We are running an online course designed to improve IELTS Academic Writing performance skills by focusing on a curated set of coded skills rather than wasting time learning English without direction. In our first (rough and unofficial) pilot, 8 of 10 learners improved by half a band in an average of 7 days; research published by the IELTS partners suggests that gain typically takes around 90 days of intensive study. The course has been improved since that first cohort based on what the data showed, and we are looking for pilot 1.1 participants to test it. There is also a bonus round: take part in research comparing the same writing course materials taught by AI alone or by AI and teacher together - we plan to publish the full anonymised dataset to fan the flames of the AI/human debate in education. Different assumptions are going to take a hit in this experiment; we're just not yet sure whose.
IELTS Laboratory is built around the Diagnostic Learning Cycle: write first, diagnose, learn, rewrite a different task, evaluate. A number of new studies have appeared in the last few months that test important assumptions more directly than anything available when I started building. This short article summarises a few.
1 - The first direct test on IELTS writing
Until recently there was very little research on AI feedback and IELTS specifically, but that has changed. Damanik et al. (2026) ran the closest design to ours yet published. Eighteen learners completed four iterative IELTS Task 2 tasks, receiving immediate feedback from an AI tool aligned to the band descriptors, with essays scored by certified examiners before and after. They found significant improvement across all four criteria over the four weeks, along with increased learner autonomy and deeper engagement with revision. The sample is small, and I treat the finding as promising rather than proven, but it is essentially the design our system runs, tested independently, and it points the same way as our own pilot data.
2 - A landmark comparison, and a warning inside it
Hartshorn and Pack (2026) ran the first controlled test of AI-delivered dynamic written corrective feedback: 41 learners over a 15-week semester, assigned by class to feedback from experienced teachers or from an AI. On accuracy the AI group matched the teacher group. Learners improved the same whether a human or a machine marked their errors, two years ago that result was not available.
3 - Why diagnosis, and why only five issues
Two findings underpin all of the above. Lee, Jang and Hannah (2025) compared automated diagnostic feedback with unaided self-assessment among 50 graduate students. The diagnostic group performed significantly better, and the authors argue that criterion-referenced diagnosis, which shows a writer how their work measures against specific standards, strengths included, supports self-regulated learning in a way that raw correction does not. Lee (2019) makes the complementary case against comprehensive marking: it overwhelms learners, damages motivation, and treats trivial and band-limiting errors as if they were equally important.
4 - The evidence loop
Every diagnosed essay adds structured information to the system. Each issue is coded against a defined taxonomy and joined to the learner's level, the task type, and whether it was later solved, improved, or persistent between drafts. Feedback is normally a disposable document, but we use it as a growing evidence base showing which problems occur most often at each band level, which lessons actually resolve which issues, and which problems continue after instruction. When we see that a lesson consistently does not move an issue from persistent to solved across learners, the lesson gets rebuilt and A/B tested.
This is where the published research has not yet reached. The 2026 studies validate the components individually: AI feedback can improve IELTS writing, models can score the criteria reliably, diagnosis outperforms correction, and prioritised feedback outperforms comprehensive marking. What nobody has yet studied is a system in which the teaching itself is continuously rebuilt from measured learner outcomes. That is the gap this project is built into, and the essays the system diagnoses are used to help check writing improvement . We are building a system based on evidence, which will be validated, audited, and improved as we progress.
5 - Where we are, and an invitation
So our rough and ready pilot 1.0 is complete. Ten learners, seventy submissions, three hundred and fifty coded writing issues. Eight of the ten gained a full half band, most within days; two did not improve, and we are working to improve this. We are not saying anything was proven in 1.0, the sample base was too small, etc. - but we felt we had to start somewhere. From 1.0 feedback data we have implemented a new system that we are keen for learners to try.
