7 Writing7 Writing estimates a whole essay in four steps: a model first extracts observable evidence, a separate step proofreads surface errors literally, code then converts the evidence into four criterion scores, and finally a model writes learning feedback around the scores that are already fixed. This article goes through each step in order, where it can go wrong, and how you should use the score.
- Extract evidenceA model reads the question and essay and records what it can observe about task response, structure, vocabulary and grammar
- Literal proofreadingA separate step lists surface errors only, and code checks that each one really appears in the essay
- Code calibrationRules convert the evidence into four scores, then verified error density limits vocabulary and grammar
- Write feedbackA model writes the problems, a model revision and the next step around the fixed scores
Step one: extract evidence instead of scoring directly
If you ask a model for one overall score directly, the result is easily swayed by essay length, tone and template sentences. So the first step asks the model only to record phenomena it can find in the text: whether each part of the question is answered, whether the position stays consistent, whether paragraphs move the argument forward, whether collocations are natural, and whether the grammar varies. The output of this step is a set of evidence labels, not a score.
Step two: literal proofreading of surface errors
While reading, a model tends to "understand" a misspelled word on the way, so surface errors get undercounted, and the weaker the essay, the bigger the bias. That is why there is a separate proofreading step. Its only job is to list errors of spelling, word form, word choice, articles and subject-verb agreement exactly as written. It does not judge the essay, and it does not guess what the writer meant.
The errors the model lists then pass a code check: each quoted fragment must really appear in the essay, and it must differ from the suggested correction. Errors that were invented are dropped. Only errors that pass the check go into the later calculation and the feedback.
Step three: code turns evidence into scores
Task Response, Coherence and Cohesion, Lexical Resource and Grammatical Range and Accuracy are each calibrated by rules from the evidence labels. The overall score is the mean of the four, rounded to the nearest half band. Two limits follow:
- Verified error density limits the vocabulary and grammar scores, so an essay with many errors does not get a high score just because its vocabulary looks rich.
- No criterion may be more than 1.5 bands above the lowest one, which prevents the pattern where only one of the four is unusually high.
The score range shown on screen is half a band either side of the overall score. Official marking averages the four criteria, and the estimate keeps the same structure, but it remains an estimate, not an official result.
Step four: feedback cannot change the score
Only after the scores are fixed does another model call write the feedback: it points out the main problems, quotes sentences from your essay, suggests the next practice step, and tries to rewrite one body paragraph. This step receives the scores that were already calculated and cannot decide them again. The aim is to keep the score and the advice separate: the advice can be written fluently, but the score does not move because the wording is attractive.
How the model revision relates to the score
A model revision is a reference for another way of saying something. It is not the score your essay "could have had". A paragraph the model writes can read smoothly without proving that your original reached the same level. When you use a revision, first check that the meaning and any chart information were not changed, then close it and rewrite the paragraph in your own words.
What the current checks can and cannot show
Our test set holds only eight essays with official scores: five Task 1 and three Task 2. There is no Task 2 at Band 6 to 7 and nothing above Band 8. The same set was used to adjust the rules and to check the result, so the average gap of about 0.4 band on these eight is not an independent accuracy test, and it cannot support a claim that the estimate is accurate for all learners. We also tried 30 outside essays that carry scores from a website; the correlation between our estimate and those labels was only about 0.15, so discrimination is limited, and weaker essays were still estimated above Band 6 on average. Those outside labels are noisy themselves, but the result shows that over-scoring weak essays is a real risk.
Even with identical input, a model's results can differ by half a band between runs. That is why the page states: the AI estimate has an error of 0.5 to 1 band, and essays with many spelling and grammar errors may be overestimated.
How to use the score
A rise in one score may come from a better essay, or from model variation. Three checks are more useful:
- Can you find the sentence in your essay that the feedback is talking about?
- After you revise, is the problem solved?
- Does the same kind of problem come back in your next piece of writing?
Scores are good for watching the trend across several practices. They are not a basis for deciding whether to sit the exam.
To see what estimate your own essay would get, paste it into the IELTS Writing AI checker: one free check, and 3 more free checks after you sign up. AI scores are for reference only.
Connecting feedback to practice
First read what to do after AI feedback and pick one problem you can check. If you only need to look at paragraphs and word count, use the structure checker, which does not call AI. The official scoring requirements are in IELTS scoring in detail.
For common problems in each criterion and the difference between Band 6 and 7, see How to move from Band 6 to 6.5 in IELTS Writing.