Translation as Infrastructure

Automation Advocacy

TAUS began as the Translation Automation User Society: a think tank for the global language industry, which was just starting to utilize machine translation (MT). To us, it was obvious that improving and implementing this technology was the key to achieving our grand vision — and that the only thing we still needed to make automatic translation better was more and higher-quality data.

In those think tank years, TAUS advocated for automated and data-driven workflows, introducing concepts like Fully Automatic Useful Translation (FAUT) and designing dynamic quality frameworks to move away from the old static quality standard. But while research and development progressed, the appetite for these concepts in the wider industry was limited. Companies were intrigued by the promise of automated workflows, yet they were cautious and hesitant; the risks seemed high, the technology untested, and the business case uncertain.

Radical changes were imminent for the translation profession, yet adoption lagged behind and the economic reality slowed progress. Business incentives hadn’t aligned with technological potential, and as a result, the industry moved cautiously.

Looking back, these delays seem almost inevitable. Innovation rarely moves in a straight line, especially in an operational, decentralized, and risk-averse industry like ours. Technology was often ready long before the mindset was. That created moments when it felt, frankly, like TAUS was speaking into a headwind. At times, our push toward automation was misinterpreted as a disregard for human intelligence and the craftsmanship in translation. But that was never the case. Even back then, our argument was that technology should protect the profession by taking over repetitive, mechanical tasks that never required human insight and by freeing linguists to focus on their creativity.

A Data-Centered World

In 2018, TAUS started preparing for the AI-driven future. The transformation was built on a massive multilingual data repository launched 10 years earlier. Natural language processing (NLP) engineers cleaned, enhanced, and tuned this data with the latest techniques, and TAUS provided it — along with in-house expertise — to nearly all MT and AI model developers.

Slowly, the revolution TAUS envisioned — a world of hyper-automated, global, data-rich language workflows — arrived and is still with us today. However, it has taken a form that is undeniably more corporate, scaled, and commercially driven than early idealists may have hoped for. But this is also the natural evolution of any technology. Once the underlying science matures, the business side eventually wakes up. That is what happened here: Business models, procurement, and content volumes caught up with the technological possibilities. The intervening years were merely a waiting period for economic incentives to align.

Unfortunately, the inefficiency of human translation didn’t disappear with the adoption of automation. It just took a new shape. Where we once spent time on repetitive, mechanical translation, we now spend it reviewing and correcting MT output that may actually already be good enough. And that is exactly why technologies like quality estimation (QE) and automatic post-editing (APE) have become essential. They reduce unnecessary review cycles and ensure that human effort is reserved for decisions that truly require human judgment. After all, the goal was never human-free translation; it was to eliminate human waste so linguists can focus on what matters most.

Yes, automation has made processes cheaper, but cost reduction was never the real promise. Automation forced the industry to confront what human intelligence in translation is actually for: expertise in nuance, meaning, culture, and intent — all the elements machines still cannot replicate. If automation is used only to drive prices down, the industry misses the opportunity entirely. But if it is used to rethink workflows, redesign roles, and elevate human contribution, then it not only makes us cheaper but also a lot smarter.

The Inflection Point

The past couple of years saw the acceleration of AI into the mainstream. Large language models (LLMs) became central to every conversation and use case. Suddenly, everyone seemed to be high on AI, and language and translation were a central test case for it. Specialized providers with decades of knowledge and expertise were sought to bypass entirely as translations, quality evaluation, post-editing, and more complex language tasks were run directly through AI in the hopes of saving time and money. Automation suddenly went from optional to “sufficient.” Of course, traditional workflows did not disappear overnight, but the old linear processes, manual quality checks, and static service models were certainly exposed as too slow, rigid, and costly to meet the speed and scale that businesses now expected.

The excitement was palpable, but reality soon tempered the hype: Not every task could be solved by simply running content through an AI model. Integrating AI and LLMs successfully into traditional workflows is far more complex. For example, LLMs struggle with QE and APE tasks, outputting unreliable and inconsistent results. Without careful training and fine-tuning, running everything through an LLM may save time in the short term but risks introducing errors, inconsistencies, and potential privacy and copyright issues.

The disruption of AI in our industry clarified the role of every stakeholder. Technology companies pushed the limits of what machines could generate, and enterprises demanded faster, cheaper, and more dynamic content pipelines. In the middle stood the translation industry and its language service providers (LSPs): highly knowledgeable, deeply specialized, but structurally unprepared for this pace of change. The opportunity was clear: not to resist AI, but to shape how it enters our space. The question shifted from, “Will AI take over our jobs?,” to, “How do we integrate AI responsibly, efficiently, and with quality at the core?” And that is where TAUS positioned itself.

TAUS’s ambition was always to have a product in the market that could help the industry realize the vision of content available in every language to every human being on the planet. That opportunity arose in 2021 when Uber requested our help creating a model that automatically filtered poor-quality output from its MT engines. The data that TAUS had spent years curating now enabled the creation of the first version of the EPIC application programming interface (API). This experience highlighted TAUS’s evolving role from provider of data and research to developer helping organizations apply AI effectively and responsibly.