State of the Language Industry 2025

Introduction

LocWorld54 — which my friend Stefan Huyghe consistently refers to as the "Super Bowl" of the language industry — is now a few days behind us. The conference was great (kudos to the organizers), and I had a wonderful time catching up with old friends and meeting new people. Nonetheless, after I spent the post-conference weekend in San Francisco reflecting on it all, I still can't get around the overall sentiment of anxiety and a fair amount of AI saturation underneath the many (politically correct) presentations, product demos, and success stories.

What was planned to be a brief LinkedIn post quickly grew into a full article, and has ultimately resulted in an open letter to the language industry. The central idea:

The technology isn't the problem; our way of implementing it is.

While I graduated as a translator (and genuinely loved translation studies), I realized on the first day of my internship at Yamagata Europe back in 2013 that this industry is, at its core, driven by technology, which I immediately prioritized as my primary field of focus. Since the eye of the tech enthusiast may be biased, all opinions in this piece are shaped by more than a decade of experience in language technology on the one hand, as well as my background as a linguist (speaking four languages) on the other.

The Quality Debate

Like many of you, I'm a frequent Uber user. I could highlight a couple of linguistic mistakes and unnatural phrasing in Dutch, my mother tongue, both in the app and in Uber's email communication. I could also compile them into a big "AI fail" post to demonstrate how "stupid" AI actually is. But I firmly refuse to do this.

Uber gives me everything I need for the task at hand: I can book a ride in a matter of minutes, chat about the weather and sports with the driver, and take care of all the post-ride admin — always with a positive user experience, despite the occasional translation glitches. While I highly appreciate having access to everything in my native language, I simply don't care about them (fit for purpose, anyone?).

Saying that the resilient and crystallized language industry is going through tough times still feels like a euphemism. The ongoing boxing match between AI optimists and naysayers has our beautiful industry in a stranglehold that seems increasingly difficult to escape. According to me, the "man vs. machine" debate is completely unnecessary, even irrelevant, and doesn't take us anywhere. On the contrary, it only reinforces the per-word rate war between LSPs and linguists — the very root of the conflict.

The Uber analogy brings me to our Pandora's box: quality. For a long time, we've associated translation quality with getting every dot and comma right. Nevertheless, in today's technology climate, quality is increasingly connected to how multilingual content performs, i.e., its impact and intent. Here's my favorite example: if 30% of your website consists of high-traffic content that drives customer acquisition and sales, that portion deserves to be a work of art, delivered by a skilled human linguist. But if a raw machine translation of the remaining 70% of low-traffic content helps attract more visitors or keeps them engaged a bit longer, you'll have achieved the exact same level of quality — from a business perspective, not linguistically of course.

Quality in localization goes much deeper than simply distinguishing business-critical from second-tier content. When I come home after a long business day and need dinner, I might consider ordering a pizza or going to a Michelin-star restaurant. While most people will instinctively claim that the restaurant offers much higher quality, I might perceive the pizza to be the better outcome based on my needs and wishes. Ultimately, quality is a matter of perception, and this also applies to translation quality. As long as we maintain the quality debate, which won't produce a winner and is as artificial as the intelligence that generated it in the first place, we'll fail to replace our obsolete QA algorithms with metrics that align with real-life KPIs and ROI.

What we're really witnessing isn't an existential crisis; it's an incremental evolution that has taken steroid proportions since the launch of ChatGPT in November 2022. Without technological progress, I would have traveled by horse from San Francisco to Monterey and back. Tech breakthroughs inevitably change how we live, often washing away things that used to feel familiar — think of public phones and movie rentals. Nonetheless, despite the disruption, telecommunications is bigger than ever. And while streaming platforms like Netflix have increased translation volumes, diversification, and standardization, they haven't exactly killed the physical movie theater either. God only knows how Uber must feel about Waymo's driverless cabs roaming the streets of San Francisco.

Waymo

The Origin of the Pain

I've given a name to the origin of our industry's pain: neural machine translation. This is highly controversial and paradoxical, especially since the transformer was introduced by Google specifically for translation before DeepL took it to new heights. The truth of the matter is that we simply weren't ready for a technology as powerful as neural MT. On the one hand, we amplified NMT's primary weaknesses such as inconsistencies and out-of-context translations through our single-sentence paradigm. On the other hand, we made our most valuable stakeholder, the human linguist, pay the price for it — literally.

