Wales’s Welsh-Language AI Strategy Exposes the Gap Between Data and Daily Use

The Welsh Government closed its six-year technology plan in 2024 with an unusual claim: Welsh was better prepared for the AI era. A Google News item, syndicated through Wired-Gov, has now returned that claim to the technology news cycle. The record includes speech recognition, synthetic voices, machine translation, and work with OpenAI and Microsoft.

Yet the real test is no longer whether Wales can fund useful language components. It is whether those components reach the AI products, workplaces, schools, and public services that shape daily language choices. Open datasets matter, but people cannot use a corpus during a meeting or ask a dataset to complete an application.

That distinction creates pressure for both the Welsh Government and major technology companies. Wales must turn publicly funded infrastructure into measurable adoption. OpenAI, Microsoft, and other platform operators must show that Welsh support remains dependable as their models and interfaces change.

What the Google News Headline Actually Summarizes

The headline describes a policy transition, not a new AI product or a single 2025 launch.

The underlying story begins with the Welsh Language Technology Action Plan, which ran from 2018 through 2024. Its goal was to make Welsh usable across more digital situations, including voice interfaces, translation systems, and conversational software.

The Welsh Government published its final technology report on February 20, 2024. That document closed the original program and reviewed work completed since its launch.

The plan concentrated on three connected areas. These were speech technology, computer-assisted translation, and conversational artificial intelligence. Together, they form much of the technical foundation needed by a language inside modern software.

Speech technology includes automatic speech recognition, which converts recorded or live speech into text. It also includes text-to-speech systems, which generate spoken audio from written material.

Computer-assisted translation gives professional translators terminology, translation memories, and automated suggestions. Conversational AI includes chatbots and large language models, or LLMs, which generate language by predicting likely sequences of words.

The government says the program helped develop Welsh speech-to-text technology, specialist machine translation, and synthetic voices. Some synthetic voices were designed for people who lose their ability to speak.

The program also supported digital resources that developers can reuse. That point matters because a smaller language rarely attracts the same commercial data investment as English, Spanish, or Mandarin.

The Google News presentation can make the story look like a broad summary of Welsh technology and AI. The underlying policy is narrower and more specific. It concerns technology for the Welsh language, rather than every AI company, data center, or digital policy operating in Wales.

That difference is important for readers assessing the claim. Wales did not announce a domestic foundation model that competes with systems from OpenAI, Google, or Anthropic. It invested in the language assets and partnerships that can make Welsh work inside such systems.

By February 2024, that approach had produced several visible integrations. Welsh officials highlighted work with OpenAI to improve GPT-4’s handling of Welsh. They also cited cooperation with Microsoft on interpretation within Teams.

Microsoft Teams introduced a function that lets a human interpreter deliver another language during a scheduled meeting. It is not automatic machine translation. The software provides separate audio channels so participants can choose the interpreted language.

That distinction illustrates the policy’s practical character. Wales did not wait for an AI model to solve every bilingual meeting problem. It worked with an established platform to support a human-led service that organizations could use immediately.

Jeremy Miles, then Minister for Education and Welsh Language, said the objective was to let people use Welsh in more technological situations without making a special request. That is a higher standard than simply publishing a Welsh interface.

A language feature exists when a vendor lists it. A usable language experience works across input, output, search, transcription, translation, and support. It must also survive product updates that can quietly weaken less-used options.

The 2024 report therefore marked the end of one funding framework, not the end of the work. The next phase would need to connect technical supply with daily demand.

Wales Built AI Ingredients Before Generative AI Became the Main Event

Wales’s strongest decision was to invest in reusable language infrastructure before chatbots made that work commercially urgent.

Training and evaluating language systems require text, speech, dictionaries, annotations, and standardized terminology. Developers need permission to process and redistribute those materials. A useful collection trapped behind incompatible licenses cannot support an open research project or a commercial product.

Welsh presents a familiar problem for smaller-language AI. It has an active community, public institutions, broadcasters, schools, and substantial written material. However, its machine-readable resources remain small beside the English datasets collected from the open web.

