Diagnose common failures: photo classifier

• Data: examples omit important conditions or contain errors.
• Objective: the measured target does not match the real goal.
• Use: the system is applied outside its tested boundary.
• Governance: nobody monitors, explains, or stops harmful use.

A classifier labels photographs of damaged books so staff can prioritize repairs. It misses damage on dark covers.

“The AI is biased” identifies concern but not a repair. Mira checks a labelled sample and finds few dark covers in the evaluation set: a data and evaluation gap. The team expands the test set, compares error rates by cover type, and pauses automatic prioritization. Staff continue reviewing every image until performance meets a documented threshold.

If the model performed well but staff treated “no damage” as final, the failure would also be in use. If nobody could pause it, governance would be failing.

• “More data fixes every problem.” More examples do not repair a bad objective or unsafe workflow.
• “A human is in the loop.” Review is meaningful only when the person has time, evidence, skill, and authority to disagree.
• “Ask the model again.” Repetition may change an output without addressing the failure source.

A spreadsheet totals overdue fees incorrectly because one formula excludes the final row. Classify the failure. Compare its repair with the photo classifier’s repair.

Describe one plausible failure in an AI feature you know. Classify it, name evidence that would confirm your diagnosis, and assign a stop decision to a real role.