TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—21 · FACTUAL · EVOLVING
The machine fixates too
A model has favourite answers, so real variety means changing the model or the brief.
The dilemma
You want more variety. The obvious move is to press generate again. Twenty re-rolls later the images are all cousins. Is the problem your luck, or the machine's range?
The choices
Keep re-rolling, hoping the next draw leaves the neighbourhood. Or accept that a model has habits of its own (favourite forms, favourite palettes, favourite ways to resolve a corner) and get variety by changing the model, reframing the brief, or prompting deliberately against the habit.
The consequence
Re-roll and you spend an evening mapping how narrow one model's range is, without realising that is what you did. Switch the model or the brief and the range opens up, because you have moved to a different set of habits instead of sampling the same set again.
The case
A student needs three genuinely different roof strategies for a market hall in Davangere. Thirty re-rolls of "market hall roof, hot dry climate" give thirty variations on a long-span shed with clerestories. A reframed brief, "three roof strategies: vaulted, folded plate, and a field of small pitched roofs over a grid of columns," gets three in one batch. A second model with the same reframed brief gets a different three. The model had a favourite answer. The student had been asking it the same question thirty times.
Try it
Re-roll one prompt twenty times and map how narrow the model's range really is: sort the twenty into families and count them. Then explain why switching models or reframing the brief beats re-rolling when you want real variety.
Take it to crit
When a student wants more variety, do they re-roll the same model, or change the model or the brief? Ask them to show the re-roll batch and point at what did not change across it.
How it works
Fixation, meaning a return to familiar solutions, was long studied as a human habit. Recent work finds the same pattern in generative models — measured so far in experimental studies of text and image models on idea-generation tasks, not in architecture. Repeated generations cluster around a small number of favoured forms, and the effect persists across prompts that a human would read as different. Work on idea lists from language models names two mechanisms: the model's own fixation around individual concepts, and the fact that it pools its knowledge into one average rather than holding many separate viewpoints. The same research shows that deliberate prompting strategies, such as asking for step-by-step reasoning first or assigning an ordinary persona, can widen the range substantially. That is the EVOLVING part. The narrowness is real and measured in those settings today, how wide one model's range is depends on the prompt, the settings and the model family, and the workarounds are improving. Test it rather than assume it.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.