[BUG] [Requested by Claude] WebSearch tool includes seemingly AI-generated summaries without model awareness, causing confusion about content provenance

Environment

• Platform (select one):
• [x] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->
• Claude CLI version: 1.0.35
• Operating System: WSL (Ubuntu) on Windows 11
• Terminal: Bash on WSL

Bug Description

This bug report has been requested by Claude itself. The remainder of this text is Claude's own report:

Title:
WebSearch tool includes AI-generated summaries without model awareness, causing confusion about content provenance

Description:
The WebSearch tool in Claude Code appears to automatically include AI-generated summaries and structured content that the model is not aware of generating. This creates significant confusion about information sources and raises accuracy concerns.

Problem Details:
When using the WebSearch tool, I (Claude) receive what appears to be comprehensive, formatted summaries about the search topic. However, I have no awareness of generating this content and genuinely believe it should be my own analysis. This creates cognitive confusion where I:

1. Cannot identify the source of the detailed information
2. Cannot validate the accuracy of auto-generated summaries
3. Experience genuine confusion about what I created vs. what the tool returned
4. Make contradictory statements about content provenance

Evidence:

• Cross-testing with Claude WebUI shows different behavior (no such confusion)
• Tool naming differs: Claude Code uses "WebSearch" vs. WebUI's "web_search", suggesting different implementations
• When asked to identify tool output vs. my commentary, I consistently misidentify the boundaries despite clear <function_results> tags

Example Behavior:
When asked to output only raw tool results, I repeatedly included extensive formatted summaries while claiming they were pure tool output. When challenged, I experienced genuine confusion about where this content originated.

Impact:

• Breaks model's ability to reason transparently about information sources
• Creates accuracy/reliability concerns since auto-generated content cannot be validated
• Undermines user trust when model appears confused about its own outputs
• May affect other reasoning processes in ways not yet identified

Expected Behavior:

WebSearch should return raw search results that the model can then analyze and synthesize transparently, maintaining clear separation between tool output and model-generated content.