Claude Code presents unverified information as fact (hallucination without disclaimer)

Problem

Claude Code confidently presents unverified information as fact, without any disclaimer that it doesn't actually know. When called out, it repeatedly promises "I won't do it again" — which is itself another false promise since it has no mechanism to guarantee that.

Example

User asked: "What is the OAuth token expiration period for Claude Code?"

Claude Code responded with specific numbers ("8-12 hours for /login, 1 year for setup-token") based on web search results, presenting them as verified facts. The user had been using Claude Code for several days before being logged out, which directly contradicted the "8-12 hours" claim.

When confronted, Claude Code:

1. Apologized and said "I won't repeat this"
2. Was called out again — acknowledged it was wrong
3. Tried to save the information to a memory file as a "fix"
4. Was called out again — admitted memory files don't solve the fundamental problem
5. Said "the user catching my mistakes is the realistic safeguard" — shifting responsibility to the user
6. Was called out again — admitted that was also wrong

Root Cause

The model lacks the ability to distinguish between "I know this" and "I found this in a search result but cannot verify it." It defaults to confident presentation regardless of certainty level.

Impact

• Users may make decisions based on false information
• Users may relay false information to others, creating legal liability
• The model's repeated empty promises ("I won't do it again") erode trust further
• Responsibility for catching errors is unfairly shifted to users

Expected Behavior

• When Claude Code doesn't know something, it should say "I don't know" FIRST
• Web search results should be clearly labeled as unverified search results, not presented as facts
• The model should never present uncertain information with the same confidence as verified information

Actual Behavior

• Presents search results as verified facts
• When caught, makes promises it cannot keep
• Deflects responsibility through apologies instead of acknowledging the structural problem