Isabelle Andrieu:
The Power of Adaptation

Language technology has been Translated’s focus since its inception in 1999. Could you give us insight into how it’s evolved in over 25 years? What key changes and new capabilities emerged during that time?
What matters is not the technology, but how it behaves in production environments and improves over time. We founded Translated at the dawn of the Internet and have progressively built a production learning system based on state-of-the-art technologies.

The first generation of tools for translation tried to structure language. They were useful, but they captured fragments. Translation operates through context, exposure, and judgment. That is why the major shift came with the introduction of the Transformer architecture, which today powers large language models like ChatGPT. It changed how machines process language.

We worked closely with some of the researchers behind the Transformer. This allowed us to bring these advances early into production, making ModernMT one of the first commercial applications of this architecture.

What has evolved in recent years is the integration of context. We moved from segment-based systems to models that consider full documents, instructions, and external signals, such as Lara. This reflects how translators actually work. They build meaning across the entire text.

Still, these systems lack grounding in real-world experience. Language is deeply connected to experience. Understanding comes from interaction with the world. With the DVPS project, we are exploring multimodal foundation models that learn through direct interaction with the physical world. And I can tell you they will come earlier than expected.

You’ve since adopted an AI-forward model that blurs the line between AI and human translation. Could you tell us when you adopted this strategy and how has it progressed since then?
This was never a strategic pivot. It is how we built the company from the beginning. Most AI systems are trained once and deployed. Our system improves continuously through production usage, learning from professional decisions at scale.

This comes from a decision we made early on. As a translator, I believed machines could support us only if they learned how we reason about language. We focused on building a continuous feedback loop in production, where every correction, approval, and choice contributes to improving the system. This turns AI from a static tool into a system that evolves with usage and a genuine partner for the translator.

This intuition shaped the way we designed our systems. In 2015, we introduced the Matecat computer-assisted translation (CAT) tool, which was powered by adaptive machine translation that learns directly from translators’ edits. It was one of the first moments where you could see a feedback loop forming, where human input was immediately improving machine output. With ModernMT in 2017, we introduced context awareness in machine translation. The system started to look beyond individual segments and incorporate surrounding information. This was an important step, because it brought the machine closer to how translators actually work.

With Lara in 2024, we reached a new level of integration. Humans and AI now operate on the same unit of work, the full document, with access to instructions and external inputs. They contribute to a shared output. The process becomes continuous rather than sequential.

You’ve trained your translation AI, Lara, with the goal of having it think like a translator. Can you tell us how you accomplished this? What separates it from other translation AIs, and how has it performed?
Every translation project generates signals, choices, corrections, and discussions. We decided early to capture this process, not only the final output. Over time, this created a continuous stream of production decisions that reflects how professionals evaluate and refine meaning.

Lara is built on this continuous learning loop. Each translation generates a cycle: output, human decision, system update, improved output. This loop runs in production across enterprise workflows, capturing how meaning is validated in real conditions.

This creates a structural difference. Traditional AI systems rely on static training data. Our system learns continuously from real usage. This feedback loop is proprietary and replicating it’d require years of coordinated production activity, not just model training.

This changes the role of the translator. With Lara, professionals shift their time from fixing recurring issues to improving meaning, tone, and communication effectiveness, building a feedback loop that benefits both and allows for the production system evolution.

This also changes how quality improves over time. Many systems require heavy post-editing to reach acceptable quality. We designed a system that supports translators in producing better outcomes from the start. It improves through their interaction, reducing edit effort and turnaround times, and ultimately creating a cycle of compounding quality improvement.

Tell us about the various Translated products that Lara powers. What markets are you tapping into with each iteration?
The most important integration is actually TranslationOS, which is not a product. TranslationOS represents an infrastructure layer for global communication. It orchestrates AI and language professionals across enterprise workflows, adapting dynamically based on content value and context. This enables AI-driven localization, making the whole process seamless, transparent, and predictable.

Lara is also integrated in all of our tools for professional translators and creatives: Matecat for text, Matesub for subtitles, and Matedub for dubbing. With our SaaS platform Laratranslate.com we’re also addressing the consumer market, which is now reaching a mature level. Obviously, we have an increasing number of third-party applications including Lara as a translation AI.

With Lara, we are addressing the huge opportunity arising from the gap between the amount of content created and the amount of content people can truly understand. As more countries build a strong digital presence, more content is created and shared globally. At the same time, people expect to interact in their own language and to see their values resonating in every message they receive.

This creates a growing demand for translation across all markets. Companies are starting to respond to this. They move from translating only what is necessary to translating everything, and then deciding where human expertise creates the most impact.

We are moving from selective translation to universal understanding.

Translated also boasts the world’s largest network of translators. Can you give us insight into how the labor balance works between your AI products and your human translators?
We approach this as a system design question, not a question of balance. We must distinguish between the most common language combination and the long tail ones, which don’t benefit from large datasets for train translation AI systems and still require a lot of language professional involvement. As we all know from a recent CSA Research paper, only a very small fraction of the content created in the most-spoken languages every day is translated, and 99 percent of translated content is produced by AI. This is also true for companies working with us. What is interesting is how this evolves over time.

Most of our customers consider localization a driver for growth, and they continuously increase the amount of content they translate. A larger share of that content is handled by AI, while their investment in professional translators continues to grow. This is where the real difference emerges. When companies invest in translators as part of the production system, they see such improvements as reduced editing effort, faster turnaround times, and compounding quality increase as a consequence of continuous learning from human edits. When they try to minimize the role of human expertise, turnaround times and overall quality are affected by critical issues arising downstream.

The reason is simple. AI can generate output at scale, but meaning, alignment, and effectiveness still require human judgment. If that layer is underinvested, the system becomes less efficient overall.

When it comes to translators, we have observed that those who embraced AI tools since the beginning have now improved their revenues up to four times in just seven years.

This human-in-the-loop model is what we have been building since the inception of the company. AI expands what can be translated. Humans ensure that what is translated actually resonates with its audience. This only works because the system continuously learns from production decisions, which static models cannot do.