AI content operations in Cardiff under Welsh language and UK rules

What to take away

How Cardiff public bodies and broadcasters adopt AI content tools

Cardiff runs on public information. The Welsh Government, Cardiff Council, NHS Wales bodies and the city's broadcasters all publish at volume, in two languages, to tight deadlines.

AI content tools have moved into that work. Typical uses are first draft summaries, translation drafts, subtitling, metadata tagging and search snippets. The tool drafts, a human checks and signs off.

Adoption is uneven. Some teams run pilots with one tool and a small content set. Others have written AI into their content standards. Fewer still have a review date for those rules.

Broadcasters in Cardiff, including BBC Cymru Wales and ITV Cymru Wales, work under the Ofcom Broadcasting Code and their own editorial guidelines. The code's rules on due accuracy, due impartiality and fairness apply to AI-assisted output just as they do to anything else, and a licence to broadcast is held by people, not software, so accountability stays with the human editor.

Public bodies add a second layer. Freedom of Information requests can reach drafts and prompts held by a public authority, so records need to be findable. The Freedom of Information Act 2000 sets that transparency duty.

A workable Cardiff pattern has four parts: a named owner, a written use case, a Welsh language check, and a data protection review. Teams that skip the last two end up redoing the work.

Cardiff's bilingual working culture helps here. Staff who already switch between Welsh and English spot weak tool output faster than monolingual teams, because they read the Welsh as Welsh.

Start with the process, not the tool. Map where AI touches content, then map the rules that apply. That order also helps when you document your data protection duties, because the data flow is already written down.

Welsh language standards and bilingual content production

Welsh language standards are the compliance framework for public bodies in Wales. They are made under the Welsh Language (Wales) Measure 2011, and the duties themselves are set out in regulations such as the Welsh Language Standards (No. 1) Regulations 2015 and the later instruments that followed. The Welsh Language Commissioner issues compliance notices and monitors them. The standards are not optional style guidance.

A compliance notice lists the standards that apply to your body by number, grouped as service delivery standards, policy making standards, operational standards, promotion standards and record keeping standards. Check your own notice before you automate any Welsh output, because the duties differ between bodies.

The standards require active offer and provision in Welsh. A body cannot wait for a request and then translate. It must treat Welsh as a language of service, not a fallback.

AI content tools can help with volume. They can draft a Welsh version fast. They cannot decide whether the Welsh reads naturally to a Cardiff audience, or whether a term is the one a community actually uses.

Most Cardiff teams work with general-purpose services: machine translation such as Google Translate and Microsoft Translator for Welsh drafts, and generative assistants such as ChatGPT and Microsoft Copilot for summaries, tagging and first drafts. None of them is a Welsh language editor, so the human check is the control that matters.

Use a five-stage workflow for bilingual content production.

Automated translation fails predictably. It flattens idiom, mangles place names and invents formal register where plain Welsh would do. Cardiff place names and institution names are common casualties.

Bilingual content also raises an equality point. If a Welsh speaker gets a worse page than an English speaker, the service is not equivalent. The Equality Act 2010 covers inclusive content and service delivery.

Budget for the Welsh edit. It is the step that protects quality and it is the step teams most often try to cut. Cutting it moves the cost to corrections later.

ICO requirements for AI content operations in Wales

The ICO is the UK regulator for data protection. Its guidance on AI and data protection and its AI and data protection risk toolkit set out what any AI content operation must show: a lawful basis, fairness, transparency and data minimisation.

Cardiff public bodies and broadcasters process personal data in content work more often than they think. Contributor names, case studies, audience data, source material and prompt logs can all contain personal data.

Before you process, answer four questions:

A data protection impact assessment is expected for high-risk processing. AI content work can qualify, especially where special category data or large-scale audience data is involved.

Supplier terms matter. If a tool learns from your inputs, that is a processing decision, not a default. Read the contract and record your choice.

Transparency notices should be plain. If AI drafts or translates published content, say so in your privacy information or editorial note. Vague wording invites complaints.

Your UK regulations reading list is short for most Cardiff teams: UK GDPR, the Data Protection Act 2018, FOI and your sector rules. Keep the record of how each one applies and review it when the tool changes.

Procurement and supplier checks for Cardiff buyers

Cardiff public bodies buy through UK public procurement rules. Above the relevant thresholds, the governing legislation is the Public Contracts Regulations 2015, now largely replaced for new procurements by the Procurement Act 2023. For cloud and software, the G-Cloud buyers' guide, published by the Crown Commercial Service, explains how to evaluate services on the Digital Marketplace, including security and data handling.

Use the framework to shortlist, then test the claims. A listing is not evidence. Ask for the data processing agreement, the sub-processor list and the retention schedule.

Supplier checks for AI content tools:

Run vendor due diligence before the pilot, not after. A pilot that produces good output but weak paperwork still leaves you exposed.

Watch for the common gaps. Tools that claim Welsh support often mean Welsh menus, not Welsh output. Tools that claim UK hosting may use sub-processors elsewhere.

Set service standards with the supplier in the contract: accuracy thresholds, turnaround, Welsh quality checks and complaint handling. They should be measurable, not aspirational.

VAT and contract value rules apply as normal. Check whether the supplier is VAT registered and how the contract value is calculated over the term.

Governance and review points for Welsh content

Governance is what keeps AI content operations stable after the pilot. Without it, tools spread through teams and rules lag behind.

Name one owner for AI content across the organisation. Give them authority to pause a tool. Give them a route to legal, data protection and Welsh language advice.

Set review points on a calendar and write the dates into the policy. A workable cycle is monthly sampling of published content, a quarterly review of incidents and output quality, a six-monthly supplier review, and an annual refresh of the DPIA, the supplier terms and staff training.

Keep a disclosure policy that editors can follow. It should say when AI use is declared, who approves it and how it is recorded. A short policy that people use beats a long one they ignore.

Train staff on the limits. They need to know what not to put into a tool, how to spot a bad Welsh draft and who to ask when unsure.

Keep the Welsh language impact check in every review. Ask whether Welsh speakers get the same quality, speed and prominence as English speakers.

Record decisions. If you choose not to use a tool for Welsh output, write down why. That record protects you and helps the next team.

Align with the wider rules that apply to your sector. The Ofcom Broadcasting Code for broadcasters, procurement rules for public bodies and data protection rules for everyone all point the same way: human accountability for published content.

Common questions