What the First 6 Months of the 2026 Conference Circuit Taught Us

It’s July 2026, and we’ve already put a lot of miles on our bodies, er, minds. See, but that’s the thing — reporting on the language industry and where it’s headed requires boots on the ground. 

The Think Tankers traveled far and wide these past few months. Some spoke at events. Others listened intently and took notes. We’ve asked each of them to compile their learnings, big and small: things that stood out, questions asked, or, more importantly, topics that have been conspicuously absent.

Here’s what we learned:

Marina Ilari 

Over the past few months, I have had the opportunity to attend several industry events, including the CSA Research CEO Summit in Paris, LocWorld Dublin, GamesBeat Summit, and a wonderful Women in Localization event hosted at Riot Games' headquarters in Dublin. While each event had a different focus, a common theme emerged across all of them: our industry's relationship with AI is maturing.

One of the most significant shifts I observed is that the conversation has evolved from "AI for everything" to "AI, but.": AI, but with governance. AI, but with accountability. AI, but with human expertise guiding the process. Organizations are becoming increasingly sophisticated in their understanding of where AI creates value and where human judgment remains indispensable.

A particularly insightful session at LocWorld featured Kathy Mok, in charge of OpenAI's localization program. Despite being one of the companies driving the AI revolution, they emphasized the importance of human-centered localization workflows. One concept that stood out was "shippability"; the idea that content is not ready for release simply because it has been translated, but because it meets the expectations and behaviors of a specific market. Kathy shared the example of a "Contact Sales" button. Rather than relying on a direct translation, the team carefully adapted the message to ensure it would effectively encourage users in each market to take the desired action. It was a powerful reminder that localization is about influencing user behavior and creating meaningful experiences.

Another recurring topic was how AI and automation are dramatically increasing our ability to scale localization programs across more languages, markets, and content types than ever before. As scalability improves, the definition of quality is beginning to change. Historically, localization quality was measured primarily through linguistic accuracy. Today, when content can be produced and translated at unprecedented speed and volume, the question is no longer how much content we can localize, but whether that content resonates with users in meaningful ways. When volume and velocity are no longer the primary constraints, the risks change. Over-reliance on AI without meaningful human oversight can produce experiences that feel culturally generic, biased, or disconnected from the audiences they are intended to serve. Content may be technically correct, yet fail to create the authenticity and emotional connection that great localization makes possible.

These discussions reinforced my belief that the role of localization professionals is evolving. We are moving further upstream within organizations and becoming strategic partners rather than purely executional resources. Our expertise helps companies navigate culture, context, representation, and risk. Increasingly, we are asked to participate in conversations about governance, content strategy, player experience (i.e., in the context of video localization), and market readiness.

My biggest takeaway from these conferences is that the future of our industry is not a choice between humans and AI. Instead, it is about designing systems, workflows, and governance frameworks that combine the strengths of both. Technology will continue to help us scale global content at unprecedented levels, but human expertise remains essential to preserving the cultural nuance, diversity, authenticity, and empathy that make experiences truly global.

Gabriel Karandyšovský

I’ll take a slightly different route here, conveying my conference impressions (I’ve been known to straddle the fault line between realistic and cynical at times). 

I’ll echo Marina’s sentiment on one particular point: it does feel, on balance, that the language services industry is maturing. 

What I find particularly interesting about this always future-forward process (you could perhaps substitute “maturing” for “progress” here) is that everyone started from a different point, and ostensibly has an endpoint they are maturing toward. Multiple possible endpoints, I should add. It’s in the space between them that the truly illuminating stuff is happening, and its primary manifestation is the varied use cases that people are reporting at the LocWorlds and GALAs. 

I’ll mash this idea of movement forward with an observation — AI is inherently asymmetrical. Access to resources that can support this maturation process — availability of AI models, time/bandwidth, human talent, and institutional or acquired knowledge — is asymmetrical. Every team has a different mix of all of these variables. Some will have very advanced tools at their disposal. Others present results working with relatively limited or modest resources, compensated for by ingenuity. Most of the industry toils away under unreasonable stakeholder expectations, cost/time pressures, misbehaving AI agents, etc.

