AI Localization Think Tank Looking Back at 2025 | Part 2

As 2025 comes to a close, the AI Localization Think Tank paused to reflect on the year that passed. It was a year that changed workflows and professional identities across the language industry. What was once speculative became practical. What felt disruptive in theory demanded real decisions in practice.

We asked each member three questions:

1. What single development in 2025 had the strongest impact on localization and translation work?

2. Which trend or shift from 2025 do you believe will have a long-lasting impact?

3. What did you personally learn in 2025 that changed how you approach your work?

Because the insights were so rich and varied, we’re presenting them in two parts.

This is the second part. The first part is available here.

Libor Safar

VP of Growth

What single development in 2025 had the strongest impact on localization and translation work?

The pressure. Specifically, the mounting internal pressure in nearly every organization to "implement AI." To run proof-of-concepts, then move from POCs to production. There's also counter-pressure: pushing back against the belief that translation is a solved problem. Internal localization teams face pressure from other departments adopting AI tools with "translation included," bypassing what localization teams have built over years. At the heart of it all is a struggle for control and ownership of the narrative.

Which trend or shift from 2025 do you believe will have a long-lasting impact?

The continuing fragmentation of the language tech space. Now, everyone can build their own brilliant custom language AI solution; a tool or an agent that meets their very particular (and absolutely unique) internal localization need, however small or large. At some point, scalability and longer-term maintenance of these solutions may become an issue, as will interoperability.

What did you personally learn in 2025 that changed how you approach your work? There's no one single thing that comes to mind.

Internal localization teams face pressure from other departments adopting AI tools with "translation included," bypassing what localization teams have built over years. 

Marina Pantcheva

Director Linguistic AI Services

For me, 2025 was the year the localization industry began moving from “System 1” AI toward something closer to “System 2” AI. In practice, this meant upgrading from solutions built on spontaneous, single-pass LLM outputs to multi-step workflows powered by reasoning models that can plan, call tools, query databases, and verify their own work.

This evolution led to the emergence of agentic AI that made more complex AI-driven automations feasible. As a result, we started designing AI solutions that can support full localization workflows, for instance, systems that coordinate content, tools, and people rather than simply generating a one-shot translation.

With the increasing model capability and autonomy, the old “human-in-the-loop” framing began to change, too. People started more openly talking about “human-on-the-loop,” and in some cases even “human-out-of-the-loop.” The conversation shifted from “Where should a human intervene?” to “What level of oversight and accountability is appropriate for increasingly autonomous systems?”

On a personal level, 2025 was the year I began to explore AI bias more seriously. The more I learned, the more I saw how deeply AI bias is rooted in human bias - cognitive, cultural, and historical. What surprised me most was how persistent these human biases are. Yet, I got the feeling that we sometimes expect higher standards of objectivity from LLMs than from humans. As a result, I’ve become more critical not only of AI models but of people, including myself, when evaluating reasoning, fairness, and neutrality. Any real progress on “fair” AI must begin with us: our institutions, our incentives, and our willingness to examine our own assumptions as rigorously as we examine the outputs of AI systems.

With the increasing model capability, the old “human-in-the-loop” framing began to change. People started more openly talking about “human-on-the-loop,” and in some cases even “human-out-of-the-loop.”

Marta Nieto Cayuela

Senior Localization Quality Manager

For the past four years, I have given each year a word. It has never been about labeling myself, but about choosing a compass. A way to manifest what I want the year to revolve around, and then, at the end, asking myself: did it reveal something true?

2025’s word was revelation, and when I think about the industry this year, that theme fits surprisingly well. I cannot say which development had “the strongest” impact, but there are a few that, to me, have changed the ecosystem.

First, agentic AI. We already knew LLMs were not a one-size-fits-all solution, but this was the year when some organizations stopped trying to force them into that role. Customization became essential, and not as a buzzword, but as a requirement. Understanding risks, designing solutions for specific regions or content types… that was the work. This year, guardrails were not abstract principles; they were targeted agents, anticipating the pitfalls we already knew.

