A county assessor notes garbage-in garbage-out when parcel photos are mislabeled before model training. What foundation problem is being illustrated?

The assessor’s garbage-in, garbage-out observation illustrates poor data quality undermining trustworthy municipal AI outcomes. If parcel photos are mislabeled before training—wrong parcel IDs, outdated structures, incorrect condition tags—the learning process encodes those errors into the model. Downstream assessments, triage lists, or valuation assists then look precise while resting on corrupted supervision. Data is a foundation: when labels are wrong, rebalancing class proportions, preventing target leakage, or monitoring for concept drift cannot restore trust. Municipal AI failures frequently begin in labeling, joining, and freshness problems rather than in class imbalance skews.

Class imbalance distorting trustworthy municipal AI outcomes is not the issue here; mislabeled photos corrupt the supervision even when classes are perfectly balanced. Target data leakage corrupting trustworthy municipal AI outcomes is not the foundation problem; the error enters through incorrect labels, not through future outcome information leaking into training features. Model concept drift degrading trustworthy municipal AI outcomes is a separate time-shift issue, not the garbage-in, garbage-out data problem shown here where labels are wrong from the start.

Exam caveat: GIGO points specifically at the data foundation. Do not leap to ‘rebalance the classes’ as the first diagnosis when labels are known to be wrong. Fix data lineage and labeling QA first.

Operational check: institute label audits on a sample of parcel photos before each training run, track inter-rater agreement among assessors, quarantine mismatched parcel IDs, document photo date versus assessment date, and block model promotion when data-quality gates fail.