Prompting Sustainability
Abstract
“Generative AI is more than a visualization tool—it is also a mirror of its training data. By asking AI to create “more sustainable” streets, we can uncover the hidden visual assumptions embedded within foundation models. Rather than evaluating urban sustainability itself, this project examines how AI interprets it.”
Motivations
- Sustainability is contested and context-dependent — a streetscape considered sustainable in one city may be inappropriate or counterproductive in another, yet AI tools apply a single learned notion of the term regardless of setting.
- Generative AI is entering planning workflows at scale — architects and planners increasingly use text-to-image models to visualize design ideas, but when sustainability is executed as a generic transformation rule, it risks collapsing into a narrow set of recurring features (greenery, bike lanes, widened sidewalks) replicated everywhere regardless of fit.
- The open question is whether this is genuine improvement or quiet standardization — do AI-generated outputs reflect context-sensitive design thinking, or do they homogenize what “sustainable” is allowed to look like across very different urban settings?
What we did
Asking AI what a “sustainable streetscape” is
We began with a simple question: if you ask a generative AI model what a “sustainable streetscape” looks like, what does it actually produce — and how does that compare to the real street it started from?
Does it come out the same?
We then varied the prompt itself — testing whether small changes in phrasing shift what the model considers “sustainable.”
How do we know it’s different?
To quantify the difference, we ran semantic segmentation on both the raw and AI-generated images, then compared the resulting label distributions.
Key findings
- A shared template absorbs many typologies — rather than each city’s streets evolving along their own path, a wide variety of raw streetscapes converge on one dominant outcome once transformed: a vegetated, multimodal movement corridor.
- Homogenization crosses city lines, not just within them — Jakarta, Melbourne, and Singapore become harder to tell apart from each other after transformation, not merely more visually consistent on their own.
- The bias is structural, not incidental — rewording the prompt changes which feature gets amplified (vegetation, bike lanes, building frontage), but doesn’t stop the underlying push toward inflation.
- Composition metrics don’t catch everything — a closer, qualitative look at the generated images turns up real design problems that segmentation numbers miss: closed access points, illogical intersection redesigns, and mobility arrangements that read as unsafe.
- Prompt specificity narrows the bias without removing it — prompts that name concrete design categories produce more balanced, integrated streetscapes, while generic prompts default to the most visually familiar “sustainability markers.”
Why this matters
None of this means AI-generated greenery is inherently the wrong instinct — more vegetation, calmer streets, and better cycling infrastructure are broadly in line with real sustainable design practice. The issue is how uniformly it’s applied: the model pushes the same recipe onto every street at roughly the same intensity, regardless of that street’s starting character or which city it’s in, whereas real design practice adapts similar ingredients to local context and function. Practically, this suggests generative AI is well suited to early-stage scenario exploration and communicating design ideas quickly, but isn’t yet a reliable stand-in for an actual design process.
Limitations & future work
- The cross-city study covers three cities and a single generation pipeline; broader geographic and cultural variation, and other model families, remain untested.
- The prompt-specificity study is limited to Jakarta scenes, so its results describe model behavior within that specific urban reference frame rather than a universal rule.
- Closing the gap identified here will likely need structural conditioning to keep street geometry and circulation logic intact, locally calibrated generation in place of one global aesthetic, and human or expert validation to confirm outputs are actually perceived as more sustainable — not just different from the original.
Funding
This research is supported by the BiSA Q1 Hi-Impact EQUITY WCU 2025 grant, Universitas Indonesia (effective 22 October 2025 – 24 August 2026). Additional funding sources will be listed as the project progresses.
Deliverables
- First (core) paper — under review
- Second paper (preprint) — under review
- Dataset v1 (Zenodo)
- Dataset v2 (Hugging Face)
BibTeX citation
@article{pradana_promptingsustainability_paper1, author = "{Mohammad Raditia Pradana and Ahmad Gamal and Jagannath Aryal}", title = "Reframing urban 'sustainable' streetscapes: Evidence from cross-city transformation and convergence", year = "2026", note = "Under review",}@article{pradana_promptingsustainability_paper2, author = "{Mohammad Raditia Pradana and Ahmad Gamal and Jagannath Aryal}", title = "Interpreting 'sustainable' streetscapes with generative AI: context-rich vs. generic prompting", year = "2026", howpublished = "\url{https://www.researchsquare.com/article/rs-9603925/v1}", note = "Preprint, under review",}