Clue before correction: ChatGPT-enhanced strategy for promoting autonomous and reflective language learning

Synthesis: Lukešová and Jennings (2026) present a generative-AI-supported revision task that promotes learner autonomy and metacognitive engagement in L2 writing through a "clue before correction" strategy: learners receive guided hints and must infer correct solutions rather than being given direct error correction by ChatGPT. With 58 university students (CEFR A1–B1), they found ChatGPT provided structured, adaptive feedback that reduced cognitive load and supported personalized revision — a design that operationalizes learning-from-errors and productive-failure principles with AI.

Core design: clue before correction

Unlike traditional chatbot use for direct error correction (which positions learners as passive recipients), this activity is intentionally structured to require active Problem Solving: students are prompted to infer correct solutions from guided hints. This is a concrete instance of learning from errors/mistakes — the learner is pushed to diagnose and correct their own errors with clue-based Scaffolding rather than being told the answer.

Findings

• Reduced cognitive load: ChatGPT provided structured, adaptive feedback that reduced cognitive load and supported personalized revision pathways, particularly for more proficient learners (those with sufficient prior knowledge).
• Prompting matters: learner prompting behavior strongly influenced feedback quality and depth — prompting is itself a skill learners develop over time, so outcomes vary with proficiency, prompting ability, and model interpretation.
• Learning-gains framing: the paper cites that elaborative Feedback produces significantly higher learning gains than verification-only feedback (Hattie & Timperley), and that feedback timing matters — supporting the value of clue-based (elaborative) over direct-corrective feedback.
• ICAP-ME framework: a new pedagogical framework extending ICAP to include metacognitive and affective dimensions, used to analyze AI-mediated learning engagement.

Design principles

The underlying, language-agnostic principles are error elicitation, clue-based feedback, and guided reasoning. The authors position generative AI as a formative learning partner rather than an "AI as corrector," whose effectiveness depends on the interaction of proficiency, task design, and the alignment of scaffolding with cognitive demands.

Relevance to the knowledge base

This paper is an empirical example of error-driven, autonomy-supporting AI pedagogy. It connects Productive Failure and learning-from-mistakes to Language Learning and Generative AI, showing how AI can scaffold learners to diagnose and correct their own errors (clue-before-correction) rather than passively receive fixes. It connects to Metacognition (reflective engagement), Self-Regulated Learning (autonomy), Cognitive Load Theory, and Feedback (formative, elaborative).

What this means for practice

• Instructors. Restructure AI revision so ChatGPT supplies clues rather than corrections, and require learners to infer and apply the fix themselves; the paper's language-agnostic design principles are error elicitation, clue-based feedback, and guided reasoning.
• Instructors. Teach prompting as a skill or embed prompt scaffolds in the task, because learner prompting behavior strongly influenced the quality and depth of ChatGPT's feedback — the 58 participants spanned CEFR A1–B1 and outcomes varied with proficiency and prompting ability.
• Instructors. Prefer elaborative, clue-based feedback over verification-only feedback, and time it deliberately: the paper cites evidence that elaborative feedback yields significantly higher learning gains than verification-only feedback and that feedback timing matters.
• Curriculum designers. Place clue-before-correction tasks where learners already hold enough prior knowledge to act on hints, since reduced cognitive load and personalized revision pathways were reported particularly for more proficient learners.
• Curriculum designers. Plan tasks against the ICAP-ME framework so that metacognitive and affective engagement are designed in alongside cognitive engagement, rather than treating AI-assisted revision as a cognitive exercise alone.

Limitations

• Fifty-eight university students at a single institution completed the task in English or Spanish; the evidence is learner–AI interaction transcripts plus open-ended reflections, with no control group or pre/post achievement measure reported, so effects on L2 proficiency are inferred rather than demonstrated.
• Benefits were uneven within the sample — reduced cognitive load and better revision pathways held particularly for more proficient learners, so the design's value for A1-level learners is not established by the aggregate findings.
• The activity used one model (ChatGPT) and two languages, and the authors caution that findings should not be assumed to generalize to languages with limited training data, non-Latin scripts, or complex morphology, calling for cross-lingual and cross-model research.

Citation

Lukešová, A., & Jennings, P. J. (2026). Clue before correction: ChatGPT-enhanced strategy for promoting autonomous and reflective language learning. Innovation in Language Learning and Teaching. DOI: 10.1080/17501229.2025.2612532.