In short: AI can give a first reading of a student's writing in minutes. Research on essays finds it often lands close to a trained human grader, but rarely on exactly the same score, which is why researchers keep it for low-stakes uses. For a language teacher, the useful question is not "AI or teacher" but which decision each one should take. CEFRHub places a written or spoken production on the CEFR scale, with a report in under 3 minutes; it is not a certifying test, and the teacher keeps the final judgment.
What the Research Says About AI Essay Grading
The Hechinger Report described a study led by Tamara Tate, a researcher at the University of California, Irvine (Hechinger Report, May 20, 2024). Her team compared ChatGPT with trained human graders on history and English essays written by middle and high school students, in three batches (943, 344 and 493 essays), scored on a 6-point scale:
- ChatGPT's score was within one point of the human score 89% of the time in the batch of 943 essays, 83% in the 344 English essays and 76% in the 493 history essays;
- it gave exactly the same score only about 40% of the time; two trained human graders agreed exactly about half the time;
- it tended to cluster its scores in the middle of the scale, where humans gave more top and bottom scores.
The researcher's conclusion: accurate enough for low-stakes uses, such as a preliminary score on a first draft, but not for a high-stakes test or a grade that counts. At the time, the paper was still under peer review. These were essays written in the students' first language, not productions in a second language assessed against the CEFR.
Her advice to schools comparing AI graders: ask how often the AI gives exactly the same score as a human on each essay, and how often it is within one point.
Where Human Grading Remains the Reference
- Official exams. At IELTS, writing and speaking are marked by examiners against published criteria (IELTS scoring in detail); see our CEFRHub vs IELTS comparison. DELF and DALF results are issued after validation by the juries; see DELF vs DALF.
- Teaching decisions. In the same article, the researcher warns that teachers who delegate too much grading lose sight of their students' recurring mistakes, which tell them what to teach next.
- High-stakes decisions. A grade that counts, an admission, a certificate: this is where the research and the official exams keep a trained human in charge.
Human grading is not perfectly consistent either: in the study above, two trained graders gave exactly the same score only about half the time. Clear criteria matter, whoever grades.
Where AI Helps a Language Teacher
- A first reading in minutes, to free time for feedback and teaching.
- Low-stakes decisions: placing a learner in a group, a score on a first draft, tracking progress between two assessments.
- The same published method for every production, which makes it easier to compare a learner with themselves over time.
For a team or a school, our article on AI language proficiency testing for institutions covers what to ask a provider.
How CEFRHub Assesses a Production
CEFRHub assesses written and spoken production, in English and French: a CEFR level from A1 to C2, with sub-levels, and a report based on the descriptors of the CEFR and its Companion Volume, in under 3 minutes. To understand each level, see the guide to CEFR language levels and CEFR B2 vs C1: the differences.
The methodology publishes three safeguards against inflated levels:
- without at least two C1 markers attested in the production, the level is capped at B2.2;
- strong spelling or pronunciation never on its own justifies a move to C1: this is the CEFR's "range over accuracy" principle;
- each level boundary carries its own exclusion checklist, to be checked before the level is awarded.
For speaking, the learner records with the microphone in the browser, or uploads an audio file in MP3 or WAV format.
What CEFRHub Does Not Do
- CEFRHub is not a certifying test. When an official result is required, the learner needs an official exam.
- CEFRHub does not take the final decision. Like the AI graders in the study, it is a tool for a first reading and for low-stakes decisions; the teacher reviews what counts.
- CEFRHub does not assess comprehension. It assesses written and spoken production.
Frequently Asked Questions
Is AI good at grading essays?
Often close, rarely identical. In the study reported by the Hechinger Report, ChatGPT was within one point of a trained human grader 76% to 89% of the time, depending on the batch, but gave exactly the same score only about 40% of the time (Hechinger Report).
Does AI grading really work?
For a first reading and low-stakes decisions, the research is encouraging. For a high-stakes test or a grade that counts, the researcher considered it not accurate enough yet: a teacher's review remains necessary.
Can AI replace a language teacher's grading?
No. AI can take on a first reading and low-stakes decisions; the teacher keeps the decisions that count, and the view of recurring mistakes that guides teaching.
Do human graders always agree with each other?
No. In the same study, two trained human graders gave exactly the same score about half the time. Clear, shared criteria reduce the gap, whoever grades.
What should I ask a provider of AI essay grading?
How often the AI gives exactly the same score as a trained human on each essay, and how often it is within one point; ask also which criteria it uses and whether they are published.
How does CEFRHub assess a production?
Against the descriptors of the CEFR and its Companion Volume, with three published safeguards against inflated levels (methodology). The report gives a CEFR level from A1 to C2, with sub-levels, in under 3 minutes.
Is CEFRHub a certifying test?
No. CEFRHub is an educational assessment tool. When an official result is required, the learner needs an official exam such as IELTS, DELF or DALF.
Does CEFRHub assess French as well as English?
Yes. CEFRHub assesses written and spoken production in English and in French.
Can CEFRHub assess speaking?
Yes. The learner records with the microphone in the browser, or uploads an audio file in MP3 or WAV format.