How accurate is our AI grading?
Tickmark's AI estimates IELTS Writing bands. Here is how it grades, how far it is from trained raters on essays they have scored, and where it can be wrong. We publish this because a band you cannot check is worth little.
Measured against human raters
Not measured yet
We have not measured the grader against bands given by trained IELTS raters, so we show no accuracy figures rather than guess. Until we do, bands are not corrected for bias, and the figures below only show how consistent the grader is, not whether it is right.
Consistency: the same essay graded again
We grade the same essays several times and compare the bands. This needs no rater, so it shows how steady the grader is, not how close it is to an examiner.
- Essays, each graded 3 times
- 10
- Same band every time
- 90%
- Within half a band
- 100%
- Average gap between highest and lowest band
- 0.05 band
Measured on October 30, 2026 with claude-opus-5-5, prompt writing-v3, on Tickmark's own model answers (bands 7 to 8).
How the AI grades
- The four public criteriaEach response is read against descriptions based on the public IELTS Writing band descriptors: Task Achievement (Task 1) or Task Response (Task 2), Coherence and Cohesion, Lexical Resource, Grammatical Range and Accuracy.
- Evidence for every bandEach criterion's band comes with quotes from your response and what they show, so you can see why it is that band and disagree with it if the quotes do not support it.
- A range when it is a borderlineWhen a response sits between two bands, the AI says so and shows a range (for example 6.0–6.5) instead of pretending to be sure. A single band means the evidence clearly fits it.
- IELTS roundingThe task band is the average of the four criteria, rounded down to the half band, as IELTS Writing scores are.
- Correcting a measured biasAn AI grader can be consistently harsher or more lenient than examiners. Once we have measured our own bias on enough rated essays, each criterion is corrected by it, and the result page says so.
- Versioned and loggedEvery result records the model and instructions it was graded with (now claude-opus-5-5, writing-v3). Any change to them is measured again before it ships.
Limits
- Near the boundary between two bands, the AI and two human examiners can all disagree by half a band.
- Grading the same essay twice can differ slightly; the range is there for that.
- It is practice feedback, not an official score: only the IELTS test gives an official band.
- Very short, off-topic or memorised responses are hard to band reliably; the result page warns about them.