The UK Is More Multilingual Than You Might Think
Here’s a number that often surprises people: according to the 2021 Census, approximately 5 million people in England and Wales, around 8.9% of the population, speak a language other than English or Welsh as their main language. Polish is the most widely spoken, with over 611,000 speakers. Romanian follows with nearly 472,000, and Punjabi and Urdu aren’t far behind.
In some London boroughs, English is actually a minority language. And this diversity isn’t only a London story, Bradford, Birmingham, Leicester and many other cities have large, concentrated communities speaking specific languages. Even in Essex, where we’re based, linguistic diversity is growing. Besides English, common languages include Polish, Portuguese, Spanish, and South Asian languages.
For organisations serving the public, whether that’s a Citizens Advice bureau, a GP surgery, a local council, or a mental health charity, this means a meaningful portion of the people who need your help may struggle to access it in English alone.
What Multilingual AI Chatbots Actually Do
Modern AI chatbots are built on large language models (LLMs) trained on vast amounts of multilingual text. These models learn patterns in grammar, vocabulary, and syntax across languages, allowing them to understand user input and generate fluent responses in dozens of languages without separate training for each one. When a user writes in Polish, Romanian, Urdu, or Arabic, the model internally maps the input into a shared semantic representation, interprets intent, and produces a response in the same language, maintaining natural phrasing and context.
Unlike traditional translation systems, which rely on explicit word-to-word rules, LLMs capture idiomatic expressions, cultural nuances, and conversational tone, enabling more human-like interactions. This architecture makes multilingual support scalable, reduces reliance on human interpreters, and allows 24/7 accessibility across diverse language communities.
Translation and Trust
There’s an important distinction between a chatbot that translates and one that truly communicates. Good multilingual chatbots don’t just swap words from one language to another, they adapt tone and context too. Speaking to someone in their own language signals that you respect them and want to include them. That matters enormously when the service you’re providing is sensitive, mental health support, debt advice, safeguarding, or healthcare.
We’ve seen this play out in our own work. When we built Hart, a safeguarding chatbot for a youth organisation, one of the key design considerations was making sure vulnerable young people could feel safe engaging with the tool. Language accessibility was part of that. When someone feels heard in their own language, they’re more likely to open up, ask for help, and actually access the support they need.
The Cultural Nuance Problem: Why Translation Isn't Enough
Apparently modern LLMs actually default to Western/English-speaking cultural values even when responding in other languages. The “WEIRD” (Western, Educated, Industrialised, Rich, and Democratic) bias refers to the disproportionate reliance on data from these populations to train AI, often with 84% of research focusing on Western samples despite them representing <12% of the world’s population.
Even highly capable multilingual chatbots can stumble when cultural and linguistic nuances intersect with sensitive topics. In English, “mental health” is a broad, neutral term. In Arabic, however, there are two distinct words (عقلي & نفسي): one for general psychological wellbeing and another for severe psychiatric illness. A chatbot using the wrong term may unintentionally convey a serious diagnosis, even if the translation is grammatically correct.
The Practical Reality for Resource-Stretched Organisations
Let’s be honest: hiring multilingual staff or professional interpreters is expensive and often logistically complex. Most charities and public sector organisations simply can’t afford to have speakers of every relevant language on their team. And even when they can, availability is limited, especially out of hours. The NHS alone spends 75.5 million per year, with the most common languages being Polish, Romanian, Chinese, and Arabic. Highlighting just how costly and resource-intensive comprehensive multilingual support can be. The lack of translation services can also cost the NHS, with medication errors costing £98 million a year, causing or contributing to 1,708 deaths.
AI chatbots don’t replace the human relationships that are central to clinical work. But they can act as a first point of contact, available 24/7, in dozens of languages, handling the initial questions that would otherwise fall through the cracks. Lack of translation became a pitfall during the COVID-19 pandemic, as non-English speakers in the UK faced significant misinformation and gaps in official guidance, which sometimes put their safety at risk.
Why 95% of NHS Clinicians Rely on Google Translate (And Why That's Dangerous)
Clinicians often fall back on informal translation methods to get by amidst the shortage of translators: a study found 95.6% of staff using friends or relatives to interpret, and 76.1% turning to online tools like Google Translate when professional interpreters aren’t immediately available. Part of this is necessity, language services can be slow due to the of the lack of interpreters, but it’s also risky. Research (from 2019, couldn’t find anything more recent) has shown that Google Translate’s accuracy for languages like Urdu can be as low as 20–40%, with error rates approaching 80% in some contexts due to sparse training data and frequent mistranslations.
NHS guidance is clear: patients have a legal right to free, high-quality, professional interpretation and translation services under the Equality Act 2010 and the NHS Act 2006. More recently NHS England issued an improvement framework in May 2025 specifically addressing the use of AI and chatbots for translation, warning that while they offer convenience, they carry significant risks to patient safety and clinical accuracy. Unsupervised online translation tools like Google Translate should not be used in clinical settings, as they provide no guarantee of accuracy and can mislead or confuse patients, potentially endangering care.
Out-of-Hours Inequality: Why Language Barriers Get Worse at 9pm
Here’s something that doesn’t get talked about enough: language support often disappears after office hours. Someone in crisis must first navigate an English-language phone system, explain their situation well enough to request an interpreter, and wait while one is connected, assuming they know this service exists at the organisation they’re calling.
Larger organisations do have solutions. Emergency services like 999 can connect with professional interpreters at any time, and specialist agencies like Language Line, ClearVoice, and Language Empire offer 24-hour telephone interpreting to bridge the gap. But most smaller charities, community organisations, and local services can’t afford these interpreter services, or they don’t know the size of the non-English speaking population that could use their services. This is where multilingual chatbots can pick up the slack, they give smaller organisations a similar language accessibility that larger services could afford.
A Few Things Worth Knowing Before You Build
Multilingual capability is powerful, but it works best when it’s set up thoughtfully. A few things we’d always flag with clients:
- Test in your target languages, not just English. A chatbot might perform brilliantly in English but produce awkward or confusing responses in a less common language. Always include speakers of your key languages in your testing process.
- Be transparent about what the chatbot can and can’t do. If it’s an AI assistant, say so, across all languages. Trust is built on honesty.
- Know when to hand off to a human. The best chatbots are designed with clear escalation paths, especially for sensitive or complex situations. This is something we think hard about with every chatbot we build.
- Think about your specific community. Not all organisations need to support 80 languages. Look at your actual user data and local demographic information to prioritise the languages that matter most for your community.
Looking to make your services more accessible to everyone, in every language? Japeto Chat is an AI-powered chatbot platform making it easy to launch multilingual support quickly and safely, so you can extend your reach, reduce barriers, and ensure no one gets left behind. Discover how Japeto Chat can help your organisation.
