AI | Artificial intelligence in translation and co.
AI is constantly changing. And extremely fast. Which is why most language services using AI technology currently still have a relatively short half-life and can/must be very different depending on the given framework conditions.
With this in mind, it would not be enough to simply list our current AI services here. We would like to present the topic of “AI and translation processes” comprehensively with regularly updated information and specific tips on the optimal use of AI. This also includes application scenarios describing how we currently use AI in projects.
Contents
Status as of 15 April 2025
- The context
- AI & co. – definition and delimitation of terms
- What should be considered before and when using AI:
– Quality requirements (risk tolerance)
– Process requirements: Data protection/IT security
– Regulatory framework conditions (software validation)
– Economic aspects & production times - In which areas is deployment possible or conceivable?
- Where is AI already used in production?
- Conclusion
1. The context
Deep-learning technologies have experienced enormous leaps in quality in recent years – also with the help of machine translation (MT). Many of our clients have been using these technologies to carry out and optimize their translation processes for decades and have seen several technological leaps. Gemino has advised and supported them in this process – and will continue to do so in the future.
Since the introduction of ChatGPT – probably the best-known Large Language Model (LLM) – artificial intelligence and the expectations associated with it have been omnipresent. At Gemino, we are constantly evaluating suitable areas of application and are already using AI specifically in project processes.
What we can say with certainty: As tempting and simple as the use of AI may seem, careful evaluation and planning should always be carried out beforehand. This is the only way to actually reap the expected benefits.
2. Definition and delimitation of terms: AI/GenAI, LLMs, (N)MT, GenAI
With “AI” or “GenAI”, we almost always refer to Large Language Models (LLMs) in the context of textual content. The best known include GPT (OpenAI), Gemini (Google), Claude (Anthropic) and LLaMA (Meta). In addition, neural machine translation (NMT) is the most important and currently (still) predominantly used technology. This is based on deep learning and – like LLMs – belongs to the “artificial intelligence” category. Other types of machine translation systems, such as statistical and rule-based systems, hardly play a role today.
Read more
NMT systems are based on neural networks. These are specifically trained for translating with suitable text corpora. In contrast, LLMs are specially trained to generate texts. These texts are then used to generate texts in another language. The distinct methodologies employed by the two systems impart unique characteristics to their “translations”, each presenting its own set of advantages and disadvantages – much like any technology. This is why LLM and NMT are suitable for different areas of application. In our experience, it is not yet possible to make any generally valid statements about usage – for example, whether one technology is better suited to a particular type of text than another.
3. What should be considered before and when using AI?
When planning (translation) processes integrating AI systems, there are various aspects that should be taken into account:
Quality requirements (risk tolerance)
What requirements/expectations describe the quality that is to be achieved? Will the requirements remain unchanged from the previous specifications? What compromises are possible and/or acceptable? The issue of risks is also an essential part of the concept of quality and should always be considered, especially in areas that must comply with legal/regulatory requirements (e.g., Medical Device Regulation (MDR/IVDR) or EU Machinery Regulation).
Process requirements
Do the processes and systems currently in use allow the integration of AI systems? What conditions need to be created? Do the achievable benefits justify changes to existing, established processes and the associated investments, which can be considerable?
Data protection/IT security
What data protection and IT security requirements must be met? Is the use of cloud-based systems conceivable? These are almost always superior to systems used in your own data center in terms of technology and training quality. However, they must also comply with data protection requirements and your own IT guidelines.
Regulatory framework (software validation)
The regulatory and legal requirements of the respective industry must be taken into account. For example, in the area of medical products: MDR and ISO 13485 require software to be validated if it is used in the production of medical devices or has a process-related influence on product quality. For a successful validation, however, it must be proven that the software delivers a defined output for a defined input: The software must therefore be able to generate the desired result reliably. To the best of our knowledge, this is not yet possible in the case of LLMs (keyword “hallucination”). However, this does not mean that AI systems cannot be meaningfully integrated into translation processes in this environment other ways.
Economic aspects & production times
One thing is clear: The appeal of AI lies in potential cost savings and faster production times.
Always check in advance whether the desired benefits can actually be achieved. This is because the integration of AI into existing processes depends on the associated business objectives.
