A European conference in Galway has put language diversity on the AI agenda. The practical challenge is to build tools that help people use their languages in everyday life, with communities able to shape the data, judge the results and sustain the work
Imagine a person who speaks Irish at home opening an online service and finding that its conversational assistant works best in English. The assistant may accept Irish text and produce an answer that looks fluent. The person still has to decide whether it understood an unusual expression, preserved an important distinction or supplied the right information. Switching to English might feel like the quickest way to get the task done. Repeated across ordinary encounters, that choice would change where a language is used, even if nobody had formally decided to exclude it. This illustrative situation captures a question now facing language communities as AI becomes an interface to work, learning and public information.
A statement published by Údarás na Gaeltachta on 10 November 2026 reports that the International Conference on Language Diversity and Technology brought more than 250 delegates to Galway on 5–6 November. Participants called for Irish and other European languages to have a place in AI development. The statement describes continued work on digital resources and cooperation. It records an agenda and commitments, with no measured change in language use attached. The event’s dates are also confirmed by the ADAPT research centre’s conference programme, which describes the meeting as a forum for under-resourced languages and AI.
The social significance extends beyond whether a particular model can translate a sentence. When software helps people write messages, search records or learn a subject, its linguistic strengths shape which choices feel easy. A language that requires repeated corrections can demand more time from its speakers than a better supported alternative. That friction can matter in a busy workplace or a short appointment. The challenge raised in Galway is how to make everyday participation possible in a person’s chosen language, with tools whose quality and limits the community can assess.
Digital scarcity and a living language
The phrase “low-resource language” describes a relationship with available technological resources. A language can have skilled speakers, a rich literature and considerable use in daily life while offering developers relatively little appropriately licensed digital material. Another may have abundant text of one kind but little recorded speech with accurate transcripts. Those differences affect what can be trained and evaluated. A collection of parliamentary documents might help a system handle formal phrasing while offering less guidance for a conversation with a child or a regional expression. Counting the volume of material therefore leaves open whether it represents the work people want a tool to do.
This creates a practical choice about where to begin. In an illustrative community project, an oral-history archive could support transcription, but its vocabulary and recording conditions might differ from those of a contemporary service desk. A team could develop a tool for helping archivists locate passages and evaluate it for that purpose. Extending it to live conversations would require new evidence. The distinction gives a community a way to pursue a manageable benefit while learning what material and expertise it needs for the next application. It also keeps a successful demonstration from silently becoming evidence for every other use of the language.
A Simon Fraser University announcement from 28 October describes a proposed sequence with community leadership at its start. The FLAIR project plans to develop curricula with Indigenous language experts, then build community-curated audio and text datasets for speech and learning tools. It also describes an ambition to support smaller tools on local devices, including offline use. These are development plans rather than published outcome measures. They illustrate how a project can begin with the people and activities it aims to support before deciding which model to build.
Who decides what the data may become?
Collecting language material can create value and new obligations at the same time. A recording might contain a speaker’s voice, family references or knowledge shared for a particular audience. Agreement to preserve that recording in a local archive may differ from agreement to make it available for unrestricted model training. The distinction is especially important when material passes through several organisations. A contributor needs a comprehensible account of where the recording will go, who can reuse it and whether an apparently local project depends on an external service. A dataset becomes easier to govern when those choices are documented while it is being assembled.
Community participation also involves deciding whose language is represented. A dataset built from one variety can make other speakers appear unusual to the resulting tool. A spelling convention can make search more consistent while making some existing records harder to find. These are choices that teachers, speakers, archivists and linguists may approach from different practical needs. A useful design process would make those differences visible and permit more than one valid form where the task allows it. Treating every variation as an error could simplify the software while diminishing the range of expression it is meant to support.
Consider a hypothetical speech transcription project with several regional varieties. Its organisers might keep metadata about recording conditions and the variety used, with appropriate permissions, so that they can inspect where recognition fails. Speakers could help select examples that test meaningful differences rather than merely familiar phrases. That work takes time and warrants resources. If it is treated as unpaid correction after a model has already been deployed, the people experiencing the worst results inherit the largest repair burden. A budget for language technology therefore includes the expertise needed to decide what counts as a faithful result.
