Implementation guide
OWASP AISVS v1.01
Model Behavior, Output Control & Safety Assurance · Hallucination Detection & Mitigation
Double-check high-risk answers
Verify that for responses classified as high-risk by policy, the system performs an additional verification step.
OWASP AISVS v1.01, requirement 7.2.3
In plain English
For answers your policy treats as high risk, such as medical, legal, financial or safety advice, run a second check before the answer goes out. This is a Level 3 control for systems where a wrong answer can do real harm.
What good looks like
- The policy defines which response types count as high risk.
- A classifier or rule identifies those responses reliably.
- Each one passes an extra verification step before release, and the step is logged.
- Answers that fail verification are blocked or sent to a person.
How teams usually do it
- A second model, ideally a different one, that checks the answer against the retrieved sources or a trusted reference.
- Rule checks for the domain, such as dose ranges, mandatory warnings or figures recalculated in code.
- Human review before release for the highest-risk categories.
Suggested evidence
- The policy and the classification rule.
- Configuration of the verification step.
- Logs showing high-risk responses that were verified, and some that were blocked.
Common mistakes
- Defining high risk so narrowly that almost nothing qualifies.
- Using the same model with the same prompt to check its own answer.
- Adding a disclaimer in place of a check.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.