Promoting Critical Thinking With Domain-Specific Generative AI Provocations

The evidence on the effects of generative AI (GenAI) on critical thinking is mixed, with studies suggesting both potential harms and benefits depending on its implementation. Some argue that AI-driven provocations, such as questions asking for human clarification and justification, are beneficial for eliciting critical thinking. Drawing on our experience designing and evaluating two GenAI-powered tools for knowledge work, ArtBot in the domain of fine art interpretation and Privy in the domain of AI privacy, we reflect on how design decisions shape the form and effectiveness of such provocations. Our observations and user feedback suggest that domain-specific provocations, implemented through productive friction and interactions that depend on user contribution, can meaningfully support critical thinking. We present participant experiences with both prototypes and discuss how supporting critical thinking may require moving beyond static provocations toward approaches that adapt to user preferences and levels of expertise.

💡 Research Summary

The paper investigates how generative AI (GenAI) can be deliberately designed to foster critical thinking rather than merely automate tasks. Drawing on two domain‑specific prototypes—ArtBot for fine‑art interpretation and Privy for AI‑privacy risk planning—the authors examine the role of “provocations” (strategic questions that introduce productive friction) in prompting users to reflect, justify, and revise their reasoning.

ArtBot is built on locally hosted Llama 3 models with Retrieval‑Augmented Generation (RAG) that accesses a curated corpus of art‑historical metadata, curatorial texts, and educational resources. Its interaction follows a Socratic tutoring style: the system asks domain‑grounded questions such as “Would your interpretation change if you knew the work was created during a revolution?” Users must articulate their own interpretations before the AI offers further prompts. In a user study (n = 13), participants produced short reflective writings after each artwork, showing deeper interpretive engagement compared with a non‑AI baseline.

Privy supports AI practitioners during early design phases by guiding them through a structured, tree‑shaped workflow embedded in a whiteboard‑style interface. Powered by GPT‑4.1 and informed by established privacy taxonomies (Lee et al., Das et al.), the system asks targeted questions like “How can you design this feature to encourage users to regularly review and update their sharing settings?” Users must assess risk relevance, severity, and propose mitigation strategies before advancing. Twelve practitioners generated privacy‑risk documents that were evaluated by experts; the documents demonstrated higher completeness and quality than those produced without AI assistance.

Both systems share three core design principles: (1) Productive friction – the AI resists providing direct answers, instead requiring users to externalize their reasoning first; (2) Domain‑specific provocation – prompts are grounded in domain knowledge, terminology, and normative frameworks, making them meaningful and actionable; (3) Facilitator framing – the AI is presented as a thinking partner rather than an answer engine. The “user‑generated content gate” forces participants to type substantive responses before receiving further AI support, operationalizing a human‑in‑the‑loop approach.

Qualitative observations reveal a mixed affective response. Some participants experienced mild frustration when the system slowed progress, reflecting expectations shaped by conventional answer‑oriented chatbots. Others reported surprise and engagement, noting that the questions surfaced considerations they had not previously entertained. Importantly, reactions varied with users’ domain expertise: experts tended to challenge the AI’s suggestions and appreciated the nuanced provocation, while novices often accepted the AI’s guidance more readily. This suggests that provocation strength and style should be adaptable to individual expertise and tolerance for friction.

The authors conclude that (a) domain‑specific provocation yields richer, more aligned user output than generic “Why?” prompts; (b) productive friction can successfully shift GenAI from automation toward a catalyst for metacognitive activity; and (c) future designs must incorporate adaptive mechanisms that tailor provocation intensity to user preferences, expertise, and mental models of AI. They propose further work on dynamic provocation models, longitudinal studies of learning outcomes, and integration of user‑controlled customization to balance challenge with usability.