AI Writing Feedback for DELF: Scoring Against the Real Criteria

Writing is the hardest skill to improve alone, because your own errors are invisible to you. You cannot spot the mistake you did not know was a mistake. AI feedback closes that gap, but only if what it checks matches what DELF actually marks, and most tools check the wrong thing.

Why writing stalls without feedback

Reading and listening give you an external signal: you either understood or you did not. Writing gives you nothing. You produce a text that feels reasonable, and without correction you have no way of knowing which parts a marker would have taken points off for.

Traditional correction solves this and introduces a different problem, which is throughput. Waiting days for one essay to come back means practising slowly and reconstructing your own reasoning after the fact, by which point the correction lands on a text you barely remember writing.

What DELF actually marks

This is where most preparation goes wrong. DELF B2 writing is scored out of 25 across four criteria, and grammar is one of them.

Roughly half the marks go to content and organisation, the other half to language. That split explains a result candidates find baffling: accurate French scoring poorly. A text with almost no errors that answers two thirds of the prompt, asserts a position without defending it, and arrives as three undifferentiated paragraphs is losing marks under three criteria while doing well under one.

A general-purpose grammar checker only sees the fourth criterion. It has no view on whether you addressed the prompt, whether your argument was developed, or whether your paragraphs connect, which is why a clean report from one tells you much less than it appears to.

What useful feedback covers

Grammar detection is the baseline: conjugation, agreement, tense logic, accents. Necessary, and the easiest part.

Above that, vocabulary. Whether a word is wrong matters less here than whether it is the precise one, and whether you are reaching for B2-level range or circling the same basic set. Register belongs here too, since a formal letter written in casual French loses marks even when every sentence is correct.

Then structure and coherence: whether an argument develops across paragraphs, whether the connectors do real work rather than decorating, and whether the reader can follow the shape of your position.

And task achievement, which is the one that is easiest to check and most often skipped. Did you produce the format asked for, cover every element of the prompt, and land inside the word count.

SavoirX runs DELF-aligned writing tasks at A2, B1 and B2 across more than thirty topic categories, with mock exams in the real format, and returns a score with corrections on every submission rather than a list of flagged words.

How to use it so it works

Write first, check second. Composing with a checker open trains a skill you cannot use on exam day. Write under real conditions, timed, with nothing open, then analyse.

Sort your errors by type, not by essay. One misplaced pronoun is noise. The same misplacement across four essays is a grammar topic you have not learned, and an hour on the topic beats four more corrections of the instance.

Rewrite rather than patch. Applying a correction teaches you almost nothing. Rewriting the paragraph with the correction integrated forces you to produce the right form, which is what actually transfers.

Compare against model answers. Corrections tell you what was wrong; a model tells you what good looks like at your level, which is different information and often the more useful of the two.

Track the rate, not the count. The number that matters is errors per hundred words on the same task type over time. If it is flat after a month, the method needs changing rather than more volume.

Which mistakes waste the practice

Leaning on assistance while writing, which produces essays that do not represent what you can do unaided.

Reading corrections without acting on them, so the same error returns essay after essay. If a piece of feedback has appeared three times, it is not feedback any more, it is a topic.

Practising untimed. DELF writing is as much a pacing problem as a language one, and an essay that would have scored well unfinished scores badly.

And optimising only for grammar, which is the trap the criteria table exists to prevent. Candidates who fix every agreement error and never practise structuring an argument are polishing one quarter of the mark.

What to realistically expect

AI feedback shortens the loop and removes the volume ceiling. It does not write for you, and it cannot tell you what you meant when your sentence did not say it. Nuance, style and whether a French reader would find your phrasing natural still benefit from a human eye.

What it does reliably is make your recurring errors visible and give you enough attempts to fix them before someone marks you for real. For most candidates that is the difference between arriving at the exam having written five essays and having written thirty.

Frequently Asked Questions

Can a grammar checker like Grammarly prepare me for DELF?

Not on its own. It catches errors, which is one of four criteria. It has nothing to say about whether you did what the prompt asked, whether your argument is defended, or whether the text is organised, and those carry roughly half the marks.

How is DELF B2 writing scored?

Out of 25, across four criteria: respect de la consigne, capacité à argumenter, cohérence et cohésion, and langue. Roughly half the points go to content and organisation, the other half to language.

How many practice essays do I need before the DELF?

Enough that your recurring errors stop recurring, which is usually fifteen to thirty with feedback read properly. The count matters far less than whether you are correcting patterns or collecting corrections.

Should I use AI feedback while writing or after?

After. Writing with a checker open turns the exam skill into a different task, since on the day you have no assistance. Write it under exam conditions, then analyse.

Why is my grammar good but my DELF writing score low?

Because accuracy is one criterion of four. Perfectly correct French that ignores part of the prompt, argues nothing, or arrives as undifferentiated paragraphs loses marks in three places while scoring well in one.