For a while, we believed we could leverage our most valuable asset, the translation memory, to train NMT models across domains, verticals, and language pairs. This never quite materialized, on the one hand because model training is a cumbersome, time-consuming and therefore expensive endeavor, but also because we missed one crucial detail: only a tiny fraction of all the translated content on this planet is ever curated by human linguists or ends up in translation memories. As a result, TMs are useful for light customization layers on top of baseline models, but arguably ineffective for training models at scale. Other drops in the customization ocean such as MT glossaries, MT routing, or blackbox QE scores are noteworthy, but do more harm than good if deployed recklessly.

Meanwhile, enterprise localization programs as well as LSPs established shiny MT departments that thrived for years, but their efforts often boiled down to comparing baseline models and having sample sets evaluated by human linguists (oh the irony). Plenty of glitter, but no real traction behind the MTPE model.

When you let all of this sink in, it's hard to ignore that we created the illusion of "more, faster, better, cheaper" ourselves. We made our buyers believe that AI could do everything, and in this process, we fueled the unstoppable rise of DeepL — arguably the only company that really got NMT right.

The current AI hype is often referred to as a bubble, expected to burst sooner or later. I believe the language AI bubble has been showing serious cracks for a long time, deep enough to take some very promising companies down, including Lengoo and Unbabel. The technology wasn't the problem; our way of implementing it was.

The AI Hype

Gartner Hype Cycle

Despite a lot of good intentions, it's hard to deny that we still have a long way to go in terms of meaningful AI implementations. Many products and features are limited to some AI "sprinkling" on top of traditional workflow steps like preprocessing or translation — processes that are typically clunky, document-based, and controlled by manual triggers.

Some of the factors that continue to slow us down on the road to true business impact include inflated expectations, insufficient preparation, and a lack of clarity around purpose. Too often, AI is adopted for the wrong reasons (cutting costs no matter what) or applied to the wrong use cases (replacing developers or translators). In other words, there still isn't enough awareness of what AI can and cannot do. At the same time, many companies underestimate the effort required to deploy a meaningful AI program, i.e., the budget, expertise, time, data, processes, and technology it demands. Deploying AI isn't a matter of unplugging a language model, plugging in a new one, and expecting it to produce miracles.

When the AI hype cools down, sooner rather than later, we'll see the dawn of a new era for the language industry: an era of full-scale transformation, where localization becomes embedded in global content operations, and business outcomes outweigh processes and tools. It will be an era that forces us to rethink our 30-year legacy — from technology to pricing models and value propositions. An era where the market will mercilessly kick out those who fail to adapt, yet one in which human expertise will be as crucial as technology to translate the universe of content that is out there waiting to be translated. The technology won't be the problem; our way of implementing it may be.

Looking Ahead

In this golden age for entrepreneurship, today's everybody's chance to lead — not follow. We're dealing with dangerously (pun intended) powerful technology, and every idea has the potential to become a full-fledged use case in the space of modern corporate localization.

Having already lived through what could be labelled as an AI bubble, we now have the opportunity to turn problems from the past into assets for the future. A personal takeaway from the NMT era is that the language industry isn't primarily an industry of "builders"; we are "aggregators," "integrators," or "orchestrators". Our true value lies in developing tailored customization on top of existing technologies and ideally, we leave the creation of foundation models to the experts: academics, developers, and big tech.

Second, even though it's possible to produce highly contextualized translations with LLMs and their flexible adaptation techniques, there's an endless inventory of other potential applications — from sophisticated workflow automation to quality assurance and multimodality. If we limit our AI ambitions to machine translation, we'll miss out on the opportunity to explore the full potential of AI. I remember and particularly appreciate Morana Perić's presentation about DeepL's AI agent and its use cases, including data analysis, research, quality checks, terminology work, and workflow optimization. Instead of romanticizing "translation agents", it's refreshing to see translation as a component in an agentic workflow, not as the complex agentic outcome. Way to go.

Finally, although a reset of our industry seems inevitable, we need tangible results and short-term wins to get geared up for much-needed innovation. A targeted LLM implementation with traditional glossaries can have a lot more impact than a grand experiment that kills 30 years of technological legacy rather than rethinking it. Just because we have access to new technology doesn't mean we have to embark on new paradigms overnight. However, while the usefulness of certain resources, such as the good old term base, shouldn't be questioned, we should keep asking ourselves critical questions (How relevant are fuzzy matches really? Are vectorized memories more valuable than segment-based TM matching? etc.) and act accordingly.

I know for a fact that many companies across different market segments — automation, LSP, TMS, specialized tooling — are doing great things and have the right vision in terms of ideas as well as execution. Apart from one (ELAN Languages 🙂), I won't call them out in this post, but be my guest to DM me if you'd like my perspective. What sets them apart isn't necessarily the technology they use, but how they choose to apply it. In the end, the technology isn't the problem; our way of implementing it is.