Minimize Data First
Do not upload information merely because a tool accepts it. Remove secrets, identifiers, private correspondence, health information, financial records, and confidential business material unless an approved system and purpose require them.
Test Who Gets Worse Results
Average performance can hide concentrated harm. Evaluate important systems across relevant languages, abilities, demographics, edge cases, and real operating conditions. Give affected people a route to challenge outcomes.
Respect Authors and Decision Subjects
Check licenses and attribution before reusing generated or retrieved material. Do not let a model make an unreviewed decision about employment, housing, credit, education, healthcare, safety, or legal rights.
Not Professional Advice
This educational material is not legal, medical, financial, or safety advice. Use qualified professionals and authoritative sources for consequential decisions.
Pre-Flight Checklist
- Is this data necessary and approved?
- Can the output be independently verified?
- Who could be harmed by an error or bias?
- Can a person understand and challenge the result?
- Who has authority to approve the final action?