EX:CHANGE 2026

Following a number of discussions across social media concerning LanguageCheck.ai, a partner sponsor of EX:CHANGE 2026, we invited the company to respond to the points raised by members and the wider language services community. We publish their statement below.

Providing opportunities to learn about the new tools that are being developed for language professionals is one of the objectives behind creating an exhibitor zone at EX:CHANGE 2026.

A statement from LanguageCheck.ai

We're grateful to ITI for the opportunity to respond directly to the members who have raised these questions. The concerns voiced over the past weeks deserve a serious answer rather than a marketing one. We'll try to give that here.

Who we are. LanguageCheck was developed by AQRATE, a team that has worked in the language industry for thirty years, in consulting, commercial language services, and software for translation workflows. We are not an AI-only company that arrived yesterday. The tool is engineered by software developers, but a large part of our team holds degrees in Translation and Interpreting and works as language analysts. We come from this profession, we depend on it, and we share the unease many feel about how MT and AI are reshaping it. That shared concern is precisely why we think it's worth being exact about what LanguageCheck is, and what it is not.

What LanguageCheck actually does. LanguageCheck is a review and quality-assurance layer for MT post-editing and human review. It does not generate translations. It does not rewrite or "correct" them autonomously, deliberately, because a system that silently rewrote a translator's work would frequently be wrong. What it does is flag segments where it detects potential issues, explain why, and leave every decision to the professional reviewer. Human judgment stays at the centre of the workflow. That is a design choice, not a disclaimer.

On the word "flawless." This term has understandably drawn scrutiny, and we want to be precise about it. When LanguageCheck labels a segment "flawless," it means the system did not detect an issue worth flagging under the current configuration. It does not mean the segment is guaranteed correct, and it does not instruct anyone to skip review. How a team uses that signal is a professional choice: some review everything but move faster because issues are surfaced first; some triage by risk; some apply sampling or domain rules. The tool informs where attention is directed; it does not replace the reviewer's responsibility for the text.

On our claims, and where we can do better. We have heard the specific objection to phrasing suggesting the system "understands meaning." We accept that some of our marketing language has compressed a more careful technical reality into words that can read as overclaiming, and we are tightening that language. What we can describe accurately is how the system works: text is not simply passed to a general-purpose LLM. It runs through preprocessing, a structured comparison of source and target, and post-processing and categorisation, hundreds of thousands of lines of software built for a constrained, comparative task rather than an open-ended generative one.

That distinction matters to a fair question several people have asked: why should an AI checker be more trustworthy than an AI translator? Because the two tasks differ in kind. Generating a translation is open-ended, word choice, register, ambiguity resolution, and a model under that pressure tends to optimise for fluency, sometimes at the expense of meaning. Evaluating a finished translation is a bounded comparison against an existing source. LanguageCheck is built to exploit that difference, not to pretend the two problems are the same.

On evidence. We built the tool over three years of R&D [research and design] using a repeatable evaluation framework drawn from ten years of real bilingual data, roughly 150 million translated segments, and a billion words across many language pairs and domains, aligned to MQM and ISO 5060 principles, with human-in-the-loop validation by language analysts and native reviewers throughout testing. In well-supported pairs and domains, our internal framework shows very high reliability on segments labelled "flawless." We also state openly that performance drops in harder contexts, decontextualised micro-segments, and ambiguous fragments, and we do not hide that.

On accountability. This is the concern we take most seriously, because translators are right to be sensitive to it. We will not pretend that any software can assume legal liability for linguistic outcomes; no MT engine, no CAT tool, no QA checker, no spell-checker does, because none can stand behind a human professional's final judgement. But the answer to that limitation is not to remove the human; it is to keep the human informed and in control. LanguageCheck is, deliberately, one of the few tools at the moment that insists the translator must remain central. The competence and the final word stay with the professional. And in high-stakes content, medical, legal, regulatory, that principle matters most, which is exactly why we position the tool as support for rigorous human review and never as a substitute for it.

On the wider pressure. We understand why frustration runs high. But we'd gently suggest that the real pressure on rates and timelines is not caused by any single QA tool; it is a market narrative, widely amplified, that "AI will soon replace humans" and that raw MT output is good enough. We do not believe that is true. MT is probabilistic by nature and will keep producing errors, including serious ones, even as the systems improve. A workflow that surfaces those errors for human attention is, we would argue, on the translator's side of that argument rather than against it.

We do not expect this statement to resolve every disagreement, nor do we think it should. We would genuinely welcome continued dialogue with ITI members, including those most critical of us, and we are happy to provide full access so professionals can test the tool on their own real materials and tell us where it helps, where it doesn't, and where our messaging needs to be clearer. That kind of feedback makes the tool better and keeps us honest.

With respect and thanks to ITI and its members,

Anthony Neal Macri 
Chief Marketing Officer, LanguageCheck.ai