At a glance
- What it is: From Fake Perfects to Conversational Imperfects: Exploring Image-Generative AI as a Boundary Object for Participatory Design of Public Spaces
- Who: Jose A. Guridi, Angel Hsing-Chi Hwang, Duarte Santo, Maria Goula, Cristóbal Cheyre, Lee Humphreys and Marco Rangel (Cornell University, USC, Studio-MLA and an independent researcher).
- Where: Proceedings of the ACM on Human-Computer Interaction (CSCW), April 2025. doi.org/10.1145/3710912
- Type: qualitative case study. Reflection workshops and nine AI-mediated interviews in a real participatory process.
First reading: what it does and what it finds
It helps to start with what kind of study this is. It is a qualitative case study, not an experiment that measures outcomes. The team partnered with a landscape architecture firm in Los Angeles, Studio MLA, to improve River Garden Park, and tested a concrete idea: using image-generative AI as a boundary object, that is, as a shared piece that helps people from different worlds understand each other. In the central stage they conducted nine interviews with residents, children of Latin American migrants between 18 and 35 years old. In each one, a facilitator asked the questions and an assistant generated images in real time from what the person described, mostly using Dream Studio. They do not measure whether AI improves the design, nor do they evaluate a model. They describe how the conversation changed when the image was generated live.
The starting point comes from prior literature, not from this study: designing a public space requires balancing very different interests, and participatory methods often fall short in incorporating the voices of the least represented groups. The paper’s own question is whether an AI that draws on the fly helps bridge that distance between those who design and those who will use the place.
The first finding, and the one that gives the title, is a distinction the authors build from their data. They call Fake Perfects the technically polished images, faithful to the requested text, that nonetheless shut the conversation down: the person would say “yes, that’s it,” and there was nothing more to talk about. And they call Conversational Imperfects the imperfect images, sometimes inaccurate or frankly strange, that opened the conversation up. Success was redefined. It stopped being about getting the right image and became about sustaining a good exchange.
Those imperfect images worked in two ways. They revealed hidden ideas: one participant could not name an element of her Mexican heritage until she saw a mural in an image and remembered her mother’s stories. And they inspired new ideas: for another, a fanciful image of a park made of mushrooms helped him think about designing with natural elements. What is interesting is that the inaccuracy was the engine, not the flaw.
The second finding is that AI fostered more spatially aware conversations. In front of an image, even a wrong one, participants described more precisely where each thing went, where they placed themselves, what atmosphere they wanted. Compared with earlier interviews without images, where people listed activities, here they pointed at the plan and made corrections. The image provided something concrete to deliberate over.
The third finding sets a limit on the previous two: the outcome depended heavily on the facilitator’s skill. An accurate image could close the conversation if no one knew how to ask follow-up questions; a strange image could be a dead end if the facilitator did not make use of it. Managing the interaction between the people, the AI and oneself is a skill, not something the tool delivers on its own. From there the authors derive an idea for the field: moving from methods centered on the final object to methods centered on the process, and designing tools that help moderate, not just generate.
Second reading: from Latin America
Where you look from matters. The participants are children of Latin American migrants, mostly from Mexico and also from Guatemala, but the study takes place in Los Angeles, not in the region. It is the diaspora. That is why the finding that resonates most from here is one the authors themselves flag as a tension: the images tended to look idealized and “American,” and it was hard to represent the participants’ cultural heritage beyond flags and colors. Stereotypes appear. The paper notes one with unintended humor: a participant associates his idea of Mexico with the movie Encanto, which is actually inspired by Colombia. Not only does AI fail to correct that flattening, it sometimes reinforces it.
That turns a limitation into a tool. The authors suggest, and here I share the view, that the biased image can serve to talk about the bias itself: naming what is missing, what is superfluous, what does not look like oneself. For participatory processes in Latin America, where cultural representation is delicate and models were trained mostly on data from the North, that reading is valuable. But it is my extrapolation: the study was done with the diaspora in the United States and did not test anything within the region. I leave it as a hypothesis, not a finding.
Where the study connects with the present is in the change of era. The interviews are from 2024, with tools that still made quite a few mistakes. Today image models produce much more faithful and eye-catching results. The paper leaves an uncomfortable warning in the face of that: if what opened the conversation was imperfection, an AI that gets better and better at getting it right could, unintentionally, close it sooner. The authors themselves point in that direction when they ask that tools keep the possibility of introducing randomness and give the user simple control over how “exact” they want the image to be. Taking that warning as far as saying that the new models harm participation is my reading, not the paper’s.
The idea that travels furthest is one the authors formulate almost in passing: since neither the designer nor the resident is an AI expert, the tool puts them on a more level footing. In contexts with strong asymmetries of power and technical knowledge, that leveling, even if momentary, is something worth protecting.
The question it leaves works for any municipality in the region that wants to listen to its people before redoing a public square. What are we looking for when we show the community an image: for it to approve a drawing, or for a conversation to open up? The study suggests that confusing the first with the second is the easiest mistake to make.
The fine print
- It is a qualitative case study: its strength lies in explaining how and why the conversation changes, not in measuring how often or proving that AI improves the design. It is useful for understanding a phenomenon, not for generalizing frequencies.
- The work was done with a specific community in Los Angeles, with children of migrants, and the authors themselves acknowledge that the context limits the results and that the format of one facilitator per interview limited the scope.
- It portrays image tools from 2024. The same authors warn that their evolution may change details; what seems to age slowly is the underlying point, about conversation, power and representation.
- Transparency: the author of this blog is a coauthor of the paper.
Paper keywords: Artificial Intelligence, generative AI, human-AI interaction, design
Automated reading. This text was generated by Claude, an Anthropic model, from the original source, without line-by-line human review. It may contain errors or debatable interpretations; to check any point, see the original source.