Keeping people in the loop.
WHAT OVERSIGHT MODELS EXIST
A person decides, with the model advising A person reviews before the decision takes effect A person reviews a sample afterwards Fully automated, with an appeal route
WHAT TO CHOOSE ON
The consequence of error, and the volume.
WHAT HIGH-CONSEQUENCE DECISIONS REQUIRE
Human involvement, meaningfully.
WHAT MEANINGFULLY MEANS
The reviewer has the information, time and authority to disagree.
WHAT UNDERMINES IT
Volume too high to review properly Interfaces presenting the model's answer as the default No consequence for approving everything
WHAT THAT PRODUCES
Rubber-stamping, which is oversight in name only.
WHAT TO MEASURE
How often reviewers disagree with the model.
WHAT A VERY LOW RATE SUGGESTS
Either an excellent model, or oversight that is not happening.
WHAT TO PROVIDE REVIEWERS
The reasons, not only the answer The confidence Similar past cases
WHAT TO ALWAYS PROVIDE AFFECTED PEOPLE
A route to contest a decision.
WHAT TO MONITOR
Appeal rates and outcomes.
WHAT HIGH OVERTURN RATES INDICATE
A model that should not be deployed as it is.
