Notation IA et commentaires pour les enseignants

AI can help with two different jobs: writing feedback and suggesting scores. Feedback drafting is the safer, more useful one. Given a clear rubric and a piece of student work, a model can produce specific comments tied to each criterion faster than most people can type them. Scoring is riskier because a number goes into the gradebook and affects students directly. Several accuracy problems are well known. Scores can change between runs on the same essay. Models can reward length and polished vocabulary over reasoning. They can be swayed by confident tone. And they can invent evidence, praising or criticizing sentences that aren't in the student's work. Structured prompts and teacher review catch most of this, but not if the teacher only skims. Fairness is a separate concern. Writing by English learners, or in dialects other than standard academic English, may be judged more harshly on surface features. AI writing detectors are a related trap. They are unreliable, and research has found that some detectors disproportionately flag non-native English writers. A detector score shouldn't be the basis for an accusation. Some grading tools use more limited AI. Gradescope, for example, can group similar answers so a teacher grades each group once. The teacher still decides the score, and the AI handles the sorting. Privacy applies as well. Student work is part of the education record. In the United States, FERPA governs how it is handled, so use district-approved tools and remove names where you can. The main misconception is that AI grading is objective because it is automated. It applies patterns learned from data, including that data's biases. The workable model is AI for drafts, teacher for decisions, and honesty with students about how feedback was produced. It works best for frequent, low-stakes formative feedback, not final grades.