Why the IELTS band descriptors aren't enough for accurate AI marking
The trick to getting accurate AI-marked IELTS feedback is understanding that the exam criteria are not enough. There is a missing layer, and that layer is the examiner's training and experience in interpreting those criteria.
The public band descriptors tell you what each band looks like. They don't tell you how a trained examiner applies those words to a real script written in 60 minutes under exam pressure. Examiners learn that from training, standardisation against marked scripts, and second marking. Teachers pick it up from years of comparing students' work with officially-marked scripts and examiner comments. An AI that reads only the descriptors, whether it's a chatbot you paste them into or a purpose-built marker, misses all of it. So does a human who has only read them.
What the descriptors leave out
1. Mistakes at the higher bands
Examiners know that people aren't machines. A slip in an exam essay usually shows a lapse in concentration, not a gap in knowledge. Only edited writing is error-free, and an exam doesn't let you edit. So trained examiners allow for occasional errors at the higher bands, roughly 7 to 7.5 and above. At Band 5 there is no such allowance, because there errors really do show what the candidate can't yet do.
No Writing descriptor states this allowance anywhere. You only see it in how officially-marked scripts are scored. A marker that reads the descriptors literally sees an error at Band 8 and marks the candidate down.
2. "Attempted" is not "clear"
For Writing Task 1 Task Achievement, Band 6 asks that
a relevant overview is attempted
Band 7 is where the higher bar sits:
presents a clear overview … main trends or differences are identified
Students are taught to put the overview near the start. That's good advice on how to write one, but it isn't a marking requirement. A marker that hasn't learned the difference between the teaching advice and the descriptor penalises a perfectly valid overview for being in the "wrong" place.
3. Hesitation in Speaking
Filled pauses like erm, ah and well are a natural part of speaking about a topic without a script. Band 8 Fluency and Coherence separates hesitation used to find words or grammar from hesitation that is content-related, so the descriptor hints at it. But knowing which is which, and how much weight to give it, comes from listening to hundreds of candidates.
What happens without the missing layer: our own measurements
While building the English Me Virtual Examiner, we marked IELTS's own officially-marked Writing Task 2 sample scripts by applying the descriptors faithfully, one criterion at a time. Each script was marked three times:
| Official band | AI band (descriptors only) | Difference |
|---|---|---|
| 4.0 | 3.83 | −0.17 |
| 5.5 | 5.00 | −0.50 |
| 6.5 | 6.00 | −0.50 |
| 7.5 | 7.00 | −0.50 |
| 8.5 | 7.00 | −1.50 |
Across every criterion mark, the AI never awarded an 8 or a 9. Its ceiling was 7. It was close at the lower bands and fell further behind as the official band rose, which is exactly where examiner interpretation matters most. We saw the same pattern in Speaking.
An honest caveat: that is 5 scripts, and only 2 of them are above band 7. It's a small sample, and we report it as one.
How English Me adds the missing layer
- Examiner interpretation is written into the marker. Rules like the higher-band error allowance and "attempted versus clear" come from over 10 years of teaching IELTS and studying officially-marked scripts, not from guesswork.
- It's checked against officially-marked material: IELTS's 2023 Academic Writing sample scripts with examiner comments, and certified Speaking clips.
- Each criterion is marked separately, each with its own evidence, the way an examiner works.
- We report a range at the top of the scale. Above band 7 a single number would claim more precision than any marker has, so you get a range and the reason for it.
Measured result: 80% of Writing marks and 83% of Speaking marks land within one band of the official score (5 Writing scripts, 12 Speaking clips; Speaking correlation r = 0.85). These are internal evaluations, not certified results, and we publish them with their sample sizes.
If you use ChatGPT or another AI to mark your IELTS practice
- Pasting in the band descriptors gives the AI the words without the interpretation. Expect it to be too harsh on strong work and to treat every slip as a weakness.
- Ask for evidence for each criterion (Task Response, Coherence and Cohesion, Lexical Resource, Grammatical Range and Accuracy), quoting your own sentences. A band with no evidence behind it isn't feedback.
- Be careful with single scores above 7. That's where descriptor-only marking is least reliable.
- Compare with officially-marked sample scripts. They're the closest thing you have to an examiner's eye.
Writing Task 1, Writing Task 2 and Speaking Part 2, marked one criterion at a time, with a free account.
Try a free Writing paper