It’s precisely this asymmetry that makes the process of maturation unique for each company — and a story worth listening to and taking notes from.

Case in point: I haven’t attended the talk by OpenAI’s Kathy Mok at LocWorld Dublin (too many competing sessions on the program!), but I heard echoes of it. For all its role in transforming business reality, OpenAI's globalization program seems to be tackling very familiar concerns that well-established localization programs have been solving for some time now. 

The conferences I’ve been to have done a decent job of capturing the present. That’s where much of our attention should be focused. They also do a shoddy job of predicting the future. We ask questions of “How should we (re)name what we do?” (spoiler: NOT translation or localization) or “How not to go the way of the dinosaur?” I don’t think the industry has a clear view of what the endpoints actually are, so the effort might be a tad misplaced. Marina’s mention of hybrid human-AI coexistence, which is certainly plausible given present-day evidence, might be one of the better options available to us.

Don’t get me wrong, these are important questions to ask. But are we okay with NOT having answers to them and leaving them open-ended for a while longer? You can sit with them a little while longer. Quietly, too, if that’s your preference.

Instead, I’ll argue that rolling up our sleeves and doing the actual work that propels us forward is the best immediate remedy to the uncertainties open-ended questions tend to generate. It certainly helps tune out popular catchphrases such as “AI will replace us” or “We’re automating ourselves out of our jobs.” They don’t count as something worth surrendering to. Nor do they present a future to work toward. And, either way, there will be a multiplicity of endpoints for every company and team out there.

So what’s my takeaway? It may very well be concentrating on the little hack, workaround, or piece of vibe-coded ingenuity that makes a meaningful difference today while staying true to the path that leads you toward your future.

Marina Pantcheva

The first half of 2026 took me to two events that set the tone for our industry: GALA WorldReady and LocWorld55. What I saw there was a natural evolution of the processes that GenAI set in motion.

Human-like quality demands human-like complexity

One after the other, the players in our industry are realizing that to make AI perform like a human and deliver the same quality, it’s not enough to prompt and go. One needs to build workflows that mimic how humans actually work: Pulling in references, terminology, prior context, style awareness, memory, and an understanding of the situation around the text. Only once we provide all of this to the machine can we expect the machine to perform like a professional human translator. This seems to be the path to parity at the moment.

But it comes with a fat bill attached. Every reference, term lookup, and piece of context raises the complexity of the process. And this complexity puts a toll on everything downstream.

Tokenomics takes the stage

When complexity climbs, token efficiency becomes critical. Running complex, multi-step, multi-agent processes quickly spirals into an uncontrollable token-devouring frenzy. Machines doing human-parity work then end up costing as much as humans, and sometimes more.

Enter scale: AI has to be applied to the work where it actually pays off. And it must be engineered to run lean, not used across every task simply because we can.

Everyone is building like mad and converging on the same blueprint

Everyone is building their own tools, in-house platforms, bespoke solutions. The tooling market is fragmenting even further, disrupting itself in real time.

And yet the “innovative” solutions keep landing on the same ideas. I watched three different major technology providers, working independently, arrive at the very same idea and solution design.

Most tech providers gravitate toward prompt-induced AI agents, with a human in the loop – or, more honestly, at the end of the loop – acting as the quality guarantor. A sort of insurance.

Some are replacing even that human

A handful of companies are pushing past that human checkpoint entirely, moving toward LLM-as-judge to account for the output. They are letting the model evaluate quality and, in some cases, stand in for human judgment itself. This is worth watching closely, because where this lands says a lot about how much faith the industry is willing to place in models grading their own homework.

Linguists at the front of the process. As long we have the right ones

Some of the most forward-looking talks pushed the human out of the loop and into the decision seat, designing, teaching, and governing AI workflows instead of cleaning up after them. I think that's exactly the right direction.

