Is AI IELTS Scoring Accurate? What the Data Says

"Is AI IELTS scoring accurate?" has a short answer and a longer one. The short answer: accurate enough to tell you which of the four criteria is holding your band down, and not accurate enough to tell you your band. The longer answer needs to start with how the real thing is scored, because "accurate" only means something against a reference, and the reference is a human examiner who is not perfectly consistent either.

How IELTS Writing is scored by humans

A certified examiner reads each task and gives a band on four criteria: Task Achievement for Task 1 or Task Response for Task 2, Coherence and Cohesion, Lexical Resource, and Grammatical Range and Accuracy. The four bands are averaged into a task band, Task 2 counts double, and the result is rounded to the nearest half band. The examiner works from band descriptors, a grid of what each band looks like on each criterion; the public version of the descriptors is published by the test organisers, and the version examiners use is more detailed.

Two things about this process matter for the question of accuracy. First, a script can be seen by more than one examiner: the organisers describe a system in which samples are double marked by senior examiners, and in which a Writing or Speaking band that is markedly out of line with the candidate's Listening and Reading bands is automatically re-marked. Second, two trained examiners do not always give the same band to the same essay. The organisers publish reliability figures for the examiner-marked sections on their test statistics page; the figures are high, and they are not perfect, which is another way of saying that a margin between two human readings exists. That margin is the yardstick for any AI: a tool that lands within it is doing what a second examiner does, and a tool that claims to do better than that is claiming something the reference cannot confirm.

How an AI checker scores

An AI checker works from the same four criteria. The essay and the question are sent to a language model with instructions that include the band descriptors, and the model returns a band and a justification for each criterion, with the sentences it based them on. On this site, the overall band is then computed by the site's own code from the four criterion bands, not taken from a number the model prints, so that the arithmetic is the same for every essay. The whole essay is read; nothing is sampled or cut (the only limit is a maximum length far above any exam answer).

What a model sees well are the things that are visible in the text itself. Grammatical range and accuracy: it can count error-free sentences, spot the tense that slipped, the article that is missing, the subject that does not agree with its verb, more consistently than a tired reader. Lexical resource: it notices repetition, collocations that do not exist, a word used in the wrong register, and it can tell a learned phrase from a natural one. Structure: paragraphs, topic sentences, whether the conclusion says what the introduction promised, whether the linking words carry meaning or decorate.

What it sees less well are the things that depend on judgement rather than on the text. Fine relevance: whether the essay answers this question or a neighbouring one, whether an example really supports the claim, whether a position stated in the introduction is the one argued in the body. Tone: an essay can be grammatically clean and still read as a list, or as a speech, and a model is a less reliable judge of that than an examiner. Originality: a model tends to reward ideas it has seen often, and an unusual but well-argued position can be under-read. For Speaking, one more limit is stated on the site in these words: Pronunciation is assessed by the AI examiner from your audio recording; treat it as the least certain of the four criteria.

What our own data says

No accuracy figure is claimed on this site, because none has been measured yet against official results. What has been measured is the gap between the band test-takers received and the band they expected, which is a different thing and is published as such. After each evaluation, one question is asked: how did the band compare with what you expected? Among 405 test-takers who answered, internal tests excluded, collected through 6 September 2026, 58% said the band was lower than they expected, 25% said it was about right, and 17% said it was higher. The figures, and the date of their last update, are on the proof page.

Read carefully, that split says two things. The scoring leans strict rather than encouraging, which is by design: it is anchored to the descriptors, not tuned to feel good. And most people expect more than they get, which is what teachers report from paper practice tests too. What the split does not say is how far the AI band sits from an official band, because expectation is not an official result.

That comparison starts at the end of September 2026, with the first pair of results published whatever it shows: the band this site gave the founder before his own exam, against the official band on his Test Report Form. Test-takers who sit the real exam after practising here can add their own pair anonymously on the same page, and this article will be updated as the pairs come in.

AI checker against human examiner

How to use an AI band well

1. Compare two essays, not one essay with the truth. Write a Task 2 answer, read the notes, rewrite it on its weakest criterion, submit the rewrite. The difference between the two bands on that criterion is the most reliable number the tool gives you, because both readings share the same bias.
2. Look for what repeats. One essay marked down on Coherence can be the topic; three essays marked down on Coherence is you. The sentences the checker quotes are the material: collect them across essays and the pattern is usually visible in a week.
3. Read the criterion bands, not the overall band. A 6.5 overall hides a 7 in Grammar and a 6 in Task Response, and those two need different work. The overall band is the least useful line of the feedback.
4. Do not aim at the number. An essay rewritten to raise the AI band, by adding long words or extra linking phrases, gets worse for a human reader and often for the model too. Rewrite to fix the note, and let the band follow.
5. Keep a human in the loop for the ideas. If a teacher or a fluent friend can read one essay a week, give them the question and the argument, and leave the grammar to the checker.

The free AI checker for Writing Task 2 page lists what the checker flags on each criterion, and the Writing checker itself takes both Task 1 and Task 2.

The limits, in a list

• The band is an estimate. It can differ from an official result by roughly half a band to a full band, and by more at the extremes: very short answers, very long answers, off-topic answers, or a style the model reads as unusual.
• No accuracy figure has been measured yet against official results; the first pair arrives at the end of September 2026 and the figure will be built from pairs as they come.
• The model reads text, not the exam. Handwriting, time pressure, nerves and the difference between a practice question and the real one are not in the loop.
• Fine relevance, tone and originality are judged less reliably than grammar and vocabulary; an essay can be marked well on the surface and still not answer the question as an examiner would read it.
• A model can be swayed by length and by rare words, and the calibration work on this site consisted largely of removing that bias; it is reduced, not gone.
• For Speaking, the pronunciation band is the least certain of the four, because it is estimated from a recording rather than from a room.
• The same essay submitted twice can come back with slightly different notes; the criterion bands are stable, the wording is not.
• None of this is official, and it cannot be made official by any amount of calibration. Only the exam gives an IELTS band.

Quick answers

Can I trust an AI band?

For direction, yes; for the number, no. Trust it to tell you which of the four criteria is lowest, whether a rewrite improved that criterion, and which sentences cost marks. Do not trust it as a prediction of your official band: it can differ from a real result by roughly half a band to a full band, more at the extremes, and no measured accuracy figure exists yet on this site.

Why is the AI stricter than I expected?

Because it is anchored to the band descriptors rather than tuned to be encouraging, and because most candidates expect more than an examiner gives. In the feedback collected on this site, most test-takers found their band lower than they expected, a minority found it about right, and a smaller minority found it higher; the exact split and its date are on the proof page. A strict estimate in practice is more useful than a kind one on the day.

Does the AI see my whole essay?

Yes. The full text you submit, with the question you answered, is sent for evaluation; nothing is sampled, summarised or cut, and the only limit is a maximum length far above any exam answer. The band and the notes are about exactly the words on your screen, and the overall band is computed by the site from the four criterion bands rather than printed by the model.

Is ParrotTalk official?

No. ParrotTalk is an independent practice site, and the bands it returns are AI estimates against the official criteria, made to guide practice. Only a certified examiner gives an official IELTS band. This site is not affiliated with, endorsed by or connected to IELTS, the British Council, IDP or Cambridge English.