Intelligenza Artificiale per la valutazione linguistica nell’ambito degli esami di certificazione Cambridge

Caso studio del B1 Preliminary

Abstract

Cambridge Assessment English sets global standards in language tests, strictly tying exams to
the CEFR framework, yet the rapid rise of AI poses new paths for test grading. While existing
literature covers proprietary automated assessment systems, there is a gap regarding the
interactional evaluation limits of off-the-shelf generative AIs. This study aims to fill this gap by exploring how commercial AIs, specifically NotebookLM and Gemini Pro, assess productive skills, comparing their outputs to human examiners’ comments in a blind study of ten B1-level Italian learners. Data suggest AI limitations in the tested models when grading candidates: in the Writing test, they demand higher-level structures, failing to stick to strict CEFR frameworks. Challenges emerge in the Speaking test, where the selected AIs relied on text transcriptions, struggling to process prosodic features like pitch contour or speech rate, missing key facts of
human interaction, and prioritizing pure grammar over it. All factors considered, AI is a highly efficient tool to track class progress regularly, but high-stakes formal exams must adhere to a human-in-the-loop methodology, for only real experts can currently grasp the deepest pragmatic
traits of speech and fix AI’s “blindness”, ensuring a fair and valid test.