This means in practice:
What type of content is affected if a translation process switches to AI? Different types of content, or source texts, work differently with LLMs or NMT systems. There are also significant differences in quality depending on the target language and language combination. Perhaps some target languages are only available in certain systems.
If I have a lot of small translation projects and need different workflows with different systems for each language combination and content type, I need to take a closer look. This can result in only minimal time and cost savings compared to the traditional translation process – in the worst case, it can even lead to additional costs. In this case, a workflow variant with a little more human involvement may be more suitable than a largely automated workflow in which the quality is primarily characterized by the AI translation.
4. In which areas is the use of AI possible or conceivable?
First and foremost, of course, is machine translation. AI-based translation systems offer specially designed large language models (LLMs) depending on their use. In many cases, they are an interesting additional alternative to NMT-based systems. Depending on the application and general conditions, they may be the better alternative. For example, when integrating reference texts into the target language (e.g., prompt augmentation, RAG, etc.). Here, the AI-generated translation can be adapted to the desired language and take into account language characteristics such as style, terminology, or other aspects.
AI-supported quality management
Another area of application that we are currently investigating is the quality management of translation processes. LLMs can support quality assurance through new types of testing that were previously not feasible in terms of time and budget. Even if fully automated checking and correction is not yet possible, we are currently testing use cases as assistance systems. The use of AI as a semi-automatic check/correction or warning system for certain types of error seems entirely conceivable.
New possibilities with agentic AI
Agentic AI is currently opening up additional application possibilities. Various AI systems are combined in workflows to form AI teams. This enables a much wider range of possible uses for AI – especially for multi-stage AI-based workflows that would be impossible to implement with a single system. One example: After translation by an LLM, the text passes through a second LLM system. In this subsequent step, the terminology is then checked and adapted to the relevant specifications and definitions.
5. Where does Gemino currently use AI “as standard”?
In addition to various new use cases for which we are currently still testing the use of AI, artificial intelligence is already regularly used in production by us and for our customers. Important here: We use the technologies exclusively in consideration of the aforementioned aspects and always in consultation with our clients.
I. Productive use in the area of translation processes
We currently offer two translation workflows with AI. We work with our clients to decide which machine translation technology to use. All relevant framework conditions are taken into account as well as the suitability of the respective technologies for the specific application.
AI-assisted Translation
- Workflow with the same workflow steps as in the “traditional” process: Translation > Revision > Quality assurance.
- In addition to hits from the translation database (TM), translators also receive supporting suggestions from an MT engine (NMT, LLM-based) during the translation step.
- Advantage: The use of existing TMs is straightforward.
- The workflow is fully compliant with the requirements of ISO 17100.
Savings potential through AI-assisted translation
Up to 15% on the pro rata translation costs of “no matches” (new words to be translated, depending on text type and complexity of terminology)
Verified AI Translation
- Workflow with the workflow steps:
Machine translation > Verification > Quality assurance. - In this workflow, the translation step is carried out automatically using an MT engine (NMT, LLM-based).
- The machine translation is then checked and verified by a translator (comparable to a revision according to ISO 17100) and final quality assurance is carried out.
- This workflow does not include revision by a second linguist and is therefore not compliant with ISO 17100.
Savings potential through verified AI translation
Up to 35% off the total translation costs (depending on text type and complexity of terminology)
II. Productive use in the area of video localization
We have also been using AI technology in the area of video localization for some time now: On the one hand, for the transcription of spoken language as the basis for a translation; on the other hand, for the creation of subtitles/speaker scripts. Speech synthesis is also increasingly being used as an alternative to voice recordings with human voices.
6. Conclusion
AI offers many new opportunities in the field of translation and multilingualism – almost every day. At Gemino, we are following this development and helping to shape it. We always try to maintain a realistic picture of opportunities and risks in order to integrate new technologies into our processes and those of our customers in a meaningful way.
What is certain: There are no fixed or “best-in-class” solutions. Every use of AI is individual and should be designed to be agile so that its use can be adapted or expanded as required. Also so that new technologies can be easily integrated at any time.
When planning and implementing new multilingual processes with AI, the involvement of several specialist disciplines in the company is required – especially from the perspective of linguistic and translation planning.
Drawing on our experience with tools and processes in the field of multilingualism and our expertise in AI, LLMs, GenAI and the like, we are happy to accompany you on your way to new, better, and individual solutions.
Artificial intelligence still works best through the use of human intelligence.