Can ChatGPT score your TCF Canada writing?

Before you read. NCLC Prep sells an AI corrector for TCF Canada writing. So we have a stake in this question. This page also says what our own tool cannot do.

The short answer is that a score out of 20 from ChatGPT, from us or from any AI tool is an estimate. It is not France Éducation international’s (FEI) mark, and we found no independent study comparing ChatGPT scores with official results.

Why ChatGPT’s score moves

Nobody knows FEI’s rule. Two markers give each of the three tasks a level, then a calculation rule FEI does not publish produces the final mark. FEI publishes no annotated scripts either. So no AI can reproduce the official mark exactly. It can only get close.

The score depends on the prompt. On forums, candidates report the same text dropping from about 15 out of 20 to 11 or 13 once they add the marking criteria to the prompt (one example on Reddit). These are reports, not a measurement, but they show the number follows the question you ask.

A chat assistant is built to help. That is our reading. A tool that encourages tends to round up, especially when you ask “is this B2?”. A candidate who gets 14 out of 20 in practice and 9 in the exam has lost weeks.

One task is not the test. The official mark combines all three tasks. A single text, even a well-scored one, tells you nothing about the task 3 you ran out of time for.

What no AI can judge

• Whether the markers will see your text as learned by heart, a case FEI lists for “A1 non atteint”.
• Whether your handwriting will be legible on paper, another listed case.
• Whether you would write this text within 60 minutes for three tasks, without help or autocorrect.
• How your six levels combine into a mark out of 20.

FEI itself uses an automated correction, FIDELIA, for some computer-based scripts, but the second correction is always human. The details are on is the TCF Canada marked by AI.

The limits of our own corrector

Our corrector runs on Gemini, Google’s model. It gives a score out of 20 with a range, and a level for each criterion. Here is what you should know.

• It is an estimate on one text, never the official mark.
• The range is real. On 2 November 2026, the same task 3 text got 12 out of 20 in the English interface and 11 in the French one. Both are inside the two-point range, but they are not the same.
• It does not know how you wrote the text, or how long it took.

A free prompt for ChatGPT

If you use ChatGPT, ask it for mistakes rather than a score. Here is a prompt we wrote. It is in French so the feedback stays in French. Paste it, then your consigne and your text.

Tu es correcteur de français langue étrangère. Je prépare la tâche [1, 2 ou 3] de l’expression écrite du TCF Canada. Voici la consigne, puis mon texte.

1. Compte les mots de mon texte, en prenant pour un mot tout ensemble de signes placé entre deux espaces, et dis si je suis dans la fourchette de la consigne.
2. Dis si chaque point demandé par la consigne est traité, oui ou non, en citant la phrase qui le traite.
3. Liste chaque erreur de grammaire, d’orthographe ou de vocabulaire, avec la correction et la règle en une phrase.
4. Dis si le registre convient au destinataire.
5. Ne donne pas de note sur 20. Termine par les deux erreurs qui reviennent le plus souvent.

Check the word count yourself, because language models often miscount. Keep the same prompt from one text to the next, or you are comparing different questions rather than your progress.

How to use an AI without fooling yourself

1. Use it to find mistakes, not to find out whether you have B2.
2. Follow the trend over several texts with the same tool, rather than one score.
3. Write the mistakes that come back in a notebook, as in the 4-week plan.
4. Compare your text with annotated corrected answers to see what a B2 answer does that yours does not.

If you want an estimated score with a range and every mistake explained, you can try NCLC Prep, with two free corrections and no account.