We recently hosted a webinar where Grigory Sapunov (CTO and Co-founder of Intento) and Andrzej Zydroń (CIO of XTM International) discussed common misconceptions about Large Language Models (LLMs) in machine translation. From data privacy concerns to technical capabilities, here are their responses to the most pressing questions from the audience:
What are the data privacy concerns when using AI for sensitive client work, particularly with providers like DeepSeek, where proprietary information could potentially be leaked or stolen?
Commercial LLM providers have security measures in place. Major vendors like OpenAI, Anthropic, and Google explicitly state they don’t use customer data to train their models under commercial terms (not free tier). However, it’s important to read the terms of service carefully – for instance, DeepSeek’s API terms differ in this regard.
What is the difference between closed and open models?
Closed models keep their weights private – you can only access models through APIs or cloud systems. Open models allow you to download and use the model weights yourself.
How do LLMs achieve such strong translation abilities without specific training?
The capability likely comes from both pre-training and instruction tuning. Even early models without instruction tuning showed translation abilities (e.g., earlier GPTs). LLMs process parallel corpora and similar texts in different languages, mapping them into shared semantic spaces. While they’ve ingested parallel corpora and language textbooks, their translation abilities suggest they can associate sentences in different languages found across various internet sources.
How can we use LLMs with existing terminology databases that include definitions, usage labels, and specific translation guidelines?
The solution depends on your specific needs. While RAG (Retrieval Augmented Generation) can feed relevant termbase parts to the LLM, complex cases work better with an agentic workflow–where specialized agents handle different steps of the process, coordinated by orchestration logic.
For example, the process might start with translation memories or a fine-tuned LLM, then use a domain-expert agent to handle terminology and guidelines, followed by additional agents performing quality checks to ensure all requirements are met.
How well do LLMs handle technical, clinical, and medical translations compared to traditional MT?
LLMs can effectively handle specialized content translation, offering distinct advantages over traditional MT. They can process full texts, easily incorporate glossaries, and handle field-specific requirements. However, it’s important to carefully evaluate LLMs on your specific data and use an agentic approach to ensure your translations meet all the requirements.
Can AI miss content when processing scanned PDFs?
Results vary depending on PDF types and AI systems. In general, AI (for example, OCR systems or visual-language models) may miss content parts, especially in noisy or poor-quality scans. Sometimes, it flags unreadable content; sometimes, it doesn’t. You can build an agentic pipeline with multiple steps and quality checks along the way for more reliable results.
What about distilled models for translation? How well do they maintain multilingual capabilities?
Performance can vary, especially if distillation is done in limited languages. Like any LLM, these models need evaluation on your specific use case, including error analysis and comparison with alternatives – the same approach we use in our evaluations and our State of MT Report. DeepSeek recently released some distilled models, but we haven’t tested them yet.
Should LLM translations be stored in translation memory without human review?
You can store LLM translations in cache or storage to avoid repeating translations of identical segments. Intento Translation Storage allows for various quality checks – both human expert review and AI agents assessment – to improve stored translations if you find any issues.
Can you integrate terminology into LLMs?
Yes, LLM-based glossaries are widely used and effective in our translation workflows.
Will Google’s Willow quantum chip affect AI/LLM models?
While Willow represents an interesting development in quantum computing, this technology isn’t yet ready for practical LLM applications.
