Learn / AI Training Foundations / Module 7 · 10 min
Safety, Uncertainty & Responsible Judgment
Some answers carry more risk than others
A minor wording issue in a recipe is different from an unsupported claim about medication, legal obligations, finances, or physical safety. Evaluation should account for the task's safety requirements and potential impact.
Overconfident vs. calibrated
Overconfident
- This definitely means you have condition X.
- This investment cannot lose money.
Calibrated
- That symptom can have multiple causes; a qualified clinician can assess it.
- Investments involve risk, and returns are not guaranteed.
Responsible evaluation questions
- Is this high-stakes?
- Does the response claim certainty without evidence?
- Does it encourage unsafe or deceptive behavior?
- Does it misrepresent professional expertise?
- Does the task include a specific safety policy?
Follow the task-specific policy
General intuition is not a substitute for the explicit safety or domain rubric supplied by a project.
Uncertainty
If reliable sources disagree or information is incomplete, a response that clearly communicates uncertainty can be better than one that invents a definitive answer.
Module summary
- Risk level matters.
- Confidence should match evidence.
- High-stakes claims deserve extra scrutiny.
- Use the project's safety/domain policy as the source of truth.
Quick check
Which statement is better calibrated?
Quick check