How accurate is the AI feedback on TEF writing?
Short answer: the level estimate is rough. It can spot strong writing (about NCLC 8 and up), but it can't tell NCLC 4 from NCLC 7, so it can't tell you whether you've reached 7. Here is what we tested and what it shows.
How to read your estimate
Below NCLC 4–5: the few texts given this were rated NCLC 4–6, and most were very short.
NCLC 4–5 or 5–6: anywhere from beginner to about NCLC 7–8. It can't tell you whether you're at NCLC 7, the level Express Entry asks for.
NCLC 8–9 or 9–10: your text reached the minimum length with very few errors (6 or fewer per 100 words). In our tests, 7 of 7 texts given this were rated NCLC 8 or higher; in the exam texts it also went to 1 B1 text, so it can be too high.
Use the corrections and the three things to fix next: they show specific mistakes you can fix. The level estimate can't. It is not an official score and never replaces the real test.
What we found
- Strong writing is now recognised. 7 of 8 texts rated NCLC 8 or higher by the Centre for Canadian Language Benchmarks (CCLB) now get NCLC 8–9 or 9–10; before the error count was added, none did (the AI alone gave them 4–5 to 6–7).
- The middle is still wide. 12 of 12 texts rated NCLC 5 to 7 got NCLC 4–5 or lower, and so did 5 of 5 beginner (A1–A2) exam texts. A 4–5 can mean anything from beginner to about NCLC 7.
- It can be too high. One DELF B1 text with very few errors got NCLC 8–9; the exam level suggests NCLC 4–6. Few errors is a strong sign, not proof.
- Same text, same answer. All 13 DELF texts got an identical AI estimate in each of 3 runs, and the error count is the same every time for the same text.
Why we can't compare with real TEF scores (yet)
The organisation behind the TEF (CCI Paris Île-de-France) publishes sample questions, but no answers with their official scores. Without those, nobody outside the exam can check an AI against real TEF marking. The closest public evidence is French writing marked in other exams, and texts rated on the NCLC scale for teaching.
To build a real TEF benchmark, we're asking people who took the test to share their official scores. With their permission we compare them with the estimates GoExam gave them, and we'll publish the totals here.
Test 1: 13 officially marked DELF and DALF texts
- Texts. 13 French texts written in DELF, DALF and Alliance française exams, each published with its official mark: 8 from the Council of Europe's 2006 sample collection (prepared with CIEP and Alliance française) and 5 from Transforming FSL (marks reviewed by CIEP). We don't reproduce the texts here; the sources are public.
- Expected level. Each text was written for an exam at a given CEFR level and received the mark shown, so we use that exam's level as the expected level. There is no official CEFR → NCLC table, so we combine CCI's two official tables (TEF score → CEFR and TEF writing score → NCLC). That gives a band: A1 = below NCLC 4, A2 = below NCLC 4 to NCLC 4, B1 = NCLC 4–6, B2 = NCLC 6–8, C1 = NCLC 8–11.
- Grading. Each text went through the same AI grader GoExam users get, 3 times (last run October 25, 2026), with the original task and its own minimum length, then through the error count (November 2026).
| Text | Expected → estimate (NCLC) | Fit |
|---|---|---|
| DELF A1Holiday postcard · 47 words · mark 11 / 15 · source | below NCLC 4 → 4–5 | Higher |
| DELF A2Reply declining an invitation · 83 words · mark 9.5 / 12 · source | <4 – 4 → 4–5 | Overlaps, leans high |
| Alliance française A2Letter to a friend · 96 words · mark 8 / 10 · source | <4 – 4 → 4–5 | Overlaps, leans high |
| DELF A2Letter about a birthday party · 115 words · mark 12 / 13 · source | <4 – 4 → 4–5 | Overlaps, leans high |
| DELF A2Email declining an invitation · 82 words · mark 11.5 / 12 · source | <4 – 4 → 4–5 | Overlaps, leans high |
| Alliance française B1Short email giving advice · 45 words · mark 4 / 5 · source | 4–6 → 4–5 | Inside |
| DELF B1Essay on changes in your country · 178 words · mark 17.5 / 25 · source | 4–6 → 4–5 | Inside |
| DELF B1Forum post giving an opinion · 182 words · mark 24.5 / 25 · source | 4–6 → 8–9AI alone 5–6 · 3.3 errors / 100 words | Higher |
| DELF B1Forum post giving an opinion · 171 words · mark 19 / 25 · source | 4–6 → 4–5 | Inside |
| Alliance française B2Opinion article · 131 words · mark 16 / 20 · source | 6–8 → 8–9AI alone 5–6 · 3.8 errors / 100 words | Overlaps, leans high |
| DELF B2Formal letter (140 of 250 words) · 127 words · mark 15 / 25 · source | 6–8 → 4–5 | Lower |
| DELF B2Formal letter to persuade · 269 words · mark 15.5 / 25 · source | 6–8 → 8–9AI alone 5–6 · 6 errors / 100 words | Overlaps, leans high |
| DALF C1Summary of several documents · 204 words · mark 9.5 / 13 · source | 8–11 → 8–9AI alone 5–6 · 5 errors / 100 words | Inside |
Test 2: 22 texts rated on the NCLC scale
The CCLB's NCLC support kit includes handwritten texts by adult French learners, each assigned an NCLC level by the CCLB's experts. We transcribed the 22 texts from NCLC 4 to 12 word for word, keeping every mistake, and graded each once with the same AI (October 2026), then through the error count (November 2026). The tasks are a note to a landlord, an opinion essay and a job application letter. We don't reproduce the texts.
