When labeling documents for a custom extraction model, what happens if you label the same field inconsistently across documents?
Inconsistent labels give the model contradictory supervision for identical field patterns, so it cannot learn a stable decision boundary. This label noise directly degrades precision and recall, meaning model accuracy will decrease — the constraint the question asks you to identify.
Why this answer
Inconsistent labeling of the same field across training documents introduces conflicting ground truth, which degrades the model's ability to learn a stable pattern. The neural trainer will still complete, but precision and recall for that field drop because the model receives contradictory signals. Consistency is a core labeling best practice in Document Intelligence Studio.
Exam trap
AI-103 often tests whether candidates understand that bad labels degrade accuracy rather than causing hard failures — the tempting wrong answer is 'training will fail immediately,' which sounds strict but is not how the service behaves.
How to eliminate wrong answers
Option B is wrong because inconsistent labels do not cause an immediate training failure — the service accepts the labels and produces a model, just a less accurate one. Option C is wrong because label quality directly drives model quality; claiming 'no impact' ignores how supervised learning works. Option D is wrong because the model has no mechanism to detect and auto-correct inconsistent human labels — it treats every label as ground truth.