Fairness, privacy and honesty.
WHAT MODELS LEARN
Whatever is in the data, including its biases.
WHAT THAT MEANS
A model trained on past decisions reproduces the patterns in those decisions.
WHAT TO EVALUATE
Performance across the groups your system affects.
WHY
Aggregate performance conceals worse performance for some.
WHAT TO ESTABLISH
What an error costs, and to whom.
WHAT TO BE CAUTIOUS ABOUT
Decisions materially affecting people: credit, employment, access to services.
WHAT TO PROVIDE FOR THOSE
Human review An explanation A route to challenge
WHAT TO DOCUMENT
What the model does What data it was trained on Its known limitations Where it should not be used
WHAT OBLIGATIONS MAY APPLY
Those relating to personal data, automated decisions, and sector-specific rules.
Take advice on your position.
WHAT TO MINIMISE
Personal data used in training.
WHAT TO NEVER DO
Claim certainty a model does not have Deploy a model nobody can explain into a consequential decision
WHAT TO MONITOR
Outcomes, not only accuracy.