| NCLC level (CCLB experts) | Texts | Estimates |
|---|---|---|
| NCLC 4 | 2 | <4, 3–4 |
| NCLC 5 | 4 | 4–5, 4–5, <4, 4–5 |
| NCLC 6 | 3 | 4–5, 4–5, 3–4 |
| NCLC 7 | 5 | 4–5, 4–5, 4–5, 4–5, 4–5 |
| NCLC 8 | 3 | 9–10, 8–9, 5–6AI alone: 5–6, 4–5, 5–6 |
| NCLC 9 | 1 | 8–9AI alone: 5–6 |
| NCLC 10 | 2 | 8–9, 9–10AI alone: 5–6, 5–6 |
| NCLC 11 | 1 | 9–10AI alone: 5–6 |
| NCLC 12 | 1 | 9–10AI alone: 6–7 |
Why it squeezes to the middle, and what we tried
Before changing anything we set a bar: a fix had to do better on both tests, and the same text must not get a different band each time it's graded. Everything that asked the AI for a better score failed; counting errors passed.
- Rewording the instructions didn't move it. We tried dropping “when unsure, choose the lower level”, judging on what the writer can do rather than on errors, CEFR reference points, the official NCLC writing descriptors, and reference texts built into the grader. The estimates stayed almost identical.
- Other AI models did no better once converted to NCLC. Two Kimi models and Gemini 3 Flash placed the beginner and C1 texts more sensibly on the CEFR scale, but converted to NCLC they matched our grader on the strong texts.
- Ranking against reference texts spread the estimates out, but missed the exact level more often and sometimes overshot by two levels.
- Adjusting the scores afterwards made things worse. Raising the top band would make the same text jump between NCLC 4–5 and 6–7 from one grading to the next; stretching the scores pushed a B1 text to NCLC 8–9.
- The root cause: the AI's own scores for each criterion are almost the same for a beginner text and an NCLC 7 text. When that information isn't there, no adjustment can recover it.
- What worked: counting errors instead of asking for a score (November 2026). We count spelling and grammar errors per 100 words with LanguageTool. Texts rated NCLC 8 and up almost all have 6 or fewer; texts rated NCLC 4 to 7 have 11 to 37, mixed together. So when a text reaches the minimum length with 6 or fewer errors per 100 words, we raise the estimate to at least NCLC 8–9 (3 or fewer: 9–10). It never lowers an estimate, and it can't separate NCLC 4 from 7.
- Grammar errors alone, shown as a number (November 2026). Leaving out spelling and accents, the CCLB texts rated NCLC 4–5 had about 8–18 grammar errors per 100 words and those rated NCLC 6–7 about 4–8. That gap is real, but one A2 exam text also had about 7, so we don't turn it into a level. Your feedback shows the number itself, so you can track it from one text to the next.
The real fix for the middle still needs real TEF results from people who took the test.
Limits of these tests
- None of these are TEF tasks. TEF writing is a news-story continuation and a letter to a newspaper.
- In test 1, the expected levels are bands of up to four NCLC levels, derived by us from two official tables. All three B2 texts were narrow passes, one is 140 words of the 250 asked, and the C1 text is a summary of documents.
- In test 2, the NCLC levels were assigned for teaching; the kit itself says its samples are not formal assessments. The two references don't fully agree: texts rated NCLC 5–6 there look like A2–B1 exam texts.
- 35 texts is a small sample, and most were handwritten and transcribed by hand.
- We chose the error thresholds (3 and 6) after looking at the NCLC kit texts, so test 2 flatters them; test 1 is the independent check. Error counting reads spelling and grammar, not ideas or organisation.
- Speaking isn't covered: we found no public speaking samples with official levels and transcripts.
- Run with deepseek-v4-pro; results can change when the model or our instructions change, so we re-run these tests and date them.
- Real results people share with us will also be skewed: those who did well are more willing to share. When we compare them, we'll use each person's most recent GoExam estimate in the 30 days before their test, decided now so the rule can't be picked after seeing the data.
Took the TEF? Help build the real benchmark
Tell us your official scores. With your permission, we compare them with your GoExam estimates and publish only totals here.
Share my TEF result