The future of the Irish language in the AI era

A language does not enter the future because a model can translate a sentence into it. It enters the future when people can work, learn, create, search, speak and build software through it without accepting a worse experience.

That is the standard Irish language AI should meet. The opportunity is not to add Gaeilge as a decorative option in an English product. It is to make Irish a capable working language in the systems that increasingly mediate education, public services, media and business.

The numbers contain both strength and warning

Census 2022 recorded 1,873,997 people aged three and over who said they could speak Irish. That was 40% of the relevant population. The same CSO release recorded only 71,968 people speaking Irish daily outside education, a 2% fall from 2016.

The gap matters. Ireland has a large population with some knowledge of the language, a smaller group using it every day, and living Gaeltacht communities where Irish is not a curriculum subject but the language of family and place. Technology must not flatten those groups into one synthetic idea of an Irish speaker.

The State's 20-Year Strategy set a goal of 250,000 daily speakers outside education by 2030. The 2022 census shows how far away that goal remains. Yet there is real institutional progress too. Since 1 January 2022, Irish has stood on equal footing with the other 23 official EU languages, after the translation derogation ended.

Status creates demand. Daily use creates a future.

AI changes the cost of building in Irish

Language software used to require a separate specialist system for each major task. Modern multilingual models can transfer knowledge across languages, then be adapted with smaller, carefully designed datasets. Research such as No Language Left Behind demonstrated the reach of multilingual training across 200 languages, while also showing that lower-resource work still depends on deliberate data gathering and human evaluation.

This lowers the cost of starting. It does not remove the hard parts. Irish still needs:

• text and speech data with clear rights and documented provenance;
• coverage of dialect, register, code-switching and contemporary usage;
• evaluation written and judged by fluent speakers;
• tests for grammar, meaning, factuality, safety and cultural context;
• products designed around real needs rather than benchmark scores.

Open resources already show what community contribution can achieve. Mozilla Common Voice's June 2026 Irish scripted-speech release documents 15,381 clips from 281 speakers, with 14.26 validated hours. That is valuable infrastructure. It is also a reminder that language coverage is never finished. Read speech alone cannot represent ordinary conversation, regional variation or every domain where people need Irish to work.

The model must answer to the language community

A capable Irish model cannot be built by scraping first and apologising later. Data rights, speaker consent and community governance are part of model quality. A technically strong dataset can still be socially weak if speakers cannot see how their words are used or challenge a harmful release.

The right process starts with listening. Teachers may need writing feedback that explains errors without replacing student work. broadcasters may need transcription that respects names and dialect. public bodies may need retrieval over approved Irish documents. Gaeltacht organisations may have entirely different priorities from a consumer app.

Those use cases should define the evaluations before training begins. If the goal is public-service question answering, success is not a generic multilingual score. It is whether the system retrieves the right Irish source, preserves its meaning, answers in appropriate Irish and admits when the evidence is missing.

What LeemerLabs plans to build

Our Gaeilge AI programme is a plan, not a claim of completion. We want to build it with fluent speakers, researchers, educators, public bodies and language organisations. The work has five practical tracks.

1. Map the need. Select a small number of tasks where a better Irish model would create repeated, useful language activity.
2. Curate lawful data. Record source, licence, consent, dialect, domain and transformation history for every dataset.
3. Build Irish-first evaluations. Use fluent human review alongside automated checks, and publish the rubric before the headline score.
4. Adapt open multilingual bases. Compare prompting, retrieval, fine-tuning and continued pretraining rather than assuming every problem needs a new foundation model.
5. Release what can be shared. Publish model cards, baselines, evaluation code and suitable datasets, while keeping sensitive or restricted material private.

The same capability can support private fine-tuning for organisations with licensed domain material. Private delivery and public benefit are not opposites. A commercial engagement can fund reusable evaluation tools, documentation and open research, provided rights and boundaries are explicit from the start.

A useful future is bigger than one model

Irish will not be secured by a single checkpoint. Models change too quickly. The durable assets are people, datasets with trustworthy provenance, evaluation suites, language tools, teaching practice and institutions able to carry the work forward.

Our contribution should be judged plainly: does the work make it easier to choose Irish in daily digital life, and does it leave more capability in the hands of the Irish-language community than it found?

If you work with Irish language data, education, public services, speech technology or evaluation, write to hello@leemerlabs.ie. We are looking for partners who want to define the work before we train the model.