LLM vs NMT: how AI is revolutionising machine translation for websites and social media

Machine translation has become a key tool for fast publishing, handling large volumes and reaching international audiences. However, large-scale, organised translation is often confused with one-off translation done manually in a tool. This difference explains why quality can vary so widely from one project to another.

Today, two major approaches stand out: neural machine translation (NMT) and large language models (LLMs). Both can translate, but they function differently and have different strengths, especially in web and social media contexts, where context, speed and the risk of misinterpretation are highly significant.

In this article, you will learn how machine translation actually works on websites and social media platforms, the difference between NMT and LLMs and current limitations, illustrated with a clear example: Luxembourgish, a “low-resource” language in terms of AI resources, and therefore more difficult to translate.

The difference between NMT and LLMs: two approaches, two outcomes

Although NMT and LLMs use some similar technologies, their initial objectives are different. NMT was specifically developed for translation, whereas LLMs were created to understand and generate text for a wide range of applications, including translation.

In this context, NMT is particularly well suited to structured content such as websites, documentation and product sheets. It also works well for projects with large volumes or tight deadlines. In a professional context, it delivers more predictable results, especially when combined with glossaries and consistency rules.

By contrast, an LLM is more versatile. It can translate, but also summarise, rephrase or adapt text to a specific tone. It is often useful when a more natural phrasing is needed or when a broader context must be taken into account. However, it needs to be carefully guided: without clear instructions, it may stray from the source text, introduce nuances or rephrase too freely.

How machine translation works for a website

It is also necessary to manage content blocks, variables (prices, currencies, units), SEO tags, menus, filters, and sometimes non-translatable elements such as product names, codes and references. In practice, translating a website using machine translation involves a multi-stage pipeline.

When it comes to websites, the challenge is not simply to translate pages one by one. Website translation must also preserve consistency across menus, product sheets, SEO keywords, calls to action and overall tone, through a structured approach that links tools, the CMS, terminology and quality control.

The content management system (CMS) or connector retrieves the website content to be translated, including visible text, product-related information and certain technical elements used for SEO and accessibility. When this extraction is done properly, it helps reduce common errors such as translating elements that should remain unchanged, breaking HTML or incorrectly segmenting content.

The content is sent to a translation engine via an application programming interface (API). Two approaches are commonly used: neural machine translation (NMT), which is well suited to large volumes thanks to its efficiency and stability, and large language models (LLMs), which are more useful for refining style, rephrasing, simplifying or adapting text to a specific audience.

More and more teams are adopting a hybrid process: neural machine translation (NMT) is used to produce an initial version, then large language models (LLMs) help to rephrase and harmonise the text, with appropriate quality checks. Sensitive content is then validated prior to publication. The aim is to improve speed while preserving terminological consistency and accuracy.

Post-editing is often underestimated. It covers all the review and correction steps performed after machine translation in order to improve the quality of the final text. This step helps make the translation more consistent, natural and aligned with the guidelines. This is where terminology rules (glossaries), format checks (tags, spaces, punctuation) and safeguards (to prevent certain elements from being translated) are applied. Without post-editing, clear inconsistencies emerge, such as the same term translated differently across pages, distorted slogans and incorrectly rendered units.

According to Google Search Central guidelines, for a multilingual website it is important to correctly structure localised versions for search engines, in particular through hreflang tags en-gb → English (United Kingdom).

How machine translation works on social media

On social media, the context is different: content is short, highly contextual, often implicit, and spreads quickly. Machine translation therefore serves a key purpose: reducing the language barrier in order to improve engagement (views, comments and shares). However, an accurate translation is not always enough: the message also needs to sound natural, be compelling, and suit the target audience. In this context, marketing translation helps preserve the brand voice while avoiding overly literal or unengaging wording.

Social media clearly illustrates the limitations of machine translation when used without human oversight. Between loss of context, inappropriate tone and risks to brand image, the user experience can quickly be affected. This is a topic we explore further in our article on the compliance of machine translation on social media.

On certain platforms such as Meta Platforms, translation can be triggered automatically or at the user’s request, without direct control over how it is displayed. Some platforms have historically shifted heavily towards NMT to improve quality and speed.

According to Gigazine, changes to the recommendation algorithm on X, combined with machine translation of content, can lead to heavy exposure to foreign-language posts, such as tweets in Japanese, transforming the overall user experience on the platform. For brands, this reinforces the importance of clear messages, a consistent tone and a validation process for sensitive content.

The limitations of machine translation: what to be aware of

Machine translation is evolving at a fast pace, but it still has structural limitations that should be understood before rolling it out at scale. The first is ambiguity: on social media, a short sentence may depend on context, including irony, innuendo and cultural references. NMT can be too literal, whereas an LLM may, by contrast, over-interpret the meaning.

The second limitation relates to terminology: in technical, legal or medical fields, an approximate translation can have consequences for compliance, safety or liability. Even with advanced AI, quality checks are necessary.

Luxembourgish illustrates a major challenge: some languages are supported by fewer resources, such as texts, parallel corpora and linguistic data. As a result, models struggle to produce reliable and consistent translations. According to Silicon Luxembourg, AI still struggles with Luxembourgish in more advanced language tasks, illustrating the limitations of current models for languages with fewer digital resources.

At the same time, a number of initiatives are seeking to improve the inclusion of less widely represented languages in AI technologies. According to a joint announcement by Microsoft and the University of Luxembourg, the LINGUA initiative aims to strengthen the inclusion of under-represented languages, such as Luxembourgish, in future AI systems. Practical takeaway: the fewer resources available for a language, the greater the risk of deploying machine translation without safeguards such as glossaries, human review, testing and monitoring.

Conclusion

The challenge is not simply choosing between NMT and LLMs. What matters is the ability to deploy machine translation designed for websites and social media, with a well-defined workflow that includes clean extraction, the right translation engine for the content, safeguards, controlled publication and appropriate quality control. Luxembourgish serves as an important reminder that AI does not perform equally well across all languages.

For companies, working with a translation company helps structure these choices, secure quality and maintain a consistent multilingual process over the long term. The lower-resource the language, the stronger the framework needs to be in order to avoid hidden errors... until they eventually become visible to the public.