Methodology
How accurate is our AI IELTS scoring?
Ollo Labs's AI IELTS scorer predicts overall band scores within ±0.5 of a certified IELTS examiner on more than 90% of writing samples. The model grades against the four official IELTS criteria - Task Achievement / Response, Coherence and Cohesion, Lexical Resource, and Grammatical Range and Accuracy - using the standard IELTS averaging rules to compute the overall band. Speaking AI scoring is similar in approach but slightly noisier (±0.5–1.0) because pronunciation and pacing are harder signals to extract reliably. Use AI band scores as a calibrated estimate of test readiness, not as a guarantee of your real exam result.
How we calibrate the model
Calibration uses a held-out set of essays graded independently by certified IELTS examiners. For each essay, we record the examiner's overall band, the four criterion sub-bands, and inter-rater agreement when more than one examiner is available. We then measure mean absolute error and 90th-percentile error of the model's predictions against this set.
What the numbers say
- Writing - overall band: within ±0.5 of the examiner band on more than 90% of samples.
- Writing - criterion sub-bands: Lexical Resource and Grammatical Range are the most accurate; Task Response is the most variable (it's also the most subjective for human examiners).
- Speaking - overall band: within ±0.5–1.0 of an examiner band, depending on audio quality and accent. We re-prompt and re-score sessions where the audio is unintelligible.
- Reading and Listening: objectively scored against the answer key - there is no AI variance, only the standard IELTS raw-score-to-band conversion.
Limits of AI scoring
Two cases the AI handles less well: essays that are technically high-band but off-topic in subtle ways, and Speaking responses with strong regional accents the model has limited training data for. In both cases, the model tends to be conservative - under-predicting rather than over-predicting the band.
Independent academic studies of LLM-based IELTS scoring (including ChatGPT-class models) report inter-rater agreement with human examiners in roughly the same range as agreement between two human examiners. Treat AI band scores the way you would treat a second examiner's opinion: useful, calibrated, and not a substitute for a real test centre's assessment.
Frequently asked questions
How accurate is AI IELTS scoring compared to a human examiner?
Across the published research and our own internal calibration, well-designed AI IELTS scorers fall within ±0.5 of a human examiner's overall band on roughly 85–95% of writing samples. Ollo Labs's own writing scorer sits at the upper end of that range against essays graded by certified examiners. Speaking AI scoring is harder - variance is closer to ±0.5–1.0 because pronunciation and natural pacing are noisier signals.
What does Ollo Labs's AI grade against?
The four official IELTS criteria, exactly as a human examiner does: Task Achievement / Response, Coherence and Cohesion, Lexical Resource, and Grammatical Range and Accuracy. Each criterion gets its own sub-band, and the overall band is computed using the standard IELTS averaging rules.
Why is AI sometimes 0.5 bands different from a human examiner?
The same essay can receive different bands from two human examiners - IELTS itself acknowledges examiner variance of around ±0.5. AI variance is similar in magnitude but distributed differently: AI tends to be more consistent on lexical and grammatical features, and less reliable on subjective task-response judgements (e.g. whether your position is 'clearly developed throughout').
Is AI scoring fair?
AI scoring removes some sources of human bias (examiner fatigue, halo effects from handwriting, time-of-day) but introduces others. We mitigate model bias by calibrating against a diverse sample of essays from candidates writing in their second, third or fourth language, across both Academic and General Training tracks.
Can I trust an AI band score before booking the real IELTS?
Use it as a calibrated estimate, not a guarantee. A reasonable rule: book the real IELTS when you consistently hit 0.5 bands above your target on AI-graded responses across three different prompts. The 0.5-band buffer absorbs both AI variance and your normal day-to-day variance.
See your AI band score
Score your first essay free and compare it against your target band.