Amis Annual Meeting Brings AI Into Language Revitalization: A Governable Indigenous-Language AI Workflow for Taiwan's 55 Indigenous Townships

Indigenous-language AI becomes most valuable when community members can use it, correct it, report errors, and understand where a piece of language data came from, who agreed to its use, and when it needs to be reviewed again.

On August 29, TITV News reported that the Taiwan Amis Language Sustainable Development Association was holding its two-day 13th annual meeting in Taitung. For the first time, the annual meeting placed AI technology at the center of its language-revitalization discussion. A National Dong Hwa University educator demonstrated the Indigenous Languages Research and Development Foundation's speech recognition, text-to-speech and basic translation services. Participants also discussed how fieldwork, elder interviews and existing audio-visual materials might become AI learning resources, and how user feedback can support continued correction. The report also recorded concerns about informed consent, data sovereignty, traditional knowledge and cultural restrictions. This moves the discussion beyond whether AI can produce an Indigenous language and toward the more practical question of how Indigenous-language AI should be governed. The event can be checked in the TITV News report on the Amis annual meeting.

The meeting connects directly with tools that are already public. The Foundation's Indigenous-language AI results site currently provides speech recognition, text-to-speech and basic translation. Its text-to-speech service explicitly notes that the system is a test result and that spelling, segmentation or content errors may still occur, so formal or consequential information needs careful human review. That warning is a useful local deployment principle: AI can assist recognition, synthesis, translation and retrieval, while language correctness and cultural appropriateness remain subject to review by speakers, teachers and knowledge holders. For local governments, this is more practical than promising a fully automatic model, because it gives users a visible path to correction and accountability.

A first local workflow can begin with a “corpus passport.” Every recording, transcript, photograph, video or teaching item should carry provenance, Indigenous nation and language or local variety, speaker or provider, collection time and place, consent method, permitted uses, whether model training is allowed, whether public release is allowed, whether commercial reuse is allowed, cultural-sensitivity level, latest human-review date, version and withdrawal procedure. Not every item needs to be public. Open teaching resources, education-only materials, research-only materials, community-internal resources and ceremonial or restricted knowledge can be separated into different access layers.

The Foundation also provides an official Indigenous-language learning word-list system with learning resources across 42 Indigenous language varieties. Resources with identifiable provenance and versions are suitable starting points for a public corpus layer. At the same time, the AI project copyright statement explains that some corpora, materials and outputs are subject to formal authorization requirements. For township offices, schools, tribal colleges and research teams, the practical rule is therefore not “if it is online, it can be put into a model.” The safer rule is to record the lawful source, intended use and community or speaker permission first, and only then decide whether a resource may enter RAG, evaluation or model-training workflows.

If this is connected to a chatbot or a Two-Eyed Seeing knowledge laboratory, a permission-first pattern can be used: authorize first, retrieve second, generate last. Questions about common vocabulary, open teaching materials or official services can retrieve only from the public layer. Questions involving a local variety, elder narratives or traditional knowledge should first be checked against permissions; the service can then decide whether to show an answer, provide only a source reference, transfer the question to a language teacher, or keep the material outside generation entirely. An answer page should display the data date, source, language or variety, human-review status and an error-report action. These service fields often improve trust more directly than adding model parameters.

The annual meeting's emphasis on high-quality corpora and user feedback also suggests a manageable MVP. A township or community does not need to collect everything at once. It can begin with one language variety or one service scene, such as common township-service phrases, older adults' everyday speech, classroom sentence patterns, or agriculture and climate vocabulary, and build a gold set of 100 to 500 human-verified records. Each record can preserve the audio, transcript, speaker information, version and correction history. When ASR misrecognizes a phrase, text-to-speech pronunciation is unnatural, or translation is inaccurate, community users should be able to submit an error that is routed to a designated teacher or language worker. The resulting correction trail becomes auditable improvement data instead of anonymous text whose editing history is unknown.

International work increasingly treats language technology and language rights together. UNESCO strengthened collaboration with Unicode in 2026 to support Indigenous languages in digital environments; its earlier work on Indigenous data and generative AI also emphasized data sovereignty, informed consent and cultural sensitivity. Local planners can use UNESCO's collaboration with Unicode and UNESCO's guidance on ethical generative-AI use of Indigenous data as international reference points when deciding how language resources are represented, governed and reused.

Where traditional knowledge or cultural material is involved, the Global Indigenous Data Alliance CARE Principles and Local Contexts add another layer of useful questions: not only whether data can be found, but who can decide how it is used, how collective benefit is protected, and whether conditions can later be changed or withdrawn. These questions are especially relevant to ceremonial knowledge, ancestral recipes, gathering routes, place knowledge, songs, prayers and elder narratives. A service can make ordinary public-language materials easier to find while keeping culturally restricted knowledge under community-defined control.

For Taiwan's 55 Indigenous townships, the most useful first-stage outcome may therefore be a maintainable Indigenous-language AI workflow rather than a newly trained large model. Public data can carry traceable sources; sensitive material can carry permissions; every corpus item can have a version; AI output can have human review; errors can have a reporting channel; and teaching materials, chatbots, AI image journals and RAG can share the same governance fields. Once these foundations are stable, a township can change models or platforms and add new speech tools without reorganizing all of its language data from the beginning.

Six fields the 55 Indigenous townships can use immediately

1. Where the corpus came from: fieldwork, elder interviews, teaching materials, audio-visual records, official dictionaries or other public resources.

2. Who agreed to which uses: public release, teaching, research, model training, commercial use or community-only access.

3. How the language is identified: Indigenous nation, language, local variety, speaker, time, place and human reviewer.

4. What AI is allowed to do: recognition, synthesis, translation, retrieval and drafting; formal teaching materials and cultural judgments remain human-reviewed.

5. How errors are corrected: preserve reports, versions, the reviewer responsible and the correction date rather than overwriting history.

6. How sensitive data can exit: access changes, withdrawal and cessation of model use need executable procedures.

A workflow built this way makes Indigenous-language AI more like a public tool that community members can maintain together and less like a black box that can only be consumed.

Sources

• TITV News | Amis language annual meeting focuses on AI for endangered-language revitalization
• Indigenous Languages Research and Development Foundation | Indigenous-language AI results site
• Indigenous Languages Research and Development Foundation | AI project copyright statement
• Indigenous Languages Research and Development Foundation | Indigenous-language learning word-list system
• UNESCO | UNESCO and Unicode strengthen collaboration for Indigenous languages online
• UNESCO | Ethical Generative AI Use of Indigenous Data
• Global Indigenous Data Alliance | CARE Principles
• Local Contexts