Second, the operationalization of quality prediction. It is still a prediction, and shared tasks continue to evaluate and rank approaches yearly, but automated scoring is now part of the workflow. It is not theoretical anymore. It is being used and is influencing how teams talk about quality. There is a long way to go, but I must admit that adopting AI evaluation methods has gotten me closer to ML teams, and to my stakeholders too.

And last, the shift I think will have the longest impact is the mindset shift.

Localization teams are becoming more critical, working with executives to explore different approaches and moving from “Let’s do this with AI” to “Does it make sense to do this?” A small change in phrasing, a huge change in purpose.

This shift in thinking is already shaping how companies approach localization, and it does not look the same for everyone. Some are unfortunately moving toward reducing or even removing localization teams and freelance translators. But others are taking the opposite approach. Some localization teams are positioning themselves as internal experts, collaborating closely with engineering, product, and design, and leading conversations instead of reacting to them. The first group will return, because the work does not disappear, and linguistic knowledge is more relevant than ever. The second group has brought quality to the center of the entire workflow—to me it always was, but AI made it visible to everyone.

2025 was the year that brought me back to the core of why I do this work. Quality, for me, has always been a way of bringing direction where there is noise, and intention where there is speed. If revelation was my theme for 2025, then the revelation is this: I stay because something in this work still lights me up. And because the people who carry that same light (yes, maybe you who are reading this) keep pushing the industry forward.

Some localization teams are positioning themselves as internal experts, collaborating closely with engineering, product, and design, and leading conversations instead of reacting to them.

Miguel Sepulveda

Globalization Director

What single development in 2025 had the strongest impact on localization and translation work?

In 2025 I saw a shift. We moved from using AI as an “AI-assisted translation” add-on like a separate tool you only use when you needed a translation to AI-orchestrated localization workflows. People stopped seeing AI as something that simply makes translation faster. Instead, they started using it as the layer that connects content, context, style rules, user data, and quality metrics.

This change is helping us address one of the long-standing weaknesses in our industry: how to measure quality in a reliable way. Some TMS platforms now include AI-driven quality scorecards, and they are giving teams a new way to measure what we deliver, at scale and with more consistency. It’s not perfect yet, but it’s the closest we’ve been to turning quality into something truly measurable rather than subjective.

Which trend or shift from 2025 will have a long-lasting impact?

What I mentioned before about measuring quality is part of a bigger trend that will stay with us. Localization is becoming a little more data driven every year. And this is putting us in a better position to talk with product owners and leaders about things like: how content performs across markets, how cultural adapted content affect the local user experience, or how our localized product influences retention and even revenue.

What did you personally learn in 2025 that changed how you approach your work?

In 2025 I started seeing something that changed the way many product owners, product directors, and C-level leaders finally understood that AI is not a plug-and-play solution for localization. You can’t drop it into a product roadmap and expect an instant localization. Last year, I saw a lot of hype and unrealistic expectations, to be honest. Now I’m starting to see that leadership has started to adjust its expectations. They realised that removing the human layer creates gaps. And this wasn’t something I noticed only in high-profile stories, such as Klarna or Duolingo. I observe this in everyday decisions made within teams. These moments demonstrated that the “AI replaces humans” narrative does not hold up in real-world product development. I can see the industry is recalibrating (slowly)

Product owners, product directors, and C-level leaders finally understood that AI is not a plug-and-play solution for localization. You can’t drop it into a product roadmap and expect an instant localization.

Monica Albini

Researcher in AI

In 2025, the real turning point was the moment when LSPs collectively accepted AI in localization as a stable part of production rather than a future possibility. With AI actually in place rather than imagined, we finally had something concrete to evaluate, compare and negotiate. An example that illustrated this transition well was the TEF conference: a space where conversations about quality, results, obstacles and success stories were openly discussed, moving the industry past assumptions into data-backed dialogue.

The trend of the year, in my opinion, was that the youngest translators were often more forward-driven than their tutors and trainers. When interacting with the new generation, I feel now a “we’ll figure it out somehow” mentality, a willingness to step into uncertainty despite reduced prospects for “language-only” roles and unclear career paths. At the same time, while companies were embracing AI implementation with pragmatism and speed, some academic environments struggled to keep the same pace. Not all, obviously, but the gap is visible.

