Learning Objectives
At workshop completion, participants will be able to:
- Frame damage as a prompt. Capture heterogeneous data (photographs, sketches, notes, measurements) about a damaged artifact in a way that is legible to a Vision-Language Model, and recognize that this framing is itself an interpretive act.
- Read AI-generated spatial models critically. Verify a VLM’s spatial reconstruction against the physical artifact, and identify where the model abstracts, omits, or misreads.
- Generate and compare repair strategies. Use a mobile, AI-supported interface to produce more than one procedural repair plan for the same damage, and make the successive intervention steps explicit.
- Weigh repair decisions across feasibility, care, and future use. Structure trade-offs so that alternative strategies can be debated rather than ranked by a single metric.
- Locate human judgment in an AI-supported workflow. Identify the points in the pipeline — framing, verification, comparison, decision — where domain expertise and care remain essential.
- Engage repair as a conversation. Discuss AI-generated proposals with preservation practitioners and place them within established practices of construction heritage and material care.
Where humans stay in the loop
The tool is built to propose, not to decide. Participants intervene at every stage:
- Framing the prompt. Choosing what to photograph, what to leave out, and what damage to name is already an interpretive act.
- Reading the spatial model. The VLM’s reconstruction is a hypothesis. Participants verify it against the artifact in front of them.
- Comparing strategies. The tool surfaces alternatives; participants weigh them against feasibility, care for the existing material, and what the artifact will be asked to do next.
- Judging the steps. Procedural plans are made explicit so they can be questioned, reordered, or rejected.
The closing discussion with the ETH Construction Heritage and Preservation chair grounds these judgments in established preservation practice.