How Eurmes uses AI

Teaching material: lessons, exercises, the audio you hear (synthetic voice) and some teaching images. The AI proposes; a teacher decides.

And what we do not generate: decisions about you. No AI system at Eurmes decides whether you get in, whether you pass, what you are worth or what you are offered. There is no profiling and no scoring of you as a person. AI is used to teach you, not to judge you.

No AI-generated content reaches a learner without a teacher having reviewed and published it. This is not an aspiration: it is how our editor works. The AI writes drafts inside the teacher's tool, and publishing is always a human, explicit act.

There is one exception, and we are telling you about it: personalised exercises are generated for you on the spot, so no prior review is possible there. That is why such content is marked as AI-generated.

We name them, because saying «we use AI» without saying whose is not transparency:

• Text and teaching images: Google (Gemini), OpenAI, Anthropic, Mistral, Alibaba (Qwen) and Moonshot (Kimi).
• Synthetic voice (text to speech): ElevenLabs, Fish Audio and Microsoft.
• Pronunciation assessment: Microsoft Azure Speech, in Ireland, within the European Economic Area.

Any task whose text comes from a learner can only be handled by models hosted in the European Union, or by providers whose transfer certification we have checked one by one. Today that rules out the Chinese models, and also OpenAI and Anthropic.

Those excluded models only generate generic teaching material; nothing you have written passes through them.

The prompts used to generate content carry no identifier of yours: not your name, not your email, not your user id. To whatever generates a lesson, you do not exist.

The only time something of yours travels is when you write it: when you personalise an exercise, your text goes to the model as it is — without your identity, and only to the providers allowed by the rule above. And we do not train any model on your data. None, ever.

When you practise pronunciation, your voice is sent to Azure through our servers to score how that sentence came out. That is all.

• It scores the exercise, not the person. The result tells you how your attempt sounded; it does not classify you or feed any decision about you.
• It assesses how you sound, not who is speaking. There is no speaker recognition.
• Your voice is not stored. It is processed to give you the result and then it is gone; we do not keep a transcript either.
• We will never use it to identify you. No voice login, no voiceprint. If we ever wanted to change this — it is in no plan — it would first require your explicit consent and a prior impact assessment. Paperwork and your permission first; never the other way round.

If you hear a voice that is not human or read a text a machine wrote, you have a right to know. Generated audio is identified as synthetic voice in the player itself, and AI-generated lessons carry their marker alongside the teacher's review. Our rule is to label everything, whether the law demands it or not.

A score that strikes you as unfair? Content that is wrong or reads oddly? Write to hola@eurmes.eu . A person reviews it, not a system.