We’ve been asked frequently how Dioma uses AI, and we aim to be transparent. There are a lot of learning tools popping up that just wrap an AI model in a sleek interface and call it learning.

Because of this, we understand why you would be cautious.

Yes, Dioma uses generative AI, the same family of technology behind tools like ChatGPT and Claude. The power of generative AI is how we’re able to deliver on-demand feedback at $13 a month.

We are incredibly intentional about where we use AI, and where we don’t.

The two jobs the AI has in a session

Generative AI does two things in Dioma, both of them in the delivery layer of your session.

It corrects what you produce. When you write a paragraph or speak your answer aloud, a language model reads your response against the relevant curriculum topics and returns the corrections and explanations you see about fifteen seconds later.

It drafts your weekly plan. A model helps choose which curriculum topics your coming week should focus on, working strictly within the level the system has already assigned you by other means.

That is the complete list. Speech recognition, the step that turns your voice into text, uses a dedicated transcription service rather than a generative model. Everything else you might wonder about, from your level to your progress score, runs on plain, traditional code.

The decisions the AI is never given

The judgment calls that shape your learning are deliberately kept out of the model’s hands.

We did not divide it this way out of caution alone. We tested every part of our system between AI solutions and traditional code solutions, and continue to test each week.

Our logic is that where a decision needs to be consistent, inspectable, and fair, it lives in rules we can read and test. Where the work requires flexibility and is generating language about your language, the model does it.

The curriculum sets the boundaries

The corrections themselves are not a model improvising from its general training. Every request that goes to the AI connects to relevant Dioma curriculum: the topic explanations, worked examples, grading notes, and common errors that our educators wrote for your level.

Think of a new teacher joining a school. They bring their own fluency and quick judgment, but they teach from the department’s syllabus and mark against the department’s marking guide. That is the model’s position at Dioma. The educators decided what a learner at each level needs, what counts as an error there, and which mistakes matter most. The model applies those notes to your sentences in real time.

This is what we mean when we say the feedback is checked against the expert curriculum, and it is why a Dioma correction differs from pasting your homework into a general chatbot. A chatbot answers from everything it has ever read, with no record of what you have covered or what your level demands. Dioma’s model answers from the same set of educator-written notes every time, and its answer has to stay attached to them. Who wrote those notes, and what it took to build them, is its own story: how we designed Dioma’s curriculum.

What happens to the words you write and say

The request that reaches the model carries your response, the exercise prompt, the languages involved, and that slice of curriculum. It does not carry your name, your email, or your account ID. The models run inside Dioma’s own cloud environment rather than through a consumer chatbot service, and the providers we run on do not keep these requests or use them to train their models.

Your practice history, the record Dioma itself keeps to track your progress, stays in Dioma’s database, where it belongs to your account. When we evaluate a new model or a change to our feedback system, we benchmark it internally against real practice submissions before anything changes for learners, because testing on synthetic sentences would tell us little about feedback quality on real ones.

How the feedback gets checked

We would rather tell you plainly what the quality system is than imply a person reads every correction. No human reviews each individual correction before you see it. What stands behind the feedback instead:

The curriculum layer is fully human. Educators wrote and reviewed every topic, example, and grading note the model works from, and that work is the deepest quality control in the product, because it constrains what the model can say.

Every correction is traceable. Each piece of feedback is stored with the exact model and prompt version that produced it, so when something looks off, we can find every learner who got feedback from that configuration and fix it at the source.

Changes are benchmarked before they ship. A new model or prompt runs against a set of real learner submissions, repeated multiple times, and gets compared against the incumbent before it ever reaches a session. Exercises go through their own testing, graded on a seven-part rubric with borderline cases queued for human review.

And when a correction still looks wrong to you, you can flag it. No system built on generative models is perfect. What we can do is keep every problem visible and traceable when it happens.

Why we use AI at all

Given all these limits, the fair question is why bring generative AI into the product at all. The answer is the one thing it makes possible that nothing else does at this price: feedback on your own production, in seconds, at any hour.

Producing the language, writing and speaking from scratch and getting corrected, is the practice that moves intermediate and advanced learners forward, and it is the piece self-study could never provide. Before this technology, the only way to get your own sentences corrected was another person: a tutor at $30 to $60 an hour, or a class moving at the group’s pace. We love classes and tutors, but intermediate and advanced learners need more practice than they provide.

Dioma’s whole design, an expert curriculum with a model delivering it, exists to put that correction loop inside a $13-a-month product. What you are paying for is the corrected sentence, fifteen seconds after you wrote it, anchored to a rule an educator chose for your level. The AI is only how it reaches you that fast.

If you want to see what that loop looks like in practice, the method page walks through a real session start to finish. And if you are deciding how AI should fit your own study more broadly, including the ways it can quietly do your practicing for you, we wrote a companion piece on how serious learners should use AI.

Frequently asked questions