How our CELPIP AI scoring works

What you get for each answer

• An estimated level range (for example 8–9) with a single best estimate and a confidence label. Ranges are deliberate: one practice answer is not enough to pin down a level.
• Four-criterion feedback using the dimensions CELPIP publishes: Content/Coherence, Vocabulary, Listenability (Speaking) or Readability (Writing), and Task Fulfillment.
• Your top fix, quoting your own words, plus a second fix.
• A model answer to the same prompt, one level above your target, and a button to try again.

What is sent, and what stays on your phone

Speaking: your phone turns your speech into text on the device. The audio stays on your phone. Only the transcript and timing measurements (how long you spoke, words per minute and pauses) are sent, with the task, to our server.

Writing: the text you wrote and the task.

Our server forwards the request to a Google Gemini model on Google Cloud’s Vertex AI, which acts as our data processor. We do not keep the content after the response, and Google’s terms say it does not use the data to train its models. Details are in the privacy policy.

How the estimate is made

1. The model receives the task, its time or word limits, and a rubric we wrote from the public CELPIP rating dimensions and level descriptions, in our own words.
2. Simple measurements are added: word count, speaking time, words per minute and pauses for Speaking; word count and paragraphing for Writing.
3. The model must answer in a fixed format: a level for each criterion, quoted evidence from your answer, two fixes and a model answer. Our server checks the format and rejects malformed results.
4. The criteria are combined into the range you see. CELPIP does not publish how it weights the dimensions, so our combination is our own.

What it cannot do

• It does not hear you. Because only the transcript is sent, pronunciation, intonation and accent are not assessed. Listenability is estimated from fluency signals (pace, pauses, self-corrections visible in the transcript) and grammar. A real rater hears much more.
• Transcription errors count against you. If your phone mishears a word, the model sees the wrong word. Speak clearly and check the transcript shown with your result.
• Estimates can be off. An estimate can differ from the level a trained rater would give by a level or more.
• It is not a CELPIP rater. Real tests are rated by trained raters (Writing with an AI-human hybrid system reviewed by people). Our estimate cannot predict your official result and is never a guarantee.

How we test it

Before release, we run the scorer on a set of 60 original practice answers written at known target levels across all Speaking and Writing tasks. Our release target is that at least 85% of estimates fall within one level of the intended level. We will publish the measured result on this page once it has been run on the release version. We do not have, and do not claim, any comparison with official CELPIP results.