How Accurate Are AI IELTS Writing Checkers? What a 97-Essay Test Shows — and How to Use Them Without Fooling Yourself

Picture a candidate three weeks before her IELTS test. She writes a Task 2 essay on whether governments should fund the arts, pastes it into three different AI checkers, and gets back three different verdicts: 5.5, 6.0 and 7.0. One tool says her vocabulary is the problem. Another praises her vocabulary and blames her paragraphing. The third gives her a seven and tells her she is ready.

She now knows less than she did before she asked. If the seven is right, she should stop worrying about writing and spend her last weeks on listening. If the 5.5 is right, she needs to change how she builds an argument, and she needs to start today. The scores are not just numbers to her; they decide how she spends the most valuable study time she has left.

AI writing feedback has become the default way many IELTS candidates practise, simply because a human teacher who marks every essay within an hour is expensive and rare. So the question is worth asking plainly: how close do these tools actually get to a real examiner? And once you know that, how should you use them?

How the test worked

The only fair way to measure a scoring tool is to give it essays whose true Band is already known. Fortunately, those exist. Cambridge and the official IELTS partners publish sample responses together with the overall Band an examiner awarded, partly so that teachers and candidates can calibrate their own judgement.

The test collected 97 of these official responses: 51 Task 1 answers and 46 Task 2 essays, covering both Academic and General Training. They came from the official IELTS sample materials, official practice tests and the Cambridge IELTS books 18 to 21. Each response was then scored by five popular tools — ChatGPT, Engnovate, Writing9, IELTSWritingChecker.org and IELTS Writing Practice — and each tool’s overall Band was compared with the examiner’s.

A word on who ran it. The test was carried out by the team behind IELTS Writing Practice, one of the five tools, and that same team wrote this article. You should read the results with that in mind. That is also the reason the full method, the source books and the limitations are published in the 97-essay IELTS writing scoring benchmark, so that anyone can check the approach rather than take a vendor’s word for it.

Three measures matter. The first is the average distance between the tool’s score and the examiner’s, regardless of direction; lower is better. The second is how often the tool lands within half a Band of the examiner, which is roughly the margin a candidate cares about. The third is bias: whether a tool tends to score too high or too low on average.

What the numbers say

Tool	Average error (Bands)	Within 0.5 Band	Within 1 Band	Average bias
IELTS Writing Practice	0.53	74.2%	93.8%	−0.27
IELTSWritingChecker.org	0.57	70.1%	86.6%	−0.38
ChatGPT	0.88	46.4%	76.3%	−0.85
Engnovate	0.94	39.2%	73.2%	−0.86
Writing9	1.72	19.6%	35.1%	−1.58x

97 official responses with published examiner Bands; tools observed through 30 August 2026.

The gap between the top and the bottom of the table is wider than most candidates would guess. The two specialised checkers at the top came within half a Band of the examiner roughly three times out of four. The purpose-built AI IELTS writing checker from IELTS Writing Practice missed by more than a full Band on only 6 of the 97 responses. At the other end, Writing9’s average error was larger than a full Band and a half, and it landed within half a Band less than one time in five.

ChatGPT sits in the middle, and its result is perhaps the most useful one for ordinary learners, because so many people simply paste an essay into a chatbot. Its general knowledge of the IELTS descriptors is good, but in this test it was within half a Band less than half the time. It is a helpful reader. It is not a reliable examiner.

One detail affects Academic candidates directly. Academic Task 1 asks you to describe a chart, map or diagram, and a tool that cannot see the picture cannot check whether your numbers and comparisons are right. In this test, Writing9’s public workflow did not accept the Task 1 image, which is part of why its Task 1 scores drifted so far. IELTS Writing Practice, IELTSWritingChecker.org, ChatGPT and Engnovate all took the image. If you are preparing for the Academic module, check this before you trust any Task 1 score.

The pattern nobody expected

The most striking finding is not in the ranking at all. Look at the bias column: every one of the five tools, on average, scored the official responses lower than the examiners did. For the best tools the difference is small, about a quarter of a Band. For ChatGPT and Engnovate it is close to a full Band, and for Writing9 it is more than one and a half.

That has a practical consequence. A candidate who practises with a harsh tool may decide she is a 6.0 when an examiner would give her a 6.5 or 7.0. She may then rebuild an essay structure that already works, or delay a test date she was ready for. A score that is too low feels like rigour, but it misleads you just as much as one that is too high.

The test also suggests that Task 2 is harder to score than Task 1 for every tool. That makes sense: a Task 1 report can be checked against the data, while a Task 2 argument asks the reader to judge how well an idea is developed, which is exactly where human examiners also need the most training.

One caveat: the test compares overall Bands only, because official samples rarely publish criterion scores. The ranking also held on the responses from the newest Cambridge books, 20 and 21, which AI models are least likely to have seen.

How to use a checker well

A score, even an accurate one, does not improve your writing by itself. What helps is what you do next. Write under timed conditions, get feedback the same day, and then rewrite only the paragraphs affected by your lowest-scoring criterion rather than trying to fix everything at once. A few days later, write on a new question and see whether that weakness has moved.

Because every tool in the test tended to score low, do not panic over a single result. Treat a score as a range of about half a Band either side, and watch the trend over five or six essays.

For model answers, use samples whose Bands are known, such as the official Cambridge samples or the Band 8–9 collections on IELTS Writing Practice, and borrow their sentence structures rather than memorising whole essays. For practice questions, recent IELTS writing questions, which IELTS Writing Practice publishes every month, are closer to what you will meet on test day than prompts from ten years ago.

Back to the candidate with three scores

Return to the candidate with three different verdicts. With these results in hand, her next step is clearer. She can stop averaging three tools and choose one that has been measured against examiners, one that accepts Task 1 images if she is taking the Academic test. She can read every score as a range rather than a sentence passed on her. And she can spend her remaining three weeks on a simple cycle: timed writing, same-day feedback, and a focused rewrite of her weakest criterion.

AI checkers will not replace a good teacher, and the best of them would not claim to. But for a candidate writing three essays a week with no teacher available, the difference between a tool that is usually within half a Band and one that is usually more than a Band out is the difference between useful practice and expensive guessing. It is worth knowing which one you are using.