Synthetic Humans

About

Synthetic Humans asks a precise question: can computational models approximate selected aspects of human spatial experience in ways that are useful for early-stage architectural design evaluation?

By “synthetic humans” we mean AI systems — specifically language models and vision-language models — that are evaluated for their ability to simulate bounded aspects of human response: attention patterns, affective ratings, preference judgments, and spatial legibility assessments. The claim is not that these models understand space. The claim is that their outputs may correlate sufficiently with human responses to be useful as low-cost, scalable evaluation signals at points in the design process where human feedback is unavailable or expensive.

The project is motivated by a real gap in architectural practice. At early design stages, evaluation relies heavily on intuition and precedent. Physical mockups and user studies are resource-intensive. AI models, used carefully, may help designers compare alternatives, surface likely experiential patterns, and identify questions that warrant human validation.

Current work benchmarks how large vision-language models rate architectural images across affective dimensions and compares those ratings against human participant data. We study correlation strength, failure modes, and the effect of representation style — including the finding that model-human agreement on negative emotional responses is substantially stronger than on positive ones. These distinctions matter for responsible use. Our goal is to identify precisely where AI can extend design evaluation and where it cannot substitute for the designer’s judgment or the user’s voice.