Contents
It was my last night in Madrid, and I wanted to get one small thing right before I left. The woman who ran the pensión near Plaza Mayor had gone out of her way all week: extra towels, a recommendation for the one restaurant near El Retiro that did not have a line out the door. I wanted to thank her properly in Spanish, not just smile and wave on my way to the airport.
I opened a translation app, typed out something like “thank you so much for everything, you made our trip,” and got a phrase back. Then, more out of habit than doubt, I opened a second app to check it. It gave me something noticeably different: more formal, almost stiff, the kind of thing you would say to a stranger rather than someone who had just spent a week looking out for you. A third app landed somewhere in between. Three apps, three answers, and no way for me to know which one actually sounded warm and which one sounded like a tourist reciting a phrasebook.
I went with my gut in the end, standing in that doorway scrolling between three screens while she waited, far too patiently, for me to just say something. She smiled either way, which is its own kind of relief. But the moment stuck with me on the flight home. If three free apps could not agree on how to say something as simple as thank you, what happens when the stakes are higher: an allergy you need to disclose, a symptom you need to describe to a pharmacist, an apology that needs to land exactly right?
Why one wrong guess is worse than it sounds
Most of us default to whichever translation app is already installed, and for good reason. It handles directions, menu items and basic small talk without complaint. The trouble is that travel has a habit of testing language in exactly the moments where a wrong guess costs the most: describing a food allergy to a server, explaining symptoms at a pharmacy counter, filling out a customs form, or thanking someone who has genuinely taken care of you, which is exactly the situation I found myself stuck in that night.
The risk is rarely that the app gets something outright wrong. It is that you have no way of knowing when it has, unless you already speak the language well enough not to need the app in the first place. Standing in that doorway, I had no real way to judge which of my three thank-yous was the right one. I just had to pick.
Turns out I am not the only one juggling apps
Travel guides have quietly worked around this for years by recommending different apps for different jobs. One recent roundup of the best travel translation apps points travelers toward Google Translate for general use and DeepL for more precise written text, while flagging region-specific tools for other language pairs entirely. Looking back, I realized I had been doing some version of this for years without ever really thinking about it: one app for the metro, another for anything I actually planned to write down.
But swapping apps by scenario only patches over the deeper issue. Even the same app, asked the same question twice, can waver, which is more or less what happened to me three times in a row in that doorway.
So why don’t they ever agree?
That question followed me home, so I finally looked into it properly. Every translation app is built on top of an AI language model, and different models trained on different data will render the same sentence differently. That is not a bug in any one app so much as a fact about language itself: plenty of everyday phrases, thank you included, have more than one correct answer depending on tone, formality and context, and no single model has a monopoly on picking the right one.
An independent industry evaluation of translation engines and large language models, covering tone of voice, terminology and full-text consistency across dozens of systems, found meaningful gaps between how individual models handled the same requirements. In practice, that is the exact experience of standing in a Madrid doorway comparing three different thank-yous on three different screens.
A separate breakdown of translation apps for international travel puts it plainly: most travelers end up juggling more than one app depending on the situation, and the constant switching creates its own kind of friction. You are not imagining it if this sounds familiar. The workaround has just become normal.
What I found out about consensus
I went down a bit of a rabbit hole after that trip and came across MachineTranslation.com, an AI translation platform developed by the translation company Tomedes, built around a different idea: instead of betting on one model’s guess, why not ask several models the same question and see where they actually land together?
Its SMART mechanism runs a given piece of text through 22 different AI models at once, evaluates the surrounding context, and returns the translation that the majority actually agree on, rather than the output of whichever single engine happens to be installed on your phone. Internal benchmarking on this approach shows up to a 90% reduction in critical translation errors compared with relying on a single model, and an aggregated quality score of 98.5 out of 100 against GPT-4o’s 94.2.
The hallucination numbers tell a similar story: under 2% for consensus-based output, compared with 10 to 18% for individual top-tier models used alone, which would have explained a lot about that doorway in Madrid.
“MachineTranslation.com is no longer just a scoring and benchmarking layer for AI outputs; it now builds a single, trustworthy translation from those outputs, end to end,” says Ofer Tirosh, CEO of Tomedes,
MachineTranslation.com’s parent company. For anything with real stakes, a document, a medical form, a formal letter, the platform also layers in Human Verification, so a professional reviewer can check the AI output before it goes anywhere important. And with the mechanism applied across more than 330 languages, it is not a fix built for one or two major destinations. It would have worked just as well that night in Madrid as it would in Mexico City or Bogotá.
Back to that doorway
A simple thank you is exactly the kind of phrase that trips up single models: short, common, and full of small register decisions a traveler has no way of judging alone. Lining up the best translation of thank you in Spanish across different models shows this disagreement laid out directly, multiple phrasings side by side instead of a single confident-sounding guess. It is the same scene from that pensión doorway, just made visible instead of silent.
What I would do differently next time
I am not about to delete the translation app already on my phone. I am just more deliberate now about which moments deserve a second opinion.
- Keep the everyday app for directions, menus and small talk. It is fine for low-stakes situations.
- For anything with real consequences, an allergy disclosure, a medical description, a formal apology or thank you, look for a translation that shows you where models agree rather than a single unverified answer.
- If a document or form genuinely matters (a visa application, a medical history, a signed agreement), have a human check it before you rely on it.
It is part of a broader shift I only started noticing after that trip. How AI is already reshaping the way people plan trips goes well beyond translation, from itinerary building to real-time customer support, and language is simply the piece of that shift travelers run into most often, usually standing in a doorway, hotel lobby or pharmacy queue. If you are building out your own pre-trip checklist, the rest of the Traveler Tips archive is a good place to keep digging.
I still think about that pensión doorway sometimes, three screens open, her waiting patiently for me to just say something. Next time I am somewhere I do not speak the language, I am not going to ask which app to open. I am going to ask whether what is on my screen is one model’s best guess, or the version several of them actually agreed on. It is a small question. It would have saved me a very silly minute in Madrid.