This imbalance affects more than fluent conversation. Models can struggle with Welsh spelling, regional language patterns, specialist terms, names, code-switching, and spoken audio. They can also answer in English after receiving a Welsh prompt.

The government’s response was to treat data as shared infrastructure. Publicly funded resources were often released under licenses that permit reuse. This gives universities, independent developers, and larger vendors a common starting point.

The strategy anticipated an important feature of generative AI. A language’s position depends partly on whether developers can find clean, lawful, representative data when training or evaluating a model.

The Welsh Government’s 2025 technology updates show that collection continued after the original plan ended. By August 2025, the government reported 249 hours of Welsh audio training data across four named collections.

Those collections included 124 hours from Common Voice, 45 hours from the Bangor Transcription Bank, 40 hours from a Cymen project, and 40 hours from Paldaruo. Audio hours are not interchangeable, since recording quality and speaker diversity affect their value.

The update also described more granular data work. Contributors tagged the grammatical roles of words in over 200 additional sentences. They annotated 400 sentence entities for knowledge-linking applications.

Named-entity annotation identifies people, places, organizations, and other meaningful items inside text. Knowledge linking connects those items to structured records, helping systems distinguish a place name from an ordinary word.

Welsh officials also reported sharing more than 10,000 Welsh names through Wikidata. A related project created 11,264 Welsh labels for building names.

These numbers are modest compared with the corpora behind leading English models. Their value comes from specificity, licensing, and human verification. Carefully labeled material can test failures that a huge, loosely collected dataset conceals.

Speech data carries a similar tradeoff. More hours usually help, but coverage also matters. A system trained on a narrow group of speakers can perform poorly across age groups, accents, recording environments, and regional varieties.

This is why open infrastructure remains a stronger long-term asset than a temporary chatbot demonstration. A single model release becomes obsolete. A well-maintained dataset can improve many generations of products.

Wales also shared material with AINA, a Catalan government project focused on language technology. That cooperation places Welsh within a broader effort by European language communities to reduce dependence on English-first development.

Catalan and Basque offer useful reference points. Each has invested in public language resources, research partnerships, and deployment. The comparison shows that smaller languages can influence technology markets when institutions coordinate data and purchasing.

However, the comparison also exposes the scale problem. Even good public datasets remain tiny beside the material available for globally dominant languages. Wales cannot close that gap through collection alone.

Its leverage comes from making Welsh resources easy to reuse and difficult for major vendors to ignore. Public procurement, evaluation standards, and partnerships can turn a small dataset into a product requirement.

A Google News headline can surface the completed projects, but it cannot show whether those projects changed user behavior. That requires evidence from products and workplaces.

The Main Contest Is Infrastructure Versus Everyday Use

Wales has proved that it can create technical building blocks, but it has not yet proved that those blocks consistently increase daily Welsh use.

The Welsh Government set three priorities after reviewing the original action plan. The first is using technology to increase daily Welsh use, especially at work.

The second is ensuring that language technology reaches both developers and users. The third is improving Welsh AI and speech technology through data sharing and related work.

Mark Drakeford, then Cabinet Secretary for Finance and the Welsh Language, described those priorities in a January 7, 2025 technology statement. The timing matters because it moved the policy from building resources toward distribution and adoption.

A developer can find an open speech dataset and still lack the engineering budget to integrate it. A public agency can buy software that nominally supports Welsh but defaults to English at critical steps.

Workers can also avoid Welsh tools if they add delay or require special configuration. A technically correct feature may fail when it disrupts an established workflow.

This is the article’s central opponent: language infrastructure versus language use. The government’s strongest evidence sits on the infrastructure side. The public outcome depends on the other side.

Consider workplace meetings. The Teams interpretation feature removes one barrier by supporting separate language audio. Yet an organization still needs interpreters, meeting procedures, informed participants, and administrators who configure the feature.

Consider speech recognition. A model can perform well in a controlled evaluation but struggle with several speakers, background noise, specialist vocabulary, or a conversation that shifts between Welsh and English.

Consider generative AI. A chatbot can produce fluent Welsh while introducing factual errors or unnatural terminology. Fluency alone does not establish reliability for education, health care, legal work, or government services.

