Corpus-informed AI-Likeness Screening for L2 Writing: A Proof-of-Concept Field Test and Two Flagged Cases

Generative AI has expanded the range of support available for L2 writing, including idea generation, language revision, feedback, and grammatical assistance, but it has also intensified concern about outsourcing in unsupervised assignments. Existing AI detectors show uneven accuracy, vulnerability to circumvention, and fairness risks in multilingual settings. In response, this proof-of-concept field test examined AI-Likeness Inspector, a teacher-facing prototype that uses corpus-informed, interpretable linguistic indices for low-stakes screening. The prototype operationalizes indices and the fixed reference threshold from Mizumoto et al. (2024), including an errors-per-100-words threshold of 3.44, and applies them to task-matched L2 writing. Under the same essay prompt and task conditions as Mizumoto et al., 142 first-year university students submitted 200–300-word argumentative essays online. Two submissions (1.4%) fell below the fixed error-rate threshold and were treated as cases for follow-up rather than as proof of AI authorship. Follow-up interviews with both students indicated that each had used ChatGPT to draft the essay and had prompted the model to include deliberate errors to appear more human. The findings therefore support a narrow conclusion: transparent, task-specific screening can help identify atypical submissions for further checking, but this study does not estimate sensitivity, specificity, or false negative rates and should not be read as validation of a general-purpose detector.