Why do AI IELTS essay scores so often stop at 6 or 6.5?

What to prepare first

We built a regression set from 17 official IELTS samples with examiner scores and comments, spanning Band 4 to 8.5. These samples provide both a credible score and examiner evidence that can be checked against the model output.

How to revise the next version

The ceiling came from stacked biases: conservative scoring, counting absolute errors rather than frequency, penalising one defect across several criteria, missing high-band anchors, and model differences. Five errors in 350 words are not equivalent to five in 180, and one expression problem should not be punished four times.

• “When unsure, score lower” compounds every ambiguous judgement downward
• Judge language control by error frequency, not raw error count
• Charge a defect only to the criterion it actually affects
• Calibration must include high-band work and allow occasional errors at Band 8

Why changing the rubric was not enough

The v3 rubric added anchors through Band 9, allowed occasional high-band errors, removed blanket downward scoring, mapped each criterion to official descriptors, and used official half-band rounding. MAE narrowed to 0.83–1.00, yet many Band 7.5–8.5 Task 2 samples still scored 5.5–6.5. A better rubric could not fully compensate for model judgement.

What improved in the Qwen regression

After cross-model comparison, the system moved to Qwen. From 16–25 July 2026, ten consecutive five-case scheduled regressions passed with MAE at 0.3–0.4; 35 fixed-sample cases ran in the most recent seven days. This shows improved stability on those calibration points and much less mechanical compression around 6–6.5.

What these numbers do not prove

A scheduled regression on five fixed samples is not a new independent validation. The full v7 17×3 engineering gate was stopped after 60 minutes without a final report, and a fully untouched blind set has not yet been completed. We therefore do not claim proven accuracy, superiority to examiners, an accuracy percentage, or guaranteed improvement.

• The score is a conservative reference range, not an official IELTS prediction
• Fixed regression demonstrates stability, not external blind validation
• Any single AI score should be checked against essay evidence

The score is an entry point, not the answer

For a learner stuck between 6.5 and 7, an isolated number is not enough. Useful AI feedback identifies the criterion holding the score back, quotes the relevant essay evidence, and gives a revision that can verify change. With ChatGPT or any other tool, provide the full prompt, essay, target band, and concern—then treat the output as clues rather than a verdict.

• Whether the score is supported by essay evidence
• Whether feedback chooses one first-priority issue
• Whether advice can become a same-paragraph revision that verifies change