CMU researchers' AI agent uses constructive conflict to improve novice design

Carnegie Mellon researchers Howard Ziyu Han and Nikolas Martelaro are flipping the script on AI design tools. Most generative AI agents help designers by adding ideas and expanding the design space. Their new system, inspired by adversarial design theory, instead plays devil's advocate — synthesizing pushback from stakeholders like clients, users, or manufacturers to force designers to reconsider their decisions.

In a between-subjects experiment with 45 novice design students, the team compared three conditions: unsupported self-reflection, written stepwise prompts using a constructive-conflict framework, and an interactive AI agent enacting that same framework dynamically. Both the stepwise and interactive groups reported significantly higher self-reconsideration and made more concrete improvements to their proposals than the self-reflection group. But the antagonistic agent stood out: it introduced more conflictual perspectives, and participants in the interactive condition generated and discarded significantly more ideas — indicating deeper exploration of divergent stakeholder needs.