Outcome-only feedback can support problem solving when learners use correctness signals to regulate subsequent attempts. GenAI systems can extend this support by providing process guidance, but open-ended assistance may also encourage uncritical advice-taking and overreliance. This raises the question: can GenAI be constrained to cue strategy use while preserving learners’ regulatory control?

We addressed this question by comparing two supports for solving verbal analogies: automated outcome-only feedback (correct/incorrect) and constrained chatbot-based process scaffolding. Verbal analogies are well suited because they require relational mapping and verification. We compared these supports in adults with and without ADHD. Because ADHD is associated with weaker executive control and metacognitive regulation, structured external cues may be especially helpful, making ADHD a useful test case for agency-preserving GenAI support.

Native English-speaking adults with and without ADHD (N = 213) completed a 2 (Group: ADHD vs non-ADHD) × 2 (Support: outcome-only vs constrained chatbot) between-participants study. All participants learned a step-by-step analogy-solving strategy and restated worked examples to ensure engagement. Training included three novel analogies. In the outcome-only condition, participants solved problems unaided and received automated correct/incorrect feedback. In the chatbot condition, participants solved analogies with a constrained chatbot that guided strategy execution through fixed multi-turn prompts with progressively specific hints, without providing the correct answer. After each training item, all participants viewed the correct answer and wrote a brief justification, equating answer exposure across conditions.

After training, participants completed a 15-item unaided test with confidence ratings; response times were used to estimate efficiency. Test accuracy was higher in the chatbot condition for both groups, with no differences in efficiency. Chatbot support also increased confidence without undermining metacognitive monitoring: confidence-accuracy alignment and discrimination between correct and incorrect responses. Overall, the findings support a human-GenAI collaboration approach in which constrained process scaffolding improves transfer while preserving monitoring accuracy.