How Langujet's Feedback Works

A score on its own doesn't help you improve. Useful feedback is explainable, specific, and tied to the official criteria: it points to the exact sentence or moment that costs you marks and names the reason. We optimise for that, not for producing an impressive-looking number.

IELTS Writing and Speaking are each marked on four criteria; TOEFL uses task rubrics scaled to a section score. Our feedback maps to those same criteria so you can see which one is the bottleneck and target it.

When you submit a response, the model evaluates it against the published band descriptors and rubric criteria, produces an estimate per criterion, and highlights specific issues — task response, coherence, vocabulary range, grammar accuracy, and, for Speaking, fluency and pronunciation features. The output is designed to be actionable: what's wrong, why, and what to do next.

Because it's automated, you can use it repeatedly and immediately, which is what makes tight practice loops possible.

It is strong at the things that recur and are pattern-based: grammar and word-choice errors, repetition, weak structure, under-developed task response, hesitation and pacing in speech, and consistency across many attempts. Used this way, it surfaces the issues most candidates can act on.

It is weaker at the things human examiners weigh with judgement: subtle tone, cultural nuance, originality of argument, and edge cases. Our estimates approximate official scoring but are not identical to it, and they should be read as guidance — a direction and a diagnosis — rather than a guaranteed result.

For attempts that matter, a qualified human corrector can review your work alongside the AI estimate. Humans add the nuanced judgement and accountability that automated scoring can't fully replace, while the AI keeps everyday practice fast and cheap.

We're deliberate about this split rather than pretending AI does everything: speed and volume from the model, nuance and high-stakes judgement from people.

We tell you when feedback is AI-generated and when it comes from a human corrector. Estimates are presented as estimates, with the criteria they're based on. We avoid inventing precision we can't back up — which is why you won't see fabricated accuracy percentages or guaranteed-score claims on this site.