Pitfalls of AI translation

Many companies use AI translation as a fast and cheap alternative to professional human translation. In the right context, AI translation can be a valuable support, for example for quickly understanding content or producing a first draft of a translation.

At the same time, it is easy to overestimate the quality. An AI translation can look correct on the surface, but contain hard-to-detect errors that affect the message, nuances and tonality. For anyone who does not master both languages and is furthermore not fully acquainted with the subject area, it is almost impossible to find such errors.

In practice, this makes AI translation a matter of risk – not just cost and efficiency. That is something we at Språkbolaget repeatedly see when we review AI-generated translations.

Here are seven common pitfalls that are important to be aware of before considering using AI in real-world contexts.

1. Different language combinations produce varying quality

AI models do not perform equally well across all languages. The results depend to a large extent on how much training data is available for each language pair, which means that certain combinations work significantly better than others.

English is heavily overrepresented in most language models' training data, which means that translations between English and other common languages are often of higher quality. However, with less common language pairs, such as Swedish to Finnish or Spanish to Japanese, the risk increases for both direct errors and more subtle flaws in wording.

This becomes particularly clear in texts with specialised language, technical concepts or industry-specific terminology, where AI lacks sufficient data to make the right choice. In such cases, we often see that the translation works on the surface, but lacks precision.

Things to bear in mind: Tailor the method to the language pair and content. All AI translation must be carefully reviewed and edited, but for rarer language pairs or more specialised texts, an expert with both linguistic and subject-matter competence is required.

2. Hallucinations and added or deleted words

One of the harder-to-detect types of error in AI translation is that the language model sometimes adds information that is not present in the original text. This can involve an extra word, a shift in meaning, or a phrasing that looks plausible but changes the message. It also happens that words disappear – it is easy to imagine the consequences of one ”not” too many or too few in a warning text.

What makes such errors particularly problematic is that they are very difficult to spot. The text may be grammatically correct and well-written, whilst at the same time no longer corresponding to the original. This is a type of error that a human translator rarely makes, and which places great demands on post-editing and proofreading.

In practice, an AI translation must always be compared with the source text at a detailed level. It is not enough to read the translation as a stand-alone text. It is therefore usually not simply a matter of polishing the AI-generated text a little; rather, it can be a rather time-consuming process to ensure that no errors have slipped through.

Things to bear in mind: Do not treat an AI translation as a finished document. It is a draft that needs to be critically reviewed and systematically edited.

3. What business risks do you take when translating with AI?

Using AI translation for external communications is not just a language issue. It is a business issue. Incorrect information, the wrong tone, omitted or added words, or an unfortunate turn of phrase can distort the message, affect how your company is perceived, create misunderstandings in customer dialogue and, in the worst case, lead to errors in contracts and product information.

This sort of deviation poses a hidden risk. In many cases, the risk only becomes apparent once something has gone seriously wrong.

This means that the choice of translation method needs to be linked to potential consequences, not just to cost or efficiency.

Things to bear in mind: Carry out a simple risk assessment before choosing a translation method. The more important the text is, the more reason there is to be sceptical about quick and cheap AI translation.

4. Cultural misunderstandings and loss of nuance

A good translation involves more than just rendering the words accurately. It needs to work within its context, with the right tone, the right cultural context and the right feel for the target audience.

AI is unable to reflect on the message of a text or see the context, and is therefore unable to handle that sort of nuance. The text may be factually correct on the surface and yet come across as flat, misguided or even misleading, depending on the market.

This is particularly evident in marketing and customer communications, where the nuances of language influence how the message is received. In such contexts, it is not enough simply to edit AI translations; they must also be refined by an expert in the target language and the target audience to ensure they work as intended.

Things to bear in mind: When communicating in new markets, simply translating is not enough. The content must be adapted.

5. AI cannot convey your brand’s voice

Many companies work actively with tone of voice and expression. It is a central part of the brand and something that builds recognition over time.

AI can create texts that are grammatically correct and easy to read, but cannot understand the brand's voice. The result is often more generic than strategic, which means the communication loses its tone and distinctiveness.

In practice, this becomes apparent when different texts from the same sender start to sound as though they have been written by different people.

Things to bear in mind: Ensure that a professional with a good understanding of and feel for the language and the brand reviews and adapts the text before publication.

6. AI produces errors with great confidence

One key aspect that distinguishes AI errors from human errors is that AI never signals uncertainty. AI presents ambiguous statements and outright inaccuracies with the same confident tone as it does correct information.

This makes an AI translation come across as true and reliable, even when it is not. For anyone who does not have the full background, it is difficult to distinguish correct information from inaccuracies.

People have a tendency to ascribe intelligence, emotions and human thought to inanimate objects such as AI. That is why we choose not to suspect that anything is amiss when we see a well-written and, on the face of it, professional text. It is in our nature, but it can prove costly later on.

Things to bear in mind: A well-formulated text is not the same thing as a correct text. Information and data must always be verified.

7. Terms may become inaccurate and inconsistent

AI translates sentence by sentence and cannot see the logic and overarching narrative in a longer piece of reasoning. This means that AI translations may use different terms for the same thing within one and the same text. This is not always immediately noticeable, but can create ambiguity and lead to misunderstandings.

Deliberate variation may be in order in certain types of text, but in technical, legal and medical texts, consistent terminology is a fundamental requirement. There, it is not enough for the text to be comprehensible. It must be precise and consistent throughout the material.

The general language models that AI translation is based on are not trained on an individual company's terminology and linguistic guidelines – and we at Språkbolaget see that this type of deviation is common in practice.

Things to bear in mind: Review terminology separately from language and fluency. Ensure that the same term is used consistently for a given concept.

Summary: AI is a tool, not a guarantee

AI translation can be an effective aid in the right context. However, AI does not work equally well for all language pairs, all text types or all purposes.

The more important the text is for your brand, your customers, or your business, the greater the need for quality assurance and human linguistic expertise in the process.

At Språkbolaget, we work daily with reviewing, editing and quality-assuring texts where AI is a component in the workflow. Our focus is to use the technology where it is useful, while simultaneously ensuring that the result maintains the standard required in professional communication.

If you would like to know more about how we can help your business, please feel free to Contact us.