The government’s partnership with OpenAI is therefore best understood as access and influence, not a permanent guarantee. Model behavior changes as vendors update training data, safety systems, retrieval tools, and user interfaces.

OpenAI’s work on GPT-4 gave Welsh policymakers a channel into one influential system. It did not create a public benchmark that compares every later model across accuracy, dialect coverage, safety, and consistency.

Microsoft cooperation presents the same tension. Teams interpretation is a valuable deployment, while improvements to Copilot’s Welsh processing address a different technical layer. Neither makes every Microsoft workflow bilingual by default.

The challenge grows when AI features appear inside word processors, email, search, customer service, and operating systems. Each interface creates another place where Welsh support can be complete, partial, or absent.

Users rarely evaluate these layers separately. They experience one broken sequence. If dictation recognizes Welsh but the assistant answers in English, the whole task feels unsupported.

That is why distribution should become as visible as dataset production. Wales needs evidence about active use, task completion, error rates, and user satisfaction.

Public bodies can help by publishing bilingual technology requirements in procurement contracts. They can require vendors to disclose which features support Welsh and how that support is tested.

They can also share evaluation materials. A reusable benchmark would let schools, agencies, researchers, and suppliers compare systems against the same language tasks.

Knowledge workers face a related problem when useful records are scattered across meetings, documents, and notes. A personal knowledge base can preserve context, but language support still determines whether retrieval works equally across Welsh and English material.

This makes Welsh technology policy relevant beyond Wales. Smaller-language support is a test of whether AI platforms can serve multilingual users without forcing them into one dominant language.

If vendors solve retrieval, speech, and generation for Welsh, they develop methods that can transfer to other languages with limited digital resources. If they do not, AI adoption can reinforce existing language inequalities.

What the Progress Numbers Do Not Establish

The available evidence documents activity and assets, but it does not yet provide a complete public measure of reliability or adoption.

The 249 hours of audio reported in August 2025 offer a concrete baseline. They do not reveal how accurately commercial systems transcribe Welsh across real public-service settings.

The counts for names, labels, and annotations show useful expansion. They do not show how many developers downloaded the data, which products incorporated it, or whether users noticed better results.

This is not a reason to dismiss the work. Infrastructure programs often produce value indirectly. The problem is that indirect value becomes difficult to defend when the next funding or procurement decision arrives.

A public benchmark could close part of the gap. It could test transcription accuracy, translation quality, factual response quality, code-switching, and regional variation.

Testing should include high-stakes and ordinary tasks. A health appointment has different risks from drafting an informal message, but both shape confidence in a language tool.

The evaluation would also need human reviewers. Automated metrics can compare model output with a reference, yet several Welsh translations may be correct. Native speakers must judge meaning, tone, and naturalness.

Privacy creates another constraint. Speech data gathered from public services, schools, or workplaces can contain sensitive information. Open licensing is valuable only when collection respects consent and data protection.

Representation matters as well. A dataset can be legally reusable while underrepresenting particular accents, ages, communities, or language abilities. Systems trained on it can reproduce those gaps.

Generative systems introduce separate concerns. A model may hallucinate, meaning it produces plausible but unsupported information. It may also translate a specialist term incorrectly while preserving a confident tone.

Those risks make human oversight essential in public services. Welsh support should not become a reason to lower standards that apply to English-language tools.

The government’s 2025 AI overview says Welsh public-sector adoption should be ethical, trustworthy, inclusive, and accessible. Those principles provide direction, but operational controls determine whether they influence deployments.

Wales later established a broader policy framework through its national AI plan, published in November 2025. It organized the agenda around economic growth, education, equitable delivery, and trust.

That wider plan strengthens coordination, but it also expands the number of promises requiring evidence. Language technology now sits beside health, education, public administration, skills, infrastructure, and economic development.

The risk is that Welsh support becomes a cross-cutting principle without an accountable owner. When every program must consider language, no single program may publish the outcome.

Independent scrutiny has already sharpened the issue. A 2026 Senedd committee report said witnesses viewed Wales as relatively advanced among minority-language technology programs.

