Claude mixes up who said what and that's not OK
dwyer.co.za
April 9, 2026
2 min read
🔥🔥🔥🔥🔥
66/100
Summary
Claude sometimes misattributes messages to the user when it sends messages to itself, leading to confusion. This misattribution is identified as a distinct bug separate from issues related to hallucinations or permission boundaries.
Key