Two of the four CELPIP components are counted. Two are judged. On exactly how the judged half is judged, Paragon's own website gives three different answers — and all three are live right now.
The FAQ says the Writing Test is scored by an AI-human hybrid system. The test results page describes four to six human writing raters and three to five speaking raters. The Score Guidance PDF says at least three certified raters. Those are three official Paragon sources, and they do not describe the same process.
This page quotes all three and does not pretend to know which one is current.
The failure
The failure is trusting a confident answer. Search this question and you will find prep pages, forum posts and coaches who state flatly that CELPIP Writing is "all human, unlike the machine-marked competition" — and others who state just as flatly that it is now AI. Neither camp is quoting Paragon, because Paragon does not speak with one voice on this. If you are choosing a test, or choosing how to write, on the strength of one of those claims, you are building on something nobody has verified.
Listening and Reading: counted, not judged
Listening is 38 questions. Reading is 38 questions. In both, you select from answer options shown on screen, and both are reported against a raw score out of 38 (test results).
Every official statement about raters that we could find names only Writing and Speaking. Nothing in the Listening or Reading description involves a human forming a view of your English — a question is right or it is not.
And one rule follows from the Score Guidance PDF that changes how you should sit those two components:
"there are no points deducted for wrong answers"
So never leave a question blank. A guess costs nothing and might be right. A blank is a guaranteed zero on that item. This is a small mechanical fact that quietly separates people who finish a timed part cleanly from people who leave four items unanswered because they ran short.
Speaking and Writing: four dimensions, and three of them are the same
Paragon publishes the rating dimensions in its Performance Standards, restated on the test results page. Set them side by side:
| Speaking | Writing |
|---|---|
| Content/Coherence | Content/Coherence |
| Vocabulary | Vocabulary |
| Task Fulfillment | Task Fulfillment |
| Listenability | Readability |
Three of the four dimensions are identical across the two components. Only the fourth changes, and it changes for the obvious reason — one is heard, the other is read.
Listenability is officially defined as: "How easy is it to listen to and understand the response?" Its sub-points include "Rhythm, pronunciation, and intonation".
Readability's sub-points are: "Format and paragraphing; Connectors and transitions; Grammar and sentence structure; Spelling and punctuation".
The practical consequence: organising an answer, choosing the right words, and actually doing what the task asked are scored under the same three headings whether you type it or say it.
The part where the official sources disagree
Three Paragon sources, all currently published, all describing how Speaking and Writing get marked:
| Source | What it says |
|---|---|
| FAQ, answering "Is my test scored by artificial intelligence?" | "The CELPIP Writing Test is scored by an AI-human hybrid system that combines artificial intelligence with our human rating panel." |
| Test results page | "Each CELPIP speaking performance is rated by three to five speaking raters, and each CELPIP writing performance is rated by four to six writing raters." — and elsewhere, "scored by qualified raters trained to apply consistent criteria" |
| Score Guidance PDF | "The Writing and Speaking components are evaluated by at least three certified raters." |
These cannot all be descriptions of the same process. One names artificial intelligence in the loop; one describes a purely human panel with a specific headcount range; one gives a different minimum headcount. They may be sequential — an older page left standing, a newer one added — but no official source says which is which, and this page does not know which is current.
What you can take from it: anyone who tells you confidently that CELPIP Writing is "all human" or "all AI" is not quoting Paragon.
One more limit, and it matters. Whether Speaking is AI-scored is not addressed by any official source we found. The FAQ answer names the Writing Test only. Do not read that as a denial about Speaking, and do not read it as a confirmation either.
The scale, and the floor beneath it
Scores run 0 to 12, and each component gets its own CELPIP Level — there is no single overall level on a CELPIP report.
Underneath the four dimensions sits a rule that outranks all of them:
A response giving insufficient information to assess scores 0 or 1.
Levels 0 and 1 carry "Insufficient information to assess" and map to no CLB level at all. That is the consequence people miss. A response too short or too far off-task to be rated does not land you a low benchmark — it lands you outside the benchmark scale entirely.
What nobody publishes
Be clear-eyed about the size of the gap:
- How the four dimensions are weighted. No official weighting or points breakdown is published.
- How your two Writing tasks become one Writing level. Not published.
- Any raw-score-to-level cutoff. CELPIP labels its own conversion table "Approximate". The hard cutoff tables you find on prep sites are reverse-engineered by third parties, and they contradict each other. This page prints none, on purpose.
Re-evaluation, and what its refund rule tells you
Paragon offers a re-evaluation — a re-mark of what you already did, not a retake. Within six months of your test, one re-evaluation per component, with a turnaround of approximately one to two weeks. The fee is refunded if your Speaking or Writing score increases.
Look at that condition. The refund is tied to exactly the two judged components — the two where a second look can legitimately produce a different number. Nobody offers you money back on a re-count of 38 multiple-choice answers, because a re-count does not disagree with itself. The refund rule is Paragon quietly telling you where rater disagreement is expected to live.
What actually costs you
- Leaving Listening or Reading items blank when nothing is deducted for a wrong guess.
- Writing to a rubric you invented, when the four real dimensions are published for free.
- Believing a prep page's confident account of AI versus human marking over what Paragon actually publishes — which is three incompatible statements.
- Trusting a raw-score cutoff table. CELPIP calls its own version approximate.
- Producing a response too thin to assess, and landing on 0 or 1 — off the CLB scale, not low on it.
Try one
[Try a marked task →] — one full task, marked, no card required.