25/08/2025

DeepL or ChatGPT promise fast translations at the touch of a button – but what if the result is disappointing? Many companies are facing precisely this challenge: although the translation by artificial intelligence sounds fluent, a closer look reveals errors in the technical terminology and inaccuracies in the content. We explain how to deal with faulty AI translations and when professional translation processes are essential.

Many people know the situation: the project deadline is approaching, the texts are ready – now “only” a translation is needed. Can’t an AI do that too? DeepL or ChatGPT promise quick assistance, because with just a few clicks any text is available in the desired target language and appears to read smoothly and coherently. The disillusionment only comes on closer inspection: as soon as competent native speakers check the result, it becomes clear that the devil, as so often, is in the detail. Technical terms are not correct, formulations are misleading and there are often errors in content.

This situation is typical for AI-generated translations. This is because ChatGPT and other large language models create texts on the basis of word probabilities from the trained text corpus. AI language models are therefore simply trained to predict the most appropriate next word in a text in a constantly repeated process. Due to the sheer unimaginable size of the training material, the AI output usually sounds very good. This is an aspect that makes it much more difficult to spot errors, because as with humans, eloquence often masks weaknesses in content.

What is clearly missing, however, is “knowledge”, i.e. an understanding of what is to be conveyed in terms of content. Since this cannot be covered by word probabilities, humans remain irreplaceable as language and technical experts. Only they understand the content, have the necessary background and world knowledge, and can decode and correctly convey implicit information or ambiguities.

Anyone who nevertheless uses ChatGPT or similar tools for translation should be aware of how these systems work. The AI has no prior knowledge of your company, your industry, your products or the way you communicate. As a rule, these systems start from scratch with every prompt: they have no memory of previous translations, nor can they establish consistency with previous projects.

This is where “context engineering” comes into play, i.e. the art of providing the AI with all the necessary background information. The more fully and precisely you define the context, the more likely the translation is to succeed. The most important contextual information includes industry and domain (e.g. mechanical engineering, medical technology, IT), text type (e.g. operating instructions, marketing brochure, technical specification), target group (e.g. specialist audience or end users), intended use and special terminology and style requirements.

Existing translations or sample texts can also help the system to recognise your corporate style, provided you clearly mark them as context examples only.

However, there is no guarantee of correct implementation, as AI systems do not work deterministically. Identical input usually produces different results, which means that the output can never be predicted with certainty.

AI translations must therefore be carefully checked before being passed on to downstream processes. Under no circumstances should unchecked texts be passed on to desktop publishing (DTP) or other downstream processing stages, as subsequently discovered errors and the necessary corrections can cause considerable additional costs.

Since the necessary checking requires both a good command of the source and target languages and specialised knowledge, this is where professional translation service providers ideally come into play.

This is because the evaluation of AI translations shows that the actual expense for the company is often not fully priced in. Identifying errors, researching correct technical terms, revising inappropriate formulations and creating internal consistency – all of these activities are necessary, but they require qualified employees who generally have other core tasks and are therefore usually occupied with the tasks for much longer than professional translators would be. The lack of reproducibility of AI translations exacerbates this problem, as even identical texts are translated differently from one prompt to the next. Accordingly, each output requires a complete review, which further increases the time required.

The solution is not to do without machine translation, but to use it professionally. As part of MTPE (Machine Translation + Post-Editing), professional translation service providers do use AI systems, but embedded in established quality processes. The decisive difference lies in systematic post-processing by experts.

They not only have native language skills, but also industry-specific expertise. They recognise where machine translation systems have misinterpreted ambiguities, where implicit information has been lost or cultural nuances have been misunderstood. First and foremost, however, they find linguistic and content-related errors and also adapt a text stylistically. The result is a high-quality text from an AI-generated raw translation.

As another plus point, in professional translation, AI is just one of several valuable resources that support the translation process. By setting up and using translation memories (TMs), texts are stored sentence by sentence and are available in this form for subsequent projects. This guarantees consistency and enables significant cost and time savings, as texts do not have to be translated and checked again and again. AI is then only used where sentences need to be completely translated from scratch. Specified terminology in the form of a database or a list also make a significant contribution to the quality of a translation and enable the technically precise and verifiable use of specialised terminology. All three resources – AI, TM and terminology – can be combined in a CAT tool for translation to achieve the best results.

Structured quality assurance offers additional security, especially for technical documentation where liability issues are relevant. A company’s individual specifications are also checked to ensure compliance with all requirements and thus the highest quality. Professional translation systems such as our oneSuite also enable translations to be checked directly in the original layout before costly rounds of corrections arise at the desktop publishing (DTP) stage.

The decision between ad hoc AI translation and professional translation management is ultimately a question of priorities. For internal communication or initial drafts, tools such as ChatGPT or DeepL may well fulfil their purpose. For technical documentation, operating instructions or legally relevant texts, on the other hand, quality, accuracy and a long-term perspective are what count.

The long-term economic advantage of professional translations is achieved through the systematic reuse of verified content, the continuous development of terminology databases and seamless integration into existing documentation processes. While with AI tools each translation remains an isolated event, professional translation management allows you to build up a wealth of knowledge that grows with your company. By combining AI technology and human expertise, you get the best of both worlds – and stay on the safe side in terms of quality.

Would you like to professionally enhance your AI translations or rely on a reliable translation strategy right from the start? Our experts analyse your specific requirements and work with you to develop a customised solution. Arrange your no-obligation consultation – we will show you how machine translation can be turned into professional translation quality.