Specify the Setting Before the Model

The concept

An AI tool sits inside a human process. Its risk depends on who uses it, what decision follows, and whether affected people can detect and challenge mistakes. Begin with a bounded purpose statement. Name the input, expected output, decision owner, and unacceptable failure before comparing vendors or impressive demonstrations.

Worked example

A fictional museum wants help labeling incoming photographs. 'Describe everything accurately' is too broad. A better specification asks for suggested dates and visible objects, marks uncertain claims, and requires a curator to approve public labels. A wrong coat color is inconvenient; inventing a person's identity may be harmful. The team therefore tests identity claims separately and makes abstention acceptable rather than forcing every image into a confident answer.

Limits and judgment

A purpose statement is not an assurance certificate. People may reuse outputs in ways designers did not intend. Explain limits where the output appears, restrict high consequence actions, and revisit assumptions after use begins. The relevant unit of evaluation is the complete task, including human review and correction, rather than the model in isolation.

RETRIEVE. REASON. REMEMBER.

Knowledge check.

Choose one answer for each question, then check your reasoning.