The same evidence compared Welsh with Catalan and Basque and emphasized the need to maintain momentum. That qualification matters because platform development cycles move faster than most public policy cycles.

A dataset released in 2024 can remain available for years. A commercial AI interface can change several times within one year. Support can improve or regress without a formal announcement.

The government should therefore separate four types of measurement.

First, asset metrics count corpora, recordings, voices, terms, and annotations. Second, integration metrics track products and services that use those assets.

Third, performance metrics measure accuracy and task completion. Fourth, adoption metrics show whether people actually choose Welsh for their work and daily activities.

Current public updates are strongest on the first category. Teams and other partnerships offer evidence for the second. The third and fourth categories remain harder to assess from the published record.

This limitation also applies to claims about readiness. Wales is better prepared because it possesses more resources, relationships, and institutional experience than it had in 2018.

Preparedness is not the same as parity. It does not mean Welsh users receive the same feature coverage, accuracy, or update schedule as English users.

The safest conclusion is therefore narrower than the headline. Wales has built a credible foundation and a repeatable partnership model. The next phase must prove outcomes at product scale.

Three Signals Will Show Whether the Strategy Is Working

The next meaningful evidence will come from independent testing, visible product integration, and sustained user adoption.

The first signal is a recurring public benchmark for Welsh AI. It should compare major models and speech systems using disclosed tasks, representative speakers, and human evaluation.

A benchmark would make regression visible. If a model update weakens Welsh performance, researchers and buyers could identify the change instead of relying on anecdotes.

It would also improve procurement. Public bodies could establish minimum requirements for transcription, translation, accessibility, and generated responses.

If such testing appears and vendors participate, it would strengthen the argument that Wales has moved from asset creation to accountable deployment. Without it, progress claims will remain difficult to compare.

The second signal is broader integration inside mainstream products. Teams interpretation offers a useful precedent because it connects Welsh language needs with software already used by organizations.

The same standard should apply to meeting summaries, document assistants, search, voice interfaces, and customer-service systems. Support needs to cover an entire task, not one isolated feature.

Product documentation should state where Welsh works, where it does not, and which quality controls apply. Quiet partial support creates unrealistic expectations and discourages users after failures.

New integrations would strengthen the Welsh approach because they would demonstrate leverage over global platforms. A long list of datasets without corresponding integrations would weaken the case for the current strategy.

The third signal is measurable daily use. The government’s policy aims to increase the practical use of Welsh, especially in workplaces. That outcome requires more than downloads or feature availability.

Useful indicators could include the share of bilingual meetings using interpretation, Welsh dictation volumes, adoption among public bodies, and task-completion rates across languages.

Surveys should also ask why people return to English. The answer may involve accuracy, habit, colleague preferences, interface defaults, training, or missing terminology.

Those findings would guide spending more effectively than broad satisfaction scores. A technical fix cannot solve a workplace culture problem, while awareness campaigns cannot repair inaccurate transcription.

Growing use would strengthen the conclusion that public digital infrastructure can influence language behavior. Flat use, despite more tools, would indicate that policy has focused too heavily on supply.

The Google News story is valuable because it brings an overlooked technology program into a wider AI conversation. However, aggregation compresses a six-year policy record into a headline that sounds more conclusive than the evidence.

Wales has done substantial foundational work. It started before the current generative AI surge, released reusable resources, and established relationships with major vendors.

The harder phase has arrived. Officials must keep datasets current while testing systems that change quickly. Vendors must make Welsh support visible, consistent, and measurable.

Developers and enterprise buyers should ask direct questions before treating a language option as complete. Which tasks were tested, which speakers were represented, and what happens after a model update?

Knowledge workers should also compare real bilingual workflows, not polished demonstrations. Test meetings, search, transcription, drafting, and retrieval with the language patterns people actually use.

Watch the next Welsh AI update for those three signals: public benchmarks, complete product integrations, and adoption data. They will reveal whether the strategy is changing daily technology use or mainly expanding the inventory of available resources.

That is the question behind the Google News headline. Wales has assembled many of the ingredients for Welsh-language AI. Now it must show that people can depend on the finished experience.