Fluency is only one part of quality
A fluent answer can make errors harder for a learner or a second-language speaker to recognise. The difficulty grows when the evaluator itself has limited competence in the language. In a 2025 study of multilingual translation benchmarks, Chihiro Taguchi and colleagues examined data in Asante Twi, Japanese, Jinghpaw and South Azerbaijani and identified shortcomings in the suitability of a widely used benchmark for multilingual evaluation. The study concerns those data and does not establish the quality of Irish-language tools. It does establish a reason to inspect what a benchmark asks a system to do before interpreting its score as a measure of everyday usefulness.
A hypothetical public-information assistant makes the issue concrete. A response could be grammatical while weakening an instruction about when a document must be submitted. A translation of a story could preserve the broad plot while losing a relationship expressed through a particular term. A learner’s exercise could be understandable but consistently prefer a form their teacher would correct. Each task calls for different assessment. A team could ask speakers to identify consequential meaning changes, compare acceptable variants and examine the system’s handling of uncertainty. An aggregate measure would then need to be accompanied by the kinds of mistakes that matter for the intended use.
A 2026 study by Rudali Huidrom and colleagues investigates the choice of automatic metrics and model judges for low-resource Indic translation, including settings where some target scripts lack suitable pretrained encoders. Its setting differs from Ireland’s. The broader methodological question is relevant wherever a service relies on automated grading to claim adequate language support. Developers need to show why the chosen evaluation can recognise the distinctions speakers care about, and how its conclusions compare with informed human review. Publishing those details gives a community something more usable than a broad claim that its language is supported.
Tools that make room for more use
The potential benefit is considerable when the application fits the need. An illustrative archive tool could help a researcher find recordings containing a topic, reducing the amount of audio they must inspect manually. A teacher might use a carefully reviewed drafting aid to prepare exercises at different levels. A small publisher could use a search tool to locate terminology across its own cleared material. These possibilities require validation in their particular settings. Their significance would lie in making more work feasible in the language, with speakers able to review and correct the outputs at a cost they can sustain.
The same tool can create different incentives depending on how an institution introduces it. A translation assistant could give staff more capacity to prepare information in Irish, or it could be used to remove the specialist review that makes the information dependable. A speech system could reduce repetitive transcription work while creating a new need to inspect names, dialect forms and noisy passages. Measuring the change means looking at the complete activity, including the corrections and editorial decisions. A large volume of generated material alone would reveal little about whether people can read, trust or use it.
For learners, the experience of correction deserves particular attention. A tutor that gives immediate feedback could make practice more accessible, especially when a human teacher is unavailable. But the learner may have little ability to challenge the feedback. A responsible pilot could give teachers access to common corrections, test the system on forms known to cause confusion and provide a clear route for reporting a disputed answer. It would also ask whether the tool encourages sustained engagement. The number of completed exercises can be observed quickly. Whether learners become more comfortable using the language with other people takes a different kind of follow-up.
From a conference agenda to durable provision
The Galway discussions bring attention to a collective-action problem. The benefits of dependable language infrastructure can reach schools, cultural organisations, businesses and public bodies, while no single institution has an obvious incentive to fund the whole of it. A shared recording standard or a well-maintained evaluation set might help many projects whose immediate goals differ. Cooperation can make that work possible, provided the organisations involved agree on permissions, maintenance and how contributors can influence later uses. The value of a shared resource depends in part on whether another team can understand its limits and use it appropriately years after its first release.
Public procurement could help translate such ambitions into observable requirements. An institution commissioning a language service could specify tasks and varieties that must be evaluated, the expertise involved in review and the way errors will be handled after deployment. It could require documentation of changes so that a software update does not quietly reduce quality for a group of speakers. These are proposed ways to make performance reviewable, rather than arrangements confirmed by the conference statement. They would let speakers and service owners assess delivery against a defined activity instead of relying on the presence of a language in a product menu.
The strongest evidence of progress would be found in ordinary use. Can a speaker complete a task in Irish without repeatedly switching to English? Can an archivist find and verify a passage more easily? Can a teacher correct the system and see that correction handled reliably? Those questions turn a broad commitment to linguistic diversity into things communities can investigate. They also connect technology to the settings in which a language remains alive, including conversations, learning and work. AI can support those activities when its design gives speakers useful capabilities and a continuing say in how their language is represented.
Image credit: Gaeltacht sign in Mionlach (Menlo) by Sheila1988, CC0 1.0. Illustrative photograph unrelated to the conference.