Pilot 1.1 runs next, with a target of 100+ participants, and this one is designed to explore not only course materials efficacy but also a bigger question. AI feedback and tutor feedback have now been compared, in writing and beyond; a recent meta-analysis found no significant difference in learning outcomes between the two - but feedback alone is not teaching. Teaching includes subtly personalised sequencing, guidance, discussion, and the judgement about what needs to be focused on next. To the best of my knowledge, nobody has published a study comparing AI-led instruction with teacher-led instruction in second language writing training using the same materials: the same lessons, the same diagnosis, the same measurement, with the only difference being whether a teacher is in the loop - our setup makes that comparison possible. IELTS Laboratory does work best as an online textbook, used with a teacher in a flipped format: the learner studies and writes between lessons, and lesson time goes to guidance and discussion rather than content delivery. However, some learners will work entirely alone with the diagnostic system, which provides the opportunity for comparison.
If you are preparing for IELTS Academic, you can take part. You get a structured writing training course built by an experienced practitioner, free, in exchange for agreeing to anonymised research use of your submissions. No identifying data is shared, ever.
And if you are a teacher, I want to make you a different offer. Bring a class, or even a handful of students, and teach them your way with the system underneath. At regular intervals you receive a private benchmark report: how your learners moved, criterion by criterion, against the learners who worked with the diagnostic alone. No identifying information about you or your students is published or shared; the report is for your own reference. We all know instinctively that good teachers add something computers cannot, and this is a chance to see it measured. My own hypothesis, for the record, is that the teachers win hands down, but it will be interesting to see in what ways they are different and by how much.
The data from this research project will be made available to the public on request. Where the data in 1.0 was very rough, we will be working with universities in BC to independently validate the data from 1.1 and publish in academic journals.
This experiment will start in August 2026 and continue until we have enough users to be confident in our course materials. If you are interested, please message me or comment below, the more the merrier!
References
Crosthwaite, P., & Sun, S. (2026). Generative AI and L2 written feedback studies: A scoping review. RELC Journal. https://doi.org/10.1177/00336882251386530
Damanik, S. F., Ariatna, Hartoyo, I., & Nasution, N. S. (2026). Enhancing IELTS writing task 2 performance through AI-generated feedback: A mixed-methods study on learner improvement and perceptions. Indonesian Journal of Applied Linguistics, 15(3), 592–604. https://doi.org/10.17509/g0n3qw05
Ebrahimi, M., Izadpanah, S., & Namaziandost, E. (2021). The impact of writing self-assessment and peer assessment on Iranian EFL learners' autonomy and metacognitive awareness. Education Research International, 2021, Article 9307474.
Hartshorn, K. J., & Pack, A. (2026). The effects of artificial intelligence-based dynamic written corrective feedback on second language writing and user sentiment. RELC Journal, 57(1), 69–88. https://doi.org/10.1177/00336882251405498
Huang, B. H., & Yan, X. (2025). Generative artificial intelligence in English language education: Potential, challenges, and the path forward. TESOL Quarterly, 59(S1), S330–S344. https://doi.org/10.1002/tesq.70042
Kaliisa, R., Misiejuk, K., López-Pernas, S., & Saqr, M. (2026). How does artificial intelligence compare to human feedback? A meta-analysis of performance, feedback perception, and learning dispositions. Educational Psychology, 46(1), 80–111. https://doi.org/10.1080/01443410.2025.2553639
Lee, I. (2019). Teachers' frequently asked questions about focused written corrective feedback. TESOL Journal, 10(3), e00427. https://doi.org/10.1002/tesj.427
Lee, M., Jang, E. E., & Hannah, L. (2025). Automated diagnostic feedback vs. self-assessment: Rethinking feedback mechanisms on academic writing development. TESOL Quarterly, 59(S1), S280–S317. https://doi.org/10.1002/tesq.70032
Liu, T., Ye, L., & Yan, W. (2026). A framework for evaluation of large language models in essay assessment: Reliability, alignment, and causal reasoning. Computers and Education: Artificial Intelligence, 10, 100565.
Mizumoto, A., & Eguchi, M. (2023). Exploring the potential of using an AI language model for automated essay scoring. Research Methods in Applied Linguistics, 2(2), 100050.
Pham, L. N. (2023). The interplay between learner autonomy and indirect written corrective feedback in EFL writing. TESOL Journal, 14(2), e694.