Responsible AI landscape
Mental model: responsible AI is an operating loop, not a review after modeling. It decides whether a use case should exist, names harms and affected people, requires evidence to release, and assigns authority to pause or repair the system.
Mechanism: values → controls → decisions
Convert every value into a harm, owner, artifact, metric, threshold, and rollback. Fairness becomes slice-error and recourse controls; privacy becomes data-flow and retention controls; safety becomes threat modeling and action boundaries. A value with no operational decision cannot protect anyone.
control = {"harm":"wrong denial", "owner":"risk lead", "metric":"appeal rate", "threshold":0.03, "rollback":"disable recommendation"}
assert all(control.values())
print("control is actionable")
Run with python3; expected output is control is actionable.
| Dimension | Evidence |
|---|---|
| Fairness | slice metrics, intervals, appeals |
| Privacy | inventory, access and deletion tests |
| Safety/security | threat model, red team, gates |
| Transparency | disclosure, documentation, provenance |
| Accountability | owner, incident path, rollback |
| Social impact | stakeholder and labor assessment |
Frameworks do not eliminate tradeoffs: privacy may limit fairness measurement and latency targets may reduce review. Record the tradeoff and accountable owner. Do not release when a material harm lacks a metric, authority, or recourse mechanism.
Production lens and failure modes
Review controls when a model, prompt, retrieval corpus, authority boundary, or affected population changes. A values statement, a vendor assurance, or aggregate accuracy is not release evidence. Preserve decisions and incidents so an affected person can challenge an outcome and the owner can reproduce or roll it back.
Exercises
- Turn “be transparent” into a disclosure, artifact, and test.
- Add a rollback owner to one product's control register.
Connects to: governance · oversight · threat modeling
Sources
- NIST AI RMF — lifecycle risk management.
- OECD AI Principles — accountability and human-centered principles.
- NIST AI risk characteristics — interacting trustworthiness dimensions.