Learning Methods
Feedback & Correction
AI & Language Tools

THE SHORT ANSWER
Good feedback corrects real errors and leaves correct language alone.
Dioma's corrections come from an AI model that reads your answer against the task, your level, and a curriculum written by educators. No person reviews each one, so we test the system on two measures together: harmful corrections shown and real errors still caught. In an October 2026 test, the harmful share in Hebrew speaking fell from 50% to 15%. French results were mixed.
When we tested a newer AI model on Dioma’s speaking feedback, one result stood out: it made far fewer corrections. Given the same instructions, the earlier model averaged 2.2 corrections per spoken answer and the newer one averaged 0.7, about a third as many.
We immediately looked into why. Fewer corrections might mean the model had stopped finding fault with sentences that were fine, which would be a real improvement. It might also mean the model had stopped noticing mistakes, which would feel better in the middle of a session and teach you less. The numbers alone couldn’t tell us which, so we looked at what the newer model had stopped correcting.
It turned out to be some of each. Some of the corrections that disappeared had rewritten language that was already acceptable, or had blamed the learner for a word the transcription system got wrong. Others had caught genuine mistakes that the newer model missed. We needed a way to remove the unnecessary corrections while keeping the ones that caught real mistakes.
A correction count can’t tell you how good the feedback is
A system that rewrites valid sentences produces a long list and looks rigorous, while a system that says little produces a short list and looks accurate. This is the core tension we spend most of our feedback testing trying to fix.
When Dioma reads something you wrote or said, it can be looking at one of four things, and each requires a different fix.
What the feedback is looking at | What should happen |
|---|---|
A real error. The grammar or word choice is wrong for what you meant. | A correction that says what to change and why. |
An acceptable alternative. Your sentence is correct, though the model would have phrased it differently. | Nothing. Your sentence was good and we should leave it. |
An uncertain transcript. In speaking, a word that may have been misheard. | Caution. You should not be blamed for a word you may not have said. |
A style suggestion. Your sentence is correct and could sound more natural. | A suggestion, presented as optional and not as a mistake. |
We’re constantly tuning our feedback engine to deliver more of “what should happen” on this list.
Educators wrote the curriculum your answer is checked against
A sentence can’t be judged in isolation. Dioma’s feedback works from four things: your answer, the task you were given, your level, and the relevant topics in our curriculum.
The curriculum is the part people wrote. Dioma’s educators wrote the topics your practice and feedback are built on, so the rules were set by teachers ahead of time, and the AI model has the narrower job of applying them to the particular sentence you produced. We describe that division in how Dioma uses AI.
No person reads each correction before you see it. And while our code checks that the topic a correction points to exists in the curriculum, that check confirms the reference and nothing more: it can’t tell us whether the model judged your sentence correctly.
In speaking, the words can change before the grammar is judged
A written answer reaches the feedback step exactly as you typed it. A spoken answer goes through one more step first: the recording is transcribed into text, and the model reads the transcript.
In a controlled recording we made to test this step, a Hebrew word meaning “they passed away” came back from transcription as a similar-sounding word meaning “they were solved.”
The sentence that reached the feedback step was no longer the sentence that had been spoken. If a model corrects the transcript at that point, the learner is told about a mistake they never made. The error can also run the other way, when a transcript tidies up something the learner did get wrong and the mistake goes unmentioned.
So the standard for speaking has two halves. Dioma’s speaking feedback is instructed to hold back where transcription is the likely cause: punctuation, repeated words, words you started and abandoned, and swaps between words that sound alike. A clear grammar or word-choice error in words you plainly said is still treated as yours and still corrected. This lowers one kind of risk without removing it, since a model that sees only text can miss a change the transcription made.
The test measured bad corrections and real errors together
In late October 2026 we took real learner answers that had been held apart from the ones we used while adjusting the system, and ran them through the earlier version of the feedback system and the newer one. For each version we measured what share of the corrections it showed were harmful, and how many real errors it found. A harmful correction is one you would be worse off following, such as a rewrite of language that was already acceptable.
The newer version is a newer model together with revised speaking instructions, so in speaking the test compares the two together and can’t credit either one alone. The writing instructions were the same on both sides. The judging was also done by AI models, which can disagree with each other.
In Hebrew speaking, across 86 responses, half of the corrections the earlier version showed were judged harmful. With the newer version that share was 15%. In Spanish speaking, across 28 responses, it went from 24% to 3%. Those drops would mean very little if the newer version had simply gone quiet, and the second measure is the check on that.
