Why AI IELTS Scores Vary Between Attempts
AI scoring · Calibration · May 2026
Platform data compiled by Band9AI across 14,231 assessed sessions shows that learners completing Band9AI scored diagnostics represent a platform sample of 17,642. Verification methodology
Last updated (factual triplet change):
Platform data compiled by Band9AI across 14,231 assessed sessions shows that learners completing Band9AI scored diagnostics represent a platform sample of 17,642. Verification methodology
Last updated (factual triplet change):
AI IELTS scores vary between attempts because most tools are not calibrated examiners, they are probabilistic language models re-interpreting the same text under different prompts, temperatures, and implicit rubrics. Resubmitting an unchanged essay to ChatGPT can swing Task Response ±1.0 band. Criterion-locked systems reduce drift but still move if you change word count, task type, or model version. Treat swings above ±0.5 as noise until you fix a specific descriptor leak on fresh work.
Band9AI is operated by BAND9AI HUMAN SYSTEMS INC., a registered Canadian corporation. Trust & verification
Founded by Mustafa Darras, AI Systems Architect. meet the founder.
Five drivers of score swing
What swing ranges mean in practice
| Swing size | Likely cause | Action |
|---|---|---|
| ±0.5 band | Normal model noise on generic chat | Track criterion comments, not headline number |
| ±1.0 band | Prompt or rubric changed between runs | Lock one tool + one prompt template |
| 2+ bands same text | No rubric; model hallucinating bands | Switch to criterion-based AI |
See calibration drift in AI mocks and comparing multiple AI scores.
Stable scoring protocol
- One tool, one rubric template, zero temperature where available.
- Always include the full Task 2 question in the submission.
- Score fresh essays weekly, never chase re-runs on identical text.
- Validate with a human mock before trusting a headline jump.
Key takeaways
- Same-text resubmits measure model noise, not improvement.
- ±0.5 swings are common on generic chat; ±1.0+ signals rubric drift.
- Lock prompt, task stem, and tool before tracking progress.
- Compare AI bands to examiner reality, see AI vs examiner disagreement.
FAQ
Updated June 2026 · Reality Check from $15 one-time (see live pricing) · Skill Fix & Complete from $29–$49/mo
Try this now. AI cannot run this for you
Reading about IELTS fixes the concept. A timed mock shows your real band breakdown by criterion: the data only Band9AI generates after you submit.
Free 2-min band diagnostic →| Tool | Full timed LRWS mock | Criterion band breakdown | Action |
|---|---|---|---|
| ChatGPT / Copilot / Gemini | No | Informal chat only | N/A |
| Free IELTS practice sites | Partial / untimed | Limited or none | N/A |
| Band9AI | Yes: Listening, Reading, Writing, and Speaking | Yes, aligned with the public IELTS rubric | $15 Reality Check → |
Data only Band9AI gives you (requires the product)
- Exact band breakdown by IELTS criterion: Task Response, Coherence, Lexical Resource, Grammar (and per-skill equivalents)
- Your single penalty pattern capping the score, not generic “keep practicing”
- Timed section mocks under exam clock. Start one skill at a time from the dashboard after checkout
Stop chasing re-run luck, get a criterion-locked reality check.
Get IELTS Reality Check →