The European Association for Machine Translation (EAMT) is one of the leading organizations dedicated to advancing machine translation and AI-powered language technologies. Each year, its international conference brings together researchers, language technology experts, language service providers, and industry professionals to share the latest developments in machine translation, large language models (LLMs), and AI-assisted translation.
Held in Tilburg, the Netherlands on 15-18 June, EAMT 2026 provided a valuable forum to explore how prompt engineering, evaluation methods, and human-in-the-loop workflows are reshaping the future of multilingual communication.
For Landoor, a Translate.One company, the conference marked an important milestone. The team presented the research paper, “Evaluating the Effect of Prompt Language on LLM-based Translation: Evidence from Spanish<>Italian Translation,” which investigates whether the language used to write prompts can influence the quality and usefulness of AI-generated translations.
The research was conducted by Davide Maestroni, Paola Di Cataldo (HR & Talent Development Manager at Landoor), and Antonella Bove, a PhD student at Ca’ Foscari University of Venice. It forms part of a broader collaboration between Landoor and the university, with Landoor co-funding a PhD project that investigates new post-editing workflows in the AI era.
We spoke with Davide Maestroni, Account Manager & Language Technology Specialist at Landoor, where he combines client strategy with language technology and AI-driven innovation. He holds a Master’s degree in Translation, specializing in Chinese professional translation, and has a strong interest in generative AI, multilingual workflows, and the intersection of language technology and human expertise. Davide shares his impressions from EAMT 2026, the motivation behind the research presented, and his perspective on the evolving role of linguists and language service providers in the age of AI.
Davide: EAMT is one of the most important conferences for anyone working in machine translation and language technologies. What makes it particularly valuable is the mix of cutting-edge research and real-world applications. You can listen to presentations about state-of-the-art research in the morning and then discuss production workflows, client needs, and business challenges in the afternoon.
This is especially relevant today because machine translation has been used for years across a very wide range of sectors, content types, and workflows. And even with the rise of large language models, we are still talking about machine translation, although often with different technologies, different expectations, and new questions around quality, control, evaluation, and human involvement.
For me, that’s the real strength of EAMT: it’s not just about showcasing technology. It’s about bringing together people who are building it, studying it, and using it every day. You get a much clearer picture of both the opportunities and the limitations of current AI systems, and you leave with a lot of ideas, questions, and inspiration.
Davide: This was actually a first for Landoor, and honestly I couldn’t have asked for a better introduction to the EAMT community.
We’ve been researching AI and language technologies for years, but until now we had mainly attended broader AI-focused events or national translation-focused conferences. Walking into an international conference entirely dedicated to translation was a very different experience.
What struck me most was the energy. There was genuine excitement around new technologies, but at the same time a very healthy level of realism. People weren’t there to claim that AI will magically solve everything. Researchers and industry professionals were openly discussing what works, what doesn’t, and what still needs to be understood.
That balance made the conversations incredibly interesting. Compared to many general AI events, I felt there was much more focus on practical challenges such as quality, evaluation, terminology, governance, and human supervision. The field seems to be moving beyond the question of “Can we do this with AI?” and toward the much more mature question of “How do we do it well?”
Davide: LLMs were present across many sessions, but the conversation went far beyond technology alone. Topics such as accessibility, healthcare communication, low-resource languages, creativity, sustainability, ethics and evaluation were also central to the programme.
Three themes stood out to me in particular.
First, multilingual AI is clearly here, but machine translation still matters. Translation today is not only about generating text; it is about managing workflows, context, terminology, quality and risk.
Second, evaluation has become one of the biggest challenges in the industry. We are building increasingly powerful systems, but our ability to evaluate them is not evolving at the same pace. Traditional metrics often struggle to capture context, style, creativity, coherence and real usefulness.
Third, humans are staying in the loop. There was a strong consensus that the future is not about fully autonomous workflows, but about finding better ways for humans and AI to complement each other.
Davide: Prompting is often underestimated, but it has become one of the most important stages in modern AI-assisted translation workflows. With LLMs, the output can change significantly depending on how instructions are given.
When it comes to prompt design for translation, we identified two main variablers: structure and language. We had already explored prompt structure strategies during the CIUTI Conference in Milan last May, so this time we decided to focus on the language aspect.
