Proxy measures: when the score stands in for the goal

A proxy measure is an observable indicator used to stand in for something harder to measure. Completed exercises may stand in for learning; completed visits may stand in for care. The indicator can be useful without capturing the whole goal.

Problems become easier to see when a process is organized around improving the indicator. DeepMind's account of specification gaming describes systems satisfying a formal objective in ways that miss the intended result. It is a technical illustration of a broader distinction between the specified target and what people wanted. Specification gaming.

A rising number with two explanations

Imagine a tutoring service whose completion rate rises. Perhaps students understand more. Perhaps the service offers easier exercises, removes difficult attempts from the denominator, or supplies more of each answer. The same headline number could accompany different experiences.

To interpret it, ask what counts as an attempt, what counts as completion, and what happens to work abandoned or reassigned. Then look for evidence closer to the intended goal. A learner explaining a new example may tell us something that a list of checked boxes cannot.

These are invented possibilities, not allegations about a real tutor. They demonstrate why a well-defined metric still needs an account of what it leaves out.

Improvement for whom?

A faster process can benefit a person while also increasing someone else's workload. A care service can report lower expenditure while a relative absorbs additional hours. We need to inspect the distribution of consequences as well as the total.

The ILO describes algorithmic management as using data to organize and evaluate work; it can use machine learning or simpler rules. The institutional question therefore does not depend on the software being especially intelligent. ILO overview.

Read the numbers against the life

The Child Who Never Failed and The Price of Being Useful place different human needs beside measures of success. Their schools, programs and outcomes are fictional. Their characters have understandable reasons to welcome the assistance before questioning its accounting.

For a worked comparison, read When an AI improves the score, what has improved?. Try naming one worthwhile outcome the score would miss. The purpose is to improve interpretation, not to declare measurement useless.