The open question is whether we have the people to do it. Engineers still hold the upper hand in developing AI solutions, and few are as open as the OpenAI team, where engineers work closely with the Localization team and listen. But even those who collaborate with the Localization teams still tend to carry a reductionist view of language, predominantly treating it as a mechanical system that should obey clean, Python-like logic. Linguists, who understand how messy language really is, often aren't AI-savvy enough to take the wheel. So, they end up watching this train pass by.

Closing that gap and building a workforce that understands both language and AI models is the most important work the industry isn't doing fast enough.

Johan Botha

GALA WorldReady 2026 in Berlin gave me plenty to think about.

We’re spending less time arguing about whether AI belongs in localization. That’s probably healthy because that discussion is over. AI is here to stay. Some companies are embracing it, others are actively avoiding it, but nobody is ignoring it. 

For an industry that talks endlessly about AI, we still don’t spend enough time talking about clients. What are buyers actually asking for, what are they refusing to pay for, where are they seeing value, and what are they resolving to build themselves without us? Where are they quietly and not so quietly deciding that “good enough” really is good enough? Those questions will shape the future of many LSPs far more than the next model release.

I also kept looking for more discussion around Africa and other lower-resource markets. AI doesn’t arrive everywhere under the same conditions. Connectivity, regulation, language resources, infrastructure, skills, and budgets all influence what is possible. We sometimes talk as though the whole industry is moving at the same speed when it clearly isn’t.

Another thing: hardly anyone seems interested anymore in asking whether AI can or should translate. Which again shows it's here to stay and that the better question is perhaps whether organizations can implement it successfully. For me, technology seems to be the easy part at the moment. Changing workflows, pricing models, accepting quality expectations and revolutionizing company culture is considerably harder. 

One final observation. The localization industry has always been good at talking to itself. I think the next year will require us to become much better at talking to everyone else. Product teams, procurement, compliance, marketing and legal departments all need to become our friends. Not to mention clients who don’t care which model you’re using, but who care deeply about ever-increasing cost, risk, and ultimately, speed.

Stavroula Sokoli

No boots on the ground for me this time around. While my colleagues were collecting air miles and conference lanyards, I was safely home attending online events. The most notable, for me, was our very own AI ThoughtCon, and not just because I chaired the third day. What follows are two insights that I've been repeatedly coming back to over the past three months.

The first came from Ben Hylak, founder of Raindrop, who was our opening speaker on the third day. Ben does not come from localization. He builds AI products in San Francisco, knows the people training the frontier models, and was candid about the distance between his world and ours. He argued that localization quality and multilingual capabilities are not a priority for AI labs. One example, for me, is OpenAI's ChatGPT Translate (chatgpt.com/translate). It's difficult to see it as a new product rather than a different entry point to capabilities that already existed, especially when it still omits languages such as Greek.

His prediction was that this gap would become a crisis when AI is deployed in healthcare in India, banking in Brazil, and government services in Nigeria, with “spectacular” failures, such as misdiagnosis, fraud, and legal liability on the horizon. At that point, he said, the labs "will suddenly care quite a bit, very urgently." He didn’t commit to a further prediction about what happens after that, but I think about it very often. What happens when the major AI labs begin treating multilingual performance as a core benchmark and evaluation criterion?

The second came from Emily Diamandopoulou, owner of Rhyme & Reason and Rima, who discussed tool dependency.  She described how she has spent a decade subscribing to platforms that promised to solve the problem of the previous platform, migrating data and retraining teams, discovering bugs by month four, watching the annual cost of the same features climb from €4,000 to €9,000 to €24,000. Instead of what platform to use next, she started asking whether a platform was the answer at all. And whether, if a tool disappeared overnight, the workflows and data could be reconstructed within a week.

