Evolution of automated feedback in SFL: Analysis of Claude Sonnet 3.7 and 4.0 as evaluators

Abstract

This research analyses the evolution of feedback capabilities between specific artificial intelligence for different Claude Sonnet language models in versions 3.7 and 4.0 as correction tools for texts written by students of Spanish as a foreign language. Through a qualitative comparative analysis of 15 texts from the CEDEL2 corpus, the study evaluates different items: accuracy in error detection, explanatory clarity, pedagogical appropriateness, and problems detected. Claude 4.0 increases error detection by 17% (189 vs 161) and develops greater sophistication in level adaptation, focusing on fundamental errors for beginners while providing comprehensive analysis for advanced students. The latest version introduces improvements in structural organisation through a tripartite format: ‘error → correction → explanation’. However, it presents worrying pedagogical setbacks: it eliminates complementary activities characteristic of Claude 3.7, degrades motivational feedback to generic third-person comments, and maintains biases towards peninsular varieties and hypercorrection. More problematic are the interlinguistic interferences generated by proposals in Spanish and English, resulting in inappropriate Spanglish. The analysis confirms that neither version can function autonomously without teacher mediation, establishing their optimal role as complementary tools with active pedagogical supervision. The findings show that technological evolution in educational AI does not constitute linear improvement, revealing complex exchanges between technical sophistication and pedagogical adequacy.