It hadn’t gone quiet. Relative to the earlier version, the newer one found 1.21 times as many real errors in Hebrew speaking (with a 95% interval from 0.96 to 1.58) and 0.97 times as many in Spanish speaking (0.79 to 1.19).
In Hebrew and Spanish speaking, the newer version showed far fewer harmful corrections and found about as many real errors. What the test measures is the quality of feedback on the tested answers, which is a narrower thing than how quickly anyone learns from that feedback.
French was mixed, and we released the change knowing that
French did not follow the same pattern. In French speaking, across 48 responses, the harmful share fell from 45% to 31%, a smaller improvement that still left nearly a third of shown corrections judged harmful. On real errors the judges disagreed, so French speaking has no settled verdict. In French writing the newer version went the wrong way on both measures: the harmful share was 10% against 5%, and it found 0.76 times as many real errors (0.57 to 1.00).
We released the newer version for topic feedback in all three languages anyway, accepting the French results as a known tradeoff and watching French feedback after release.
Open Practice got its own test, and French got a follow-up
Open Practice was tested separately, because results on topic exercises don’t automatically carry over to a different kind of exercise. In a development round on 27 spoken answers, the version we selected showed 1 bad correction where the earlier one showed 15, and 14 real corrections where the earlier one showed 21. On 30 written answers it showed 26 real corrections and 1 bad, against 44 and 2. The tradeoff has the same shape as in topic feedback: far fewer bad corrections in speaking, and fewer real errors flagged in both. That round was run during development, so it carries less weight than a test on answers held apart.
In the same round we also tried a follow-up candidate, one that changes the curriculum material the model is given, on a small set of French answers.
In speaking, the candidate showed far fewer harmful corrections and found more genuine errors. In writing, the two were close. The numbers are small, and they don’t replace the October result, because the answers are a different set. The writing version also still needs a confirming test, since its final instructions differ from the ones we tested.
A correction also determines what you practice next
Your results on each topic are recorded, and those results help decide which topics you practice next. A bad correction therefore does damage twice. It teaches you to doubt a sentence that was fine, and it counts against a topic you handled correctly, which can bring you practice you don’t need. A missed error does the reverse and leaves a weak topic looking stronger than it is.
If a newer version grades the same work differently, your recorded accuracy can move even though nothing about your Spanish, French, or Hebrew changed. For that reason every saved piece of feedback records which model and instructions produced it, so a shift that comes from the grader can be told apart from a shift in you. The companion articles explain how Dioma chooses practice topics and how Dioma suggests level changes from those results.
Three questions to ask about a correction you doubt
No amount of testing makes every correction right, so the last check is yours. When a correction surprises you, I’d ask three questions.
Did it respond to what you meant? In speaking, check whether the word being corrected is one you said. If the feedback is about a word you don’t remember saying, as with “solved” in the recording above, transcription is the likely cause.
Does the explanation fit your sentence? A real error comes with a reason. Read it against the sentence you produced and the task you were given, and see whether the reason applies.
Is it an error or a preference? If your version was correct and the feedback offers another way to say the same thing, you can take it as an option.
If the correction holds up, use the rule again in a new sentence of your own while the reason is still fresh. If it doesn’t hold up, report it to us. A correction that misread your meaning is the case our tests exist to catch, and you are the only person who knows what you meant to say.
For the rest of how Dioma works, see our method.
See how Dioma builds your practice
Speaking and writing practice guided by your results and an educator-written curriculum.
Explore the method
Frequently asked questions
Why would Dioma correct something I said correctly?
Two things can cause it. In speaking, your answer is transcribed before it is graded, and a misheard word can reach the feedback step as if you had said it. In speaking or writing, the model can also treat an acceptable alternative as an error. We test for both.
Does a person review every Dioma correction?
No. Educators wrote the curriculum the feedback is built on, and an AI model applies it to your answer. No person reads each correction before you see it, which is why we test the feedback system on real learner answers held apart from the ones used to adjust it.
How does Dioma tell a mistake from another acceptable way to say something?
The standard is that a real error gets corrected, an acceptable alternative is left alone, and a more natural phrasing is offered as an option. The model applying that standard works from your answer, the task, your level, and the educator-written curriculum. It will sometimes get this wrong, so we measure how often.
Does getting fewer corrections from Dioma mean I am improving?
Not on its own. The number of corrections also depends on the system doing the correcting.
Related Articles