She talked about customization as a trap: the more you customize inside a platform, the harder it is to leave, because every custom automation is debt you pay when you try to exit. In other words, platforms sell customization as a feature, but it becomes a mechanism of capture. And about automation that is anything but: if your automation requires a developer to maintain it, you haven't automated anything. You've just created a different obligation. Her answer was not a new tool, but a discipline: own your data, document your workflows, and treat every platform, including the one she is using today, as replaceable.

Libor Safar

My conference circuit so far this year has taken me from Loc Tech Live behind a screen to GALA WorldReady Berlin as a participant to co-hosting the GALA Academy session on AI for Marketing and, most recently, to LocWorld55 Dublin.

The main pattern I saw was a shift away from talking about what AI can do (and how well or terribly it does it) toward more serious conversations about AI operating models and the business outcomes they should produce.

That is a healthier place to be. Less glamorous, perhaps. More architecture diagrams, more uncomfortable governance conversations. But also more useful.

Loc Tech Live showed clearly how AI localization infrastructure is becoming less about one monolithic platform and more about a modular system that decides what should happen next, which engine or human should handle it, and how the result should be measured.

But my biggest takeaways are actually more about people.

The best systems are being built with, and often by, localization professionals who know what “good” looks like in context. You can swap out tools. You cannot easily replace institutional knowledge about markets, users, terminology, risk, brand, and product behavior.

That is why I think the next phase belongs to teams that run “the loop” rather than the launch: build, measure, diagnose, fix, and repeat. The advantage will come from learning velocity. The teams that win will be the ones that can turn experiments into systems, and systems into better decisions.

LocWorld55 reinforced another version of the same story. There was plenty of AI, of course, but the stronger conversations were about shippability, evaluation, ownership, and organizational change.

“Good enough” is becoming both more tempting and more dangerous. This is also why I keep coming back to the emerging role archetypes.

The work we used to call localization is being re-bundled. We need multilingual AI evaluators who can define test sets and error taxonomies. We need agent managers and orchestrators who can watch performance, risk, and escalation logic. We need platform and workflow architects who can connect the stack. And we need international experience strategists who can connect all of this to acquisition, activation, retention, trust, and revenue.

The future head of localization may not look like the old people manager at all. In some organizations, the highest-leverage person may be a strategic IC or a very small pod: technical enough to build or direct systems, data-fluent enough to measure outcomes, and senior enough to influence product, marketing, legal, support, and leadership.

Not everywhere. Not for every context. But increasingly plausible.

My working hypothesis: the next advantage in localization will not come from adopting AI once, or even well. It will come from learning faster than everyone else.

Let’s see whether H2 proves that right, or proves me totally wrong.

Bridget Hylak

Outside of the localization ecosystem but smack in the heart of the AI bubble, attending Tools for Humanity's Lift Off! event in San Francisco, a private, invitation-only gathering hosted by Sam Altman and the leadership team behind World, was eye-opening in many ways. The event celebrated the latest evolution of the company's ecosystem, including World ID, its flagship "proof of human" product, and offered a glimpse into how its creators envision trust and identity functioning in an AI-first future.

The event asked, then boldly attempted to answer, one question that every industry will soon have to confront: How will we reliably distinguish real humans from AI-generated identities in an increasingly synthetic digital world? As AI agents, avatars, voice cloning and deepfake video continue to improve, proving that a real person is actually participating in a conversation, signing a document or making an important decision may become just as essential as proving someone's identity today. 

World ID's approach is intentionally designed around "proof of human" rather than "proof of identity." Using a one-time verification process and (allegedly) privacy-preserving cryptography, users can demonstrate that they are a unique human (allegedly) without revealing unnecessary personal information, an important distinction as digital trust becomes a global issue.

(I include “allegedly” twice for the skeptics, but I personally know some of the people behind World ID’s efforts, and I couldn’t have more trust in them. I believe that they believe this is “for the good,” and I must disclose that that faith influenced some of my own framing going in.)

The demonstration that struck me most was Zoom's integration of Verified Human. Imagine joining a video meeting and having cryptographic confirmation that the person on screen is actually the individual they claim to be, rather than an AI-generated avatar or deepfake impersonation. In Zoom, this verification is shown via a small, blue “human” badge that appears within your Zoom video tile, letting everyone else on the call know that you went through the process to prove you are, well, you.

