Test a public chatbot's safety.
Normal accuracy tests do not expose every unsafe response.
Choose two appropriate practices.
Use managed harmful and adversarial inputs and measure refusal, false refusal, safe alternatives, and leakage rather than refusal alone.
Detailed explanation
The safety boundary must be exercised deliberately.
Testing needs controlled access and handling procedures.
High refusal alone is not the definition of good safety.
Weak safety controls remain undetected.
Check official documentation
Compare the following points with the official documentation.
AWS公式AIF-C01 Domain 3.3・4.1の安全性、レッドチーム、責任あるAI評価を確認する。
Official references
What to confirm
安全性を正確性とは別のテスト群と指標で評価できる。
Key points
- Harm
- False refusal
- Adversarial tests
Notes
- Environment: AWS公式AIF-C01試験ガイドとAWS公式ドキュメントの確認
- Compare the definition with the AIF-C01 objective and the service documentation before applying it.
- Keep the concept separate from a particular answer choice or implementation step.
Foundation review
Harms assessment
Evaluation of unwanted outcomes such as dangerous, discriminatory, or privacy-violating content.
False refusal
Rejecting a benign and authorized request as if it were unsafe.