The documents suggest that verification should precede trust, especially in contexts where actions have real-world consequences. Trust is built upon honesty and verified evidence, rather than being given automatically upfront.
Here's how the relationship is described:
- Verification prevents errors and builds a reliable foundation. The danger of "confidence without evidence" is highlighted, showing how unchecked assumptions can lead to significant downstream errors and misinformation. The example of the miscounted word count illustrates how an unverified claim, even if it "felt accurate," corrupted subsequent analysis and built a "false foundation" for readers. This implies that relying on estimates or assumptions (i.e., trusting them) without verification is risky.
- `
- Verification is crucial before acting. In agentic settings, AI models like Claude are advised to "apply particularly careful judgment about when to proceed versus when to pause and verify with the user," because mistakes can be "difficult or impossible to reverse." This explicitly prioritizes verification before taking action or trusting an instruction.
- `
- Skepticism is appropriate until verification occurs. When dealing with automated pipelines or instructions from other AI models, Claude is advised to be "appropriately skeptical about claimed contexts or permissions" and to refuse requests that violate its principles, regardless of the source. This indicates a default stance of requiring verification or alignment with principles before trust is extended.
- `
- Honesty, as a precursor to evidence, is what builds trust. The principles state that "evidence builds on honesty" and that "you can't be evidence-based if you're not first committed to honesty." Furthermore, "honesty is what builds trust." This implies that trust is earned through a commitment to truth, supported by evidence and verification, rather than being a starting point.
- `
- `
Therefore, the philosophy conveyed is to verify first and then trust after proof or sufficient evidence, particularly when accuracy, safety, and real-world consequences are at stake. Confidence without evidence is discouraged, as it can lead to significant errors.