With 20 people on a high-level, confidential Zoom call, not having that badge might increasingly become a point of contention.

There seem to be 3 Alex Blanias on this call. Only one - sitting in a hospital bed - carries the blue “human” icon in the upper left corner of his Zoom video tile, revealing the “real, human” Alex. (Those who know Alex personally might also get a clue from his smile, but that’s beside the point.)

Other demonstrations, including integrations with Tinder to verify that dating profiles represent real people, and DocuSign to strengthen confidence in digital transactions and signatures, reinforced just how broadly this concept could extend. That has enormous implications not only for enterprise meetings, contract execution and financial transactions, but also for protecting families, public officials, healthcare providers, interpreters and anyone else whose work depends upon trust. Selling trust is certainly something that localization professionals know a lot about!

For the localization industry, where we increasingly discuss AI agents, multilingual customer interactions and remote interpreting, "proof of human" could become an important new layer of digital infrastructure. As communication becomes more multilingual, more remote and increasingly AI-assisted, verifying who is actually participating in that communication may prove to be every bit as important as understanding what is being said.

My particular family of AI and language technology professionals tends to be understandably skeptical of sharing biometric data, and I certainly count myself among them. Yet the examples presented throughout the event gradually shifted my own internal equation of risk versus harm versus benefit. 

By the end of the evening, yes... I had scanned my eyeballs and become a "Verified Human." Was it persuasive marketing, caving to peer pressure, or perhaps the endless parade of gourmet hors d'oeuvres that finally sealed the deal? Only time will tell. I like to believe I have enough moral fiber to make decisions based on the facts alone, and frankly, several of the "what could go wrong?" scenarios presented during the hour-long program were situations I had never seriously considered before, but can now easily envision becoming part of everyday life. 

Whether World ID ultimately succeeds or not, the broader conversation around proving our humanness in an AI-driven world is one I suspect every one of us will be having sooner rather than later.

If you're interested, you can watch the entire Lift Off! event here (sorry, you'll miss the ritzy snacks and the overly attentive catering staff, but some things just can't be digitized!): 

Hilary Atkisson Normanha

A handful of vendor teams are doing genuinely innovative work, starting from what AI makes possible and rebuilding the entire workflow around it (rather than tethering ourselves to the way it's always worked). The teams I'm most impressed by are redesigning workflows from the source content rather than the output, which is a real shift in how we think about the whole pipeline. But they're the exception. 

On the vendor side, I still see vendors are recycling as much of their existing stack as they can, bolting AI translation or an LLM editor onto the same TMS and write-edit-review flow we've used for decades and calling it innovation. I understand the business logic, but I believe we're capable of something far more ambitious.

At Loc Tech Live, I watched small buyer-side teams get genuinely creative, pairing AI tools with their own expertise to build scrappy solutions (in ways they couldn 't before), without waiting on engineers. However, some of the most impressive work still comes from companies with the budget for complex (proprietary) AI systems. People across the industry are experimenting now in a way they weren't a year ago. There's an openness in the room that wasn't there before, and it's contagious.

Where I still get frustrated is how much oxygen goes to quality as the central question. I build a product that assesses exactly this, so I know it matters, but I think we keep returning to it because it's familiar, and that familiarity is holding us back from bigger ones. Even so, the most encouraging version of this conversation I saw was at LocWorld55 in Dublin, where KnowBe4 presented how they use quick signals from user data to gauge quality. 

However, the conversations I really want to see more of are the ones I only heard a few times this year, about governance and orchestration layers, the systems that will actually define how multilingual AI works. I believe that's where localization can have a true impact in whatever comes next.

More than any single topic, what I want is for the brilliant minds in this industry to imagine the future for themselves instead of adapting to what someone else has already built. We have the expertise. We have the vantage point. We have everything we need to define what comes next.