Model Dignity Check

Where you would use it:

·       Before any AI system or feature goes live

·       When reviewing existing AI implementations

·       During design phases to catch problems early

·       When updating or retraining AI models

The tool: Five Questions to ask before launch:

·       Who becomes invisible when we optimize? Name specific people, not categories—“elderly residents in walk-ups” not “some users”

·       What “normal” is baked into the training data? Every dataset tells a story about who matters—whose reality shaped this system?

·       How does this perform for our most vulnerable users? Test on edge cases—the users with least power, resources, or technical literacy

·       Can affected humans understand and contest decisions? Opacity breeds distrust—is there a real path to challenge the algorithm?

·       Does this strengthen or erode human agency? Are we augmenting human judgment or replacing it?

How to use it: Before launch, document written answers to all five questions. Be specific—vague answers hide real problems. If any answer troubles you, redesign before deploying. This isn’t a compliance checkbox but a discipline for catching what pure optimization misses. Run this check again whenever you update or retrain the system.