Abstract
Design fixation - the unconscious adherence to familiar design patterns that limit creative exploration - remains a challenge for creative practitioners despite advances in the development of creativity support tools. While recent work in design research has explored how large language models can augment human creativity, evidence suggests these systems may exacerbate fixation. We propose machine unlearning as a novel approach to mitigating fixation in human-AI creative collaboration. Unlike fine-tuning methods that expand a model's knowledge base, unlearning removes specific concepts to create productive gaps in the model's representational space. Responding to findings that semantic constraints on prompt composition alleviate fixation, we apply this method in our pilot study presenting a modified large language model in which the concept of 'the chair' has been strategically removed. We find that interacting with this model forces users to re-articulate design problems in novel ways, preventing convergence on familiar directions during ideation.
Keywords
Design Fixation, Generative AI, Unlearning
DOI
https://doi.org/10.21606/drs.2026.2424
Citation
Disley, M., and Khan, M. (2026) Unlearning to Rest: Machine unlearning as a method of mitigating design fixation in human-AI creative collaboration, in Simeone, L., Gray, C. M., Verhoeven, A., de Götzen, A., Bakırlıoğlu, Y., Zohar, H., Stead, M., and Buwert, P. (eds.), DRS2026: Edinburgh, 8–12 June, Edinburgh, United Kingdom. https://doi.org/10.21606/drs.2026.2424
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License
Included in
Unlearning to Rest: Machine unlearning as a method of mitigating design fixation in human-AI creative collaboration
Design fixation - the unconscious adherence to familiar design patterns that limit creative exploration - remains a challenge for creative practitioners despite advances in the development of creativity support tools. While recent work in design research has explored how large language models can augment human creativity, evidence suggests these systems may exacerbate fixation. We propose machine unlearning as a novel approach to mitigating fixation in human-AI creative collaboration. Unlike fine-tuning methods that expand a model's knowledge base, unlearning removes specific concepts to create productive gaps in the model's representational space. Responding to findings that semantic constraints on prompt composition alleviate fixation, we apply this method in our pilot study presenting a modified large language model in which the concept of 'the chair' has been strategically removed. We find that interacting with this model forces users to re-articulate design problems in novel ways, preventing convergence on familiar directions during ideation.