Spotting Toxic Clouds (IJmondCAM)

What can I learn?

A community project by Waag and researcher Yen-Chia Hsu in which residents of the IJmond region take part in labelling data — deciding what counts as a suspicious smoke cloud — to build the training dataset for a monitoring model, giving people direct influence over the choices the model makes.

Core insight

The "ground truth" a model learns from is a human judgement, not a fact of nature. When the affected community labels the data, it holds power over what the system sees and flags — a hands-on demonstration of where bias enters and how it can be steered.

How to use it in daily work

A tangible, local example of citizens shaping an AI system rather than merely being subject to it, useful for demystifying how training data and labelling work.

• Use IJmondCAM to explain to clients, in concrete Dutch terms, that someone decides what an AI is trained to recognise — and that it can be them.
• Draw on the labelling activity to show how datasets are built and why who labels the data matters for fairness.