handling, negation, temporal sequence, referral urgency, dialect variation and the ability to stop when the input is outside scope. The healthcare model should never earn a general trust badge from one average score. Every test should name the task, population, conditions, reference standard, acceptable error, reviewer expertise and intended human oversight. Benchmarks also need protected material. If the complete test set becomes training data, the score becomes a memory quiz wearing a lab coat. Maintain versioned public examples for transparency and protected evaluation sets for credible comparison. Document who can access the protected material and how leakage is investigated. Community review belongs beside technical review. Linguists can identify grammar and variation. Domain professionals can evaluate education and healthcare meaning. Community reviewers can spot stereotypes, disrespect, missing context and uses that violate local expectations. None of those groups can replace the others. The contract should define how reviewers are selected, paid, credited and protected. It should record disagreement instead of forcing every cultural question into one answer. Southeast Asia does not need a machine-readable fiction that everybody agreed. It needs a record of where the views differ and who decides for a specific use. Benefit sharing is the part most data projects leave until the slide after the budget. If a community contributes language material, the return should not be limited to a model that may or may not work for them later. Benefits can include paid collection and review, grants to local institutions, accessible datasets and tools, training, compute access, authorship, governance seats, community archives, support for local educators and health workers, or revenue terms when a commercial product depends materially on the contribution. The appropriate form will vary. The obligation to discuss it should not. Open release also needs maintenance. A dataset is not finished when the download link appears. It needs a version, change log, license, contact, correction process, security-reporting path, known limitations and a retirement policy. Every model trained from it should record the dataset version and transformations. Every benchmark result should identify the test version. Every correction should say which releases and models are affected. That creates a lineage people can follow. The contract should be readable in two directions. From a model output, a reviewer should be able to trace the model version, training and adaptation sources, benchmark evidence and human approval. From a community contribution, a reviewer should be able to trace the dataset version, permitted uses, projects that consumed it, published evaluations, known harms, corrections and benefits returned. This is not a demand to publish private records. It is a demand to keep accountability connected while protecting the people inside the data. The agreement also includes campus-wide AI tools, educator training, a scientific research sandbox and work on privacy, security, post-quantum cryptography, deepfake detection and model protection. Those programs may produce useful infrastructure around the language effort. Still, the regional language and health ambition will be judged by more than technical access. Can a Thai researcher challenge a category designed in Singapore? Can a Filipino community see how its speech data was used? Can a Burmese-language reviewer report a harmful output without writing the complaint in English? Can a patient know that a model assisted a decision? Can a student refuse to donate classwork? Can an institution move the model when the vendor relationship changes? Can the public see which languages and communities remain missing? Those are data-contract questions. The memorandum sets a promising regional agenda. The partners want models that fit Southeast Asian languages and local uses instead of asking the region to squeeze itself into systems trained elsewhere. Good. Now give the people represented in those systems more than a place in the corpus. Give them a place in the contract.