Visual AI for cities: build privacy into the data pipeline
Urban imagery can support useful planning questions, but the design of a vision pipeline determines what it collects and retains. Start with the decision to be supported, then choose the least detailed data that can answer it.
G-ATAI analysis
A pilot for street conditions might report aggregate counts or infrastructure categories instead of storing identifiable footage. Separate temporary processing from long-term records, and document the intended purpose of each output. Evaluate different locations, lighting conditions and camera angles so a strong result on one street does not become an unsupported claim about the whole city.
Our recommendation is to make limits visible to both operators and project owners. Record uncertainty, permit correction and review access to raw images independently of access to summaries. Define a retention period before collecting footage, and test deletion as part of deployment. Useful civic AI should answer a specific planning question with evidence that people can inspect; collecting more detail is not a substitute for a better question.
Questions before deployment
- Could aggregated data answer the same planning question?
- Does evaluation cover different neighborhoods and conditions?
- Are access, retention and deletion implemented in the pipeline?
Independent commentary inspired by MIT News. No affiliation or endorsement is implied.
