Designing AI Systems to Support Productive-Failure-Based Learning

Synthesis: Kim, Lin, Yu and Detrick (2026) investigate how adult learners perceive AI applications for supporting Productive Failure-based learning and what design principles AI systems should follow. Through focus groups with 35 graduate students who developed AI application storyboards and paper prototypes, they map AI support onto the phases of productive failure (generation & exploration; consolidation & knowledge assembly) and derive five design principles: human-AI collaboration, usability, reflective design, emotional design, and open knowledge/resource utilization. The core message: AI should preserve productive struggle while offering targeted, non-directive support — not substitute for it.

Core Finding

Generative AI can support productive-failure-based learning across problem exploration, solution generation, collaborative problem-solving, comparison/contrast, knowledge reorganization, and knowledge transfer — provided it is designed to preserve struggle and give non-directive help. Five design principles emerge for AI systems serving adult learners engaged in productive failure.

The productive failure framework

Productive Failure (PF), grounded in constructivist principles, engages learners with problems targeting concepts they have not yet learned, struggling to generate solutions BEFORE receiving direct instruction (Kapur, 2008; Kapur & Bielaczyc, 2012). Because learners generate solutions without cognitive support, they rely on prior knowledge and produce suboptimal or incorrect solutions. The initial struggle and failure is a powerful catalyst: it activates and differentiates prior knowledge and prepares learners to learn better from subsequent instruction. PF benefits include enhanced knowledge transfer, durable skills (Critical Thinking, resilience, communication), reduced fear of mistakes, increased engagement, and positive attitudes toward lifelong learning.

How AI supports each PF phase (adult-learner perceptions)

Phase 1 — Generation & exploration:

• Problem exploration without indirect instruction: AI (conversational agents like ChatGPT, Gemini, Claude) as "thinking partners" prompting learners to define success/failure conditions, adopt multiple perspectives on failure, and trace root causes of ill-structured problems. Chatbots can generate failure-based quiz questions to surface misconceptions in a safe, low-stakes space before formal instruction.
• Solution generation via prior-knowledge activation: AI-powered brainstorming, scenario-planning/"what-if" analysis, graphic organizers/mind maps, prototyping tools, and Socratic-style questioning that connects prior experience to new contexts.
• Collaborative problem-solving: AI facilitating equitable group engagement through collaborative workspaces (Notion, Miro, Trello).

Phase 2 — Consolidation & knowledge assembly: AI supporting comparison and contrast, knowledge reorganization, and knowledge transfer across contexts.

Five design principles for AI supporting PF

1. Human-AI collaboration — complementary roles; interpretability; human-in-the-loop integrating learners' and instructors' judgement across the AI lifecycle; feedback loops. Rather than AI giving a final answer, use AI to help learners identify areas for improvement and make final decisions themselves.
2. Usability — accessibility/inclusivity; graceful error handling and ambiguity resolution (ask clarifying questions, admit uncertainty rather than fabricate); visual cues distinguishing learner input from AI contribution; controllability (opt in/out); data interoperability; seamless AI integration into the existing tech stack (LMS, VR/AR labs).
3. Reflective design — AI that reflects on its own errors and self-improves: correction mechanisms, follow-up questions, context-aware prompts, error tracking and prevention, continuous learning loops.
4. Emotional design — AI personality and conversation design that support learners' emotional experience during struggle.
5. Open knowledge and resource utilization — shifting from pre-trained data to accessing/integrating real-time external information; breaking down data silos.

Relevance to the wiki

This paper anchors the wiki's Productive Failure concept with an AI-specific lens: how generative AI can be designed to support (not short-circuit) productive-failure pedagogy. It connects to Cognitive Offloading (the risk that AI erases the struggle), Scaffolding (non-directive support that preserves struggle), Socratic Method (questioning to activate reasoning), Feedback (non-directive loops), and the broader design question of AI as "thinking partner" rather than answer-giver. Its adult-learner focus connects to Adult Learning and andragogy principles.

Connected Concepts

• Productive Failure
• Generative AI
• LLM
• Adult Learning
• Human In The Loop AI
• Cognitive Offloading
• Scaffolding
• Socratic Method
• Feedback
• Prior Knowledge
• Learning Theories
• Higher Ed
• Transfer Of Learning

Connected Articles

• Puech Pedagogical Steering LLM Productive Failure 2025 — Pedagogical Steering of LLMs for Productive Failure
• Rhaimi Productivemath 2025 — ProductiveMath: AI to Support PF Problem Design
• Wang Safety Gap Productive Struggle 2026 — The Safety Gap: Restoring Productive Struggle
• Lukesova Clue Before Correction 2026 — Clue Before Correction: ChatGPT for Autonomous Learning
• Principled AI Education — Principled AI in Education

Citation

Kim, J., Lin, X., Yu, S., & Detrick, R. (2026). Designing AI systems to support a productive-failure-based learning: insights from adult learners on AI applications and AI system design principles. Educational Technology Research & Development. DOI: 10.1007/s11423-026-10655-6.