Dissonance Dosing: An affect- and proficiency-aware AI feedback policy that withholds closure to strengthen human critical reasoning

This research proposes a novel AI feedback policy called Dissonance Dosing that strategically withholds closure to provoke deeper reflection and critical reasoning. Inspired by Cognitive Dissonance AI, the system adaptively controls the "dissonance dose" by monitoring user proficiency and affect through two personalized controllers: a proficiency-aware controller that adjusts friction intensity based on user skill level and argumentation quality, and an affect-aware controller that maintains users within a safe emotional zone using lightweight, privacy-preserving proxies and optional self-reports. The policy learner decides when to present counter-arguments, withhold direct answers until justification is provided, or require critique of AI proposals before revealing evidence. Learning is driven by reinforcement learning from AI and human feedback, optimizing a human-centered "reflection reward" that includes argumentation quality gains, reduced cognitive offloading, and safe affect metrics. The policy is encoded as auditable instructions and nudging templates that self-update with user ratings and are periodically distilled into model weights. This approach challenges the assumption that more guidance is always better, aiming instead to foster deeper learning through calibrated contradiction and uncertainty. The research includes multi-domain evaluation in academic writing, clinical reasoning, and public deliberation, using randomized trials and mixed methods to measure argument quality, affect trajectories, openness to opposing evidence, and polarization drift. The expected impact is a new design principle for AI-augmented reasoning that optimizes human reflection quality, reduces cognitive offloading, increases argumentative rigor, and maintains psychological safety.

References:

1. PokeeResearch: Effective Deep Research via Reinforcement Learning from AI Feedback and Robust Reasoning Scaffold. Yi Wan, Jiuqi Wang, Liam Li, Jinsong Liu, Ruihao Zhu, Zheqing Zhu (2025).
2. Human-AI Interaction for Augmented Reasoning: Improving Human Reflective and Critical Thinking with Artificial Intelligence. Valdemar Danry, Pat Pataranutaporn, Christopher Cui, Jui-Tse Hung, Lancelot Blanchard, Zana Buçinca, Chenhao Tan, Thad Starner, Pattie Maes (2025). CHI Extended Abstracts.
3. Cognitive Dissonance Artificial Intelligence (CD-AI): The Mind at War with Itself. Harnessing Discomfort to Sharpen Critical Thinking. Delia Deliu (2025). arXiv.org.
4. Generative AI for medical education: Insights from a case study with medical students and an AI tutor for clinical reasoning. Amy Wang, Roma Ruparel, Anna Iurchenko, Paul Jhun, Julie Anne Séguin, Patricia Strachan, Renee Wong, A. Karthikesalingam, Yossi Matias, A. Hassidim, Dale R. Webster, Christopher Semturs, Jonathan Krause, Mike Schaekermann (2025). CHI Extended Abstracts.