A training team includes negative examples of bad leaked-PII summaries in the prompt so the model knows what to avoid. What technique is this?

Privacy-safe prompting benefits from showing failure modes. Including negative examples of bad leaked-PII summaries teaches the model what to avoid—negative constraints and counterexamples as a prompt technique. For municipal training teams, illustrating forbidden patterns (exposed SSNs, full birthdates, unnecessary medical detail) clarifies boundaries more sharply than success samples alone.

Using negative retrieved precedent examples as a prompt technique for privacy-safe civic outputs fails because retrieved precedent examples imply pulling prior cases via retrieval to ground generation, while the stem describes author-provided bad leaked-PII summaries placed in the prompt as constraints with no retrieval step. Using negative anonymized paraphrase examples as a prompt technique for privacy-safe civic outputs fails because it narrows to failed anonymized paraphrases of a source, while the stem describes generic bad leaked summaries used as negative constraints to show what to avoid, not paraphrase work. Using negative contrastive preference examples as a prompt technique for privacy-safe civic outputs fails because contrastive preference framing requires a paired chosen-versus-rejected comparison to express a preference, while the stem describes only standalone bad examples used as stop-sign constraints, not a paired comparison.

Exam caveat: negative examples must be carefully anonymized in training materials so the teaching set itself does not become a leak. The technique is about constraint clarity, not circulating real PII “for realism.”

Operational check: maintain approved negative examples with synthetic or redacted PII; place them in the prompt with explicit “do not produce outputs like these” language; scan generated summaries with a PII detector before any training artifact is shared.