Divergence on Demand
First week of term. The brief: a lamp for a mountain hut. On Thursday, each student brings sixty images. Warm wood, brass, snow outside the window, an awful lot of evening glow. The first question in the crit is no longer “Why this one?” but “Which prompt?” This is not an anecdote about laziness but a shift in what the studio talks about.
The friendliest reading comes from Ma, Cai and Chen (2026). They analysed the portfolios of 35 third-year interaction design students in New Zealand. For two weeks, they had to use generative tools at every stage of design thinking for a personal brand mark, and document the process. The study identifies a recurring loop: generate, compare, revise. AI appears in it not as an external aid but as “generative mediation” that takes part in judgement. Their framework assumes that evaluations otherwise made in the head become partly visible in selection and rejection.
The fine print is more revealing. Around 28 per cent treated the outputs as provisional endpoints across several rounds of prompting rather than as starting points (p. 22). For one student, the images kept reverting to stylised templates despite repeated refinement. At such moments, several put the machine aside and worked by hand. The authors themselves warn of “a potential risk of attenuated judgment”, in which “design judgment becomes progressively substituted by the fluency and surface plausibility of algorithmic language” (p. 23). Admittedly, more than half named this tendency themselves, and in the prototype stage, according to the instructors, many increasingly asked “Does this result support my judgment?” rather than whether the output was “good enough”. But the data come from portfolios that were part of the coursework, and one of the researchers was on the teaching team. A well-documented loop is not yet good judgement.
The experimental picture is less charming. Wadinambiarachchi and colleagues (2024) had 60 people sketch a robot avatar in three groups: no support, Google image search, image generator. On fixation, the result is clear: the AI group copied more features of the example than the other two groups. The generator brought no gain in ideas, variety or originality; if anything, scores were lower. There is also what the team calls “fixation displacement”: fixation does not disappear but migrates from the example to the generated images.
The basic pattern is old. Jansson and Smith (1991, pp. 8–9) gave final-year mechanical engineering students an example coffee cup that relied on a straw and a mouthpiece—and leaked. The brief explicitly asked for designs “to have no straws or mouthpieces”. The group with the example drew more of them anyway, and more cups that leaked. What’s new is that the example now arrives on demand, four at a time. Chulvi and colleagues (2026) tested the four-pack. Given the same lunch-box brief as twenty students, six image generators showed markedly less variety; novelty and quality did not differ significantly.
It gets less comfortable at group level. Doshi and Hauser (2024) found that short stories written with AI ideas were rated more creative individually but resembled one another more. Anderson, Shah and Kreminski (2024) compared ChatGPT with a deck of cards from 1975, Brian Eno and Peter Schmidt’s Oblique Strategies. Individually, both were similarly divergent. As a group, ChatGPT users ended up closer together and felt less responsible for their ideas. In a studio, twenty students would each deliver more while the class as a whole narrows. Individual divergence, cohort convergence.
Which leaves critique. Schmitt-Fumian, Tauscher and Thoring (2025) had 25 students compare feedback from GPT-4 and from lecturers. In the students’ eyes, the machine did well: better structured, more encouraging, without an edge. Yet the authors note the model’s “propensity to soften its replies when confronted with contradiction”. For critique that genuinely challenges assumptions, lecturers remain essential. Park (2025) describes how critique itself changes in two master’s studios. It turns transactional and fixates on surfaces. Processes become opaque, students work in parallel rather than together, and a new hierarchy forms within teams, sorted by AI literacy.
Schön (1983, pp. 78–79) described designing as a conversation with the situation: “the designer’s moves tend, happily or unhappily, to produce consequences other than those intended.” The situation “talks back”, and the designer responds with the next move. Goldschmidt (1991), studying architects as they sketched, showed that a sketch does not depict a finished image held in the mind. It creates displays that call up new images, oscillating between two modes of argument. The rendered image talks back too, but in finished form. It leaves little room for the unintended, which design education has lived on for decades. The productive misunderstanding gets polished away.
None of this calls for a ban, only for a different object of assessment. What counts is not the best image but the best comparison. Show the rejected versions and say why they died. Write down the criterion before the first prompt, so that language comes before the spec rather than afterwards as an excuse. Set a boundary early—a material, a dimension, a site—and let the machine push against it instead of roaming open country. And defend one version against the model’s praise. Only then has judgement been exercised rather than merely shopped for.
Divergence was never the number of versions but the distance between them. And judgement does not begin with generating; it begins where you throw a version away and can say why.
Sources
Anderson, B. R., Shah, J. H., & Kreminski, M. (2024). Homogenization effects of large language models on human creative ideation. In Creativity and Cognition (C&C ’24) (pp. 413–425). ACM. https://doi.org/10.1145/3635636.3656204 (full text)
Chulvi, V., Ruiz-Pastor, L., Berni, A., Royo, M., & Castillo-López, A. (2026). Creative potential of image-generative AI models for conceptual engineering design tasks. Artificial Intelligence for Engineering Design, Analysis and Manufacturing, 40, e8. https://doi.org/10.1017/S0890060426100286 (full text)
Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290 (full text)
Goldschmidt, G. (1991). The dialectics of sketching. Creativity Research Journal, 4(2), 123–143. https://doi.org/10.1080/10400419109534381 (abstract only)
Jansson, D. G., & Smith, S. M. (1991). Design fixation. Design Studies, 12(1), 3–11. https://doi.org/10.1016/0142-694X(91)90003-F (full text)
Ma, Y., Cai, Y., & Chen, J. (2026). AI-assisted divergent pedagogy: Understanding human-AI co-creativity in design thinking education. Design Studies, 105, 101407. https://doi.org/10.1016/j.destud.2026.101407 (full text)
Park, H. (2025). Generative AI in studio-based design education: Human–AI relationships, collaboration and assessment. In IASDR 2025: Design Next. https://doi.org/10.21606/iasdr.2025.845 (full text)
Schmitt-Fumian, T. K., Tauscher, S., & Thoring, K. (2025). AI vs. human: Exploring the potential of generative AI as a feedback tool to support ideation in design education. Proceedings of the Design Society, 5, 429–438. https://doi.org/10.1017/pds.2025.10057 (full text)
Schön, D. A. (1983). The reflective practitioner: How professionals think in action. Basic Books. (full text)
Wadinambiarachchi, S., Kelly, R. M., Pareek, S., Zhou, Q., & Velloso, E. (2024). The effects of generative AI on design fixation and divergent thinking. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI ’24) (pp. 1–18). ACM. https://doi.org/10.1145/3613904.3642919 (full text)
