How accurate is our IELTS writing checker?

Results

Each response was checked once by the same checker learners use, and its overall band estimate was compared with the band the examiner gave.

The fresh set is the fairest test, because none of those responses influenced the scoring. Five of the six estimates were within one band of the official score.

Where the estimate is accurate, and where it is not

• Bands 3 to 5.5: every response officially marked in this range was estimated within one band.
• Bands 8 and 8.5: four of the five responses in this range were estimated between 6.5 and 7.5. The checker is conservative with very strong writing.
• Similar-looking responses: the AI sometimes gives near-identical scores to two responses that an examiner placed a band or more apart. Our calibration cannot separate those; it needs better judgement from the model itself.

A difference of one band matters in IELTS: it is the gap between 6 and 7. Read the estimate as a range, and pay more attention to the evidence and the next steps in your report than to the number. For one criterion in detail, see how accurate AI band scores are for Coherence and cohesion.

How we tested

1. Responses with an official band. We used sample responses published by IELTS, such as the Academic and General Training writing samples on ielts.org, and by IDP. They cover Academic and General Training, Task 1 and Task 2, with official bands from 3 to 8.5. We typed each one from the published scan, so a small reading error is possible, and we do not republish them.
2. Targets set before testing. An average difference of 0.75 of a band or less, at least 80% within one band, at least 90% usable reports and no unexplained difference above 1.5 bands. Every target was met.
3. A calibration step. Before calibration the AI rarely gave any criterion above 6.5. We added a fixed adjustment that spreads scores above band 5. It was tuned on the development and reserved responses, so those figures are partly in-sample.
4. A fresh check. Six newly typed responses were opened once, after the adjustment was fixed, and the adjustment was not changed afterwards.

The checker scores each response against the four criteria in the public IELTS Writing band descriptors. Our plain-English guide to the descriptors explains what each band asks for.

What the results mean for you

• Track direction, not a test result. Compare estimates across several responses. One estimate is not a prediction of your test-day band.
• Around band 5 or 6: the estimate has usually been within one band, so a change of a band or more over several reports is a meaningful signal.
• Around band 7 or above: the estimate may be lower than an examiner's. Use the detailed feedback on each criterion to decide what to revise.
• Close to a target score: if a university or visa route depends on half a band, consider a human examiner as well. Our guide to AI checking or human correction compares the options.

Limits of this test

Twenty-seven responses is a small sample, and published samples may not look like your writing. Most of the responses helped tune the calibration, which is why we report the fresh set separately. Every usable response with an official band that we found has now been used, so the next check needs newly published ones. We will update this page when we run it.

Questions about accuracy