How Intellecta's AI feedback works

Intellecta uses two different scoring systems, and they are reliable in very different ways. This page explains which is which, so you know how much weight to put on any given result.

Objective questions: answer-key scoring

Reading and Listening answers are checked against a verified answer key stored with the practice set. No language model is involved. That means:

• The same answer always produces the same result.
• Supported question types award partial credit where the official format does, including multi-select and prose-summary questions.
• If an answer is marked wrong and you believe the key is mistaken, that is a bug we can fix — please report it.

What these results are not is a band score. Our practice sessions are shorter than a full official section, so we report practice accuracy rather than converting to a band.

Writing and Speaking: model-generated feedback

Writing and Speaking responses are assessed by a large language model prompted with the published assessment criteria for the specific task — Task Response, Coherence and Cohesion, Lexical Resource, and Grammatical Range and Accuracy for IELTS Writing, and the corresponding criteria elsewhere. Speaking responses are first transcribed from audio, then assessed from the transcript together with timing and delivery signals.

The feedback is designed to be reviewed, not accepted. It is presented as specific observations and suggested revisions rather than as a verdict.

Where the feedback is reliable

• Identifying concrete errors. Grammar mistakes, wrong collocations, missing articles, and tense inconsistency are found accurately.
• Structural observations. Missing overview in Task 1, absent topic sentences, an unclear position, a missing conclusion.
• Repetition and range. Overused connectives and repeated vocabulary are detected consistently.
• Actionable revisions. Suggested rewrites of specific sentences are usually genuinely better than the original.

Where it is least reliable

• The exact band number. Any estimate should be read as a range of roughly plus or minus half a band, and a trend across several attempts is more meaningful than a single result.
• Judging ideas. A model can tell whether an argument is developed; it is weaker at judging whether it is convincing.
• Pronunciation. Assessment runs partly through transcription, so accent features that a human examiner would accept can be represented imperfectly.
• Unusual but correct English. Valid choices that are simply uncommon are occasionally flagged as errors.
• Very short responses. With little text to work from, feedback becomes generic.

Why we do not claim to predict your band

Official band scores come from trained examiners applying detailed descriptors under controlled conditions, with standardisation processes behind them. No automated system that scores a single practice answer can reproduce that, and platforms claiming precise band prediction are overstating what the technology does.

The useful question is not "what band is this?" but "what specifically is holding this answer back?" — and that is what the feedback is built to answer.

How to get the most out of it

1. Write or speak under real time limits. Feedback on an untimed answer tells you little about exam performance.
2. Read the criterion-level comments before the score. That is where the actionable information is.
3. Log the errors that repeat. The mistake notebook exists for this — three recurring errors fixed is worth more than fifty one-off corrections.
4. Redo the same task after reading the feedback. The second attempt is where the learning happens.
5. Get human feedback too, if you can. Use automated feedback for frequency and consistency, and a teacher for judgement.

Your data

Responses you submit are processed by a third-party AI provider in order to generate feedback. What is stored and for how long is described in the privacy policy.