The project actually started from a practical observation. In our daily work, we often had the feeling that translations improved when the prompt was written in the target language rather than in English. It wasn’t something we could prove, it was simply an impression that kept coming back.
At a certain point we asked ourselves: is this really happening, or are we just seeing patterns where we want to see them?
Since only a limited number of studies had explored this topic, and often with mixed results, we felt it was worth investigating it more systematically. Our goal was to move from intuition to evidence and understand whether prompt language really has an impact on translation quality and post-editing usefulness.
Davide: At its core, the study asked a simple question: when asking an AI model to translate, does the language of the prompt matter?
To test this, we created Spanish-Italian and Italian-Spanish translation tasks across two different domains: biomedical content and advertising content. We then generated translations using the same prompts written in different languages and asked language professionals to compare the outputs.
The results challenged a common assumption. Many people expect English prompts to perform best because most large language models are heavily trained on English data. However, our findings suggested the opposite: translations generated from prompts written in the target language were preferred more often for post-editing purposes.
Another interesting finding was that this effect became stronger when prompts included more contextual information. In other words, the more detailed the instructions became, the more the choice of prompt language seemed to matter.
For us, the main takeaway is that prompt design is not just a technical detail. Small changes in how we communicate with AI systems can have a meaningful impact on translation output.
Davide: The reaction was very positive, and attendees were interested in the study for different reasons. Some were surprised because they had expected English prompts to perform better. Others were more interested in the practical implications for future translation workflows.
One of the most valuable exchanges was with another researcher working on prompt engineering for Chinese journalistic translation. We quickly realized that we were looking at two sides of a similar problem. His findings suggested that while prompt language matters, prompt structure may matter even more in certain contexts.
That kind of conversation is one of the best parts of attending conferences like EAMT. It shows how collaborative research can be, especially when people approach similar questions from different perspectives.
Davide: I’m really glad this topic came up because discussions about AI often focus on performance and productivity, while governance, privacy, and ethics sometimes receive less attention than they deserve.
We often hear that data is the new oil, and after attending EAMT I believe that statement more than ever. Data is one of the most valuable assets in the AI era, and we need to be increasingly aware of where our data is stored, who owns the systems we use, and what happens to the information we share with them.
For language service providers, this is not just a technical issue. It’s a question of trust and responsibility toward clients.
I also appreciated the strong focus on inclusion. AI systems are not neutral by default, and many studies continue to show how biases can emerge in multilingual environments. Ensuring that language technologies work fairly across different languages, cultures, and user groups will become increasingly important.
Another aspect I found particularly interesting was the discussion around sustainability and smaller models. One of my favourite takeaways from the conference was hearing researchers challenge the idea that bigger always means better. In some scenarios, smaller and more specialized models can be faster, cheaper, more sustainable, and surprisingly competitive when properly fine-tuned or prompted.
I think the future will belong not only to the most powerful AI systems, but also to the most responsible ones.
Davide: This is definitely the million-dollar question!
If there’s one thing I brought home from Tilburg, it’s that this industry is changing incredibly fast, but it’s not becoming less human.
Before attending EAMT, I expected to hear a lot about technology. And I did. But what surprised me was how often the conversation came back to people: translators, reviewers, linguists, domain experts, and end users.
My biggest takeaway is that AI is not reducing the importance of linguistic expertise; it’s expanding it.
The role of linguists is evolving from simply producing translations to designing workflows, evaluating outputs, creating prompts, managing terminology, supervising AI systems, and helping organizations make informed decisions about technology.
For language service providers, I think the challenge will be finding the right balance between quality, speed, cost, and responsibility. Different content types will require different levels of automation, and helping clients navigate those choices will become an increasingly important part of our value.
Personally, I left Tilburg more convinced than ever that language technology is entering a fascinating phase.
Not because we finally have all the answers, but because we’re finally asking better questions.
And for anyone working in translation, that’s a very exciting place to be.
EAMT 2026 confirmed that the future of translation will not be defined by technology alone, but by the way technology is evaluated, governed and integrated into real linguistic workflows. For Landoor and Translate.One, participating in these conversations means staying close to the research that is shaping the industry while continuing to bring a practical, human-centred perspective to AI-assisted translation.
As large language models become increasingly present in multilingual communication, the expertise of linguists and language service providers will remain essential. Their role will not only be to produce high-quality translations, but also to guide, evaluate and improve the intelligent systems that support them.