The pressing question has changed: first, it was “How do young translators align themselves with what companies want?”, but in 2025, the real question was: “How do academic institutions align themselves with what the industry is already doing, and ensure they are creating graduates who are competitive in today’s market?”. Students today must often bridge that gap independently, by finding resources, tools and hands-on experience outside traditional academic frameworks, which involves extra cost and effort.

If 2025 was the year we acknowledged this gap, the hope for 2026 is to see more aligned, collaborative and future-ready training ecosystems that better prepare emerging linguists.

My hope for 2026 is then simple but transformative. In details: a closer, more transparent collaboration between universities and industry, with educators equipped to explain both the promise and the challenges of localization. Also, it’s important to share real stories of this mutable sector, with successes and failures shared openly, with updated training that prepares students for the upcoming years.

The trend of the year, in my opinion, was that the youngest translators were often more forward-driven than their tutors and trainers.

Veronica Hylak

Product Innovation Strategist and AI Vlogger

2025 was the year the world’s relationship with AI flipped.

We went from collective excitement to something closer to a low-grade resentment. Not rejection, but more like a messy, emotional love-hate phase. And honestly, I felt it too.

Two shifts defined the year for me.

First, the rise of Chinese AI models. Not just “catching up,” but reshaping the competitive landscape in a way the West wasn’t prepared for. That gap will become even more obvious in 2026.

Second, the cultural impatience with AI prototypes.

People finally stopped clapping for demos. Everyone wanted real products that worked reliably in production, and the uncomfortable truth is that most companies didn’t know how to make that jump.

Too many teams wanted to be part of the AI gold rush without doing the unglamorous engineering or stepping aside for people who could. So we saw billions poured into pilots that went nowhere, endless talk about AI replacing jobs, and workers being laid off anyway.

Yet, even with all the frustration, I still find myself drawn to the architectural breakthroughs happening under the hood. We’re watching engineers experiment with systems that mimic the human brain, and exploring designs that could reshape everyday life as dramatically as the car reshaped transportation only 100 years ago.

We saw billions poured into pilots that went nowhere, endless talk about AI replacing jobs, and workers being laid off anyway.

Yota Georgakopoulou

Media Localization Consultant

Speaking about media localization specifically, the most important trend in 2025 has been the maturation of AI dubbing. We no longer see it only in demos but in real-world production – fully or partially dubbed with AI, with varying levels of human involvement or none at all. It’s still the exception of course rather than the rule, but this trend will only continue in 2026, as more service providers build serious, production-ready AI pipelines and standards are developed for this new product line.

I believe this type of automation marks the beginning of a new era in media localization, where AI plays a central role in subtitling and dubbing alike. Ideally, one that democratizes access to subtitled and dubbed content, rather than one that uses AI just as a blunt cost (and I’m afraid also quality) cutting tool.

Yes, I am all for automatic subtitles that make content accessible to all. But I also want the option of creative ones that really bring out the director’s intent and help me immerse myself in the story, that complement it rather than flatten it into a mere transcription. The same applies to dubbing. Sometimes I simply want to understand the dialogue in my own language. Other times I want to enjoy the performance and let it transport me in the story itself.

My hope is that, as an industry, we finally acknowledge the different quality levels that have always existed (even when it was just people doing the work) and that not all translation is the same. That we shine the spotlight on the “wow factor”, the creativity in translation when we see it, that we celebrate it and seek it out.

In my ideal world, the introduction of this new generation of language technologies could help us better define achievable quality levels and be transparent about the level of professional effort in each, so that pricing models can also evolve accordingly and fairly. So that “fit for purpose” can truly encompass the full spectrum of quality and purposes, including excellence and not just the bare minimum. So that high-quality professional work stands out, is recognized and compensated for the impact it has. So that the “wow factor” in translation continues to wow us for generations to come and (to paraphrase Änne Troester) our hearts can continue to be touched by what comes out of another human soul.

The introduction of this new generation of language technologies could help us better define achievable quality levels and be transparent about the level of professional effort in each, so that pricing models can also evolve accordingly and fairly.