CYMRAEG AI: Ministers try again to teach artificial intelligence to speak Welsh

Ask an AI assistant a question in Welsh and the reply often comes back in English, in mangled Welsh, or not at all.

The Plaid Cymru government elected in May says it wants that fixed, and on Friday morning named the company it is backing to do it: Cosine, a London firm whose main model, by its own description, is built to write software.

Ministers will hand over data, resources and expertise to improve how Cosine’s Lumen model understands and processes Welsh, according to the Welsh Government.

What the announcement does not include is a price. It names no sum going in either direction, no timescale, no list of which Welsh datasets are being shared, and no date by which anyone using AI in Welsh should expect it to work better.

It is also not the first time, and not the first government to try. In June 2024, with Labour in power in Cardiff Bay, the then Cabinet Secretary for the Economy, Energy and Welsh Language, Jeremy Miles, announced a data partnership with OpenAI to improve how AI technologies work in Welsh, with the promise of open data archives for the research community.

Anna Brychan, the Cabinet Minister for Education and the Welsh Language, opened her statement much as Miles did in 2024, with the line that the language belongs to us all.

A model built for banks and defence firms

Cosine was named by the UK government among the first companies backed by its Sovereign AI programme, and is developing a model it calls Lumen Sovereign, trained in Britain on the Isambard-AI supercomputer in Bristol.

The company says that model is aimed at defence, national security, banking and other regulated industries, and that BT, HSBC, Lloyds, NatWest, BAE Systems, Babcock, Thales UK, PwC and the Alan Turing Institute signed up in June to help design it. It is targeting readiness by the end of this year.

Chief Executive of Cosine, Alistair Pullen, said Welsh was an essential part of the UK’s linguistic landscape.

“This work with the Welsh Government helps us build Lumen on a stronger foundation, with better representation of the languages and public context that matter across the UK,” he said.

Ms Brychan said the government wanted more chances for people to use Welsh in every part of their lives.

“It’s important that Welsh is fully represented in the technologies people use to access information, learn, work and communicate,” she said. “This new partnership helps to strengthen Welsh-language capability in AI.”

Cabinet Minister for Enterprise, Connectivity and Energy Adam Price, the former Plaid Cymru leader, said working with sovereign AI companies would help build the skills and capability Wales needs, and that it mattered for the language.

39% in Carmarthenshire, 11% in Swansea

The numbers underneath all this are not moving the government’s way. The 2021 census recorded 526,016 Welsh speakers across Wales, down from 538,296 a decade earlier. In Carmarthenshire that meant 72,838 speakers, against 25,986 in Swansea and 18,662 in Neath Port Talbot.

That is 39% of Carmarthenshire, 13% of Neath Port Talbot and 11% of Swansea, set against a target of a million speakers by 2050. The last time a hard figure was attached to the Welsh language it came from the outgoing Labour government in February: an £8m package that included £5,000 bursaries to move Welsh-speaking teachers into secondary schools.

The pressure sits on schools such as Ysgol Gymraeg Ystalyfera Bro Dur, whose newly qualified teachers are on S4C this month, rather than on any language model. Local readers are not uniformly warm on artificial intelligence either: Port Talbot, Llanelli and Ammanford ranked among the most anxious towns in Wales for searches about AI taking their jobs earlier this year.

Whether a better-trained model changes anyone’s day is an open question. Welsh language technology policy has been promising it since 2018.