Designing cognitive friction with concept suppression in large language models

Large Language Models (LLMs) are increasingly deployed in tools for thought and creativity support systems, yet their generative fluency can collapse early-stage ideation into selection and curation, shifting cognitive labor away from concept formation and divergent thinking toward homogenized outputs. We introduce concept suppression as a subtractive approach to induce cognitive friction, treating the learned conceptual manifold of a pre-trained model as a design surface for human–AI interaction. We operationalize this with Unlearning to Rest, a prototype that applies weight-level concept suppression to Llama3.2:3b, suppressing a canonical attractor concept (“chair”) to create navigational impediments in the solution space. We evaluate the approach in a within-subject study (N = 37) where participants were set a conceptual ideation challenge. Quantitative results show a consistent workload–ownership trade-off: the unmodified model reduces mental demand and effort but also reduces perceived ownership, while Unlearning to Rest yields ratings closer to unaided ideation. Qualitative analysis indicates that the constrained model prompts user re-articulation, supporting a stage-fit workflow in which constrained assistance benefits early ideation. We contribute (1) representational intervention as a lens for LLM-based tools for thought, (2) a system instantiation via weight-level concept suppression, and (3) empirical evidence motivating stage-sensitive evaluation of LLM assistance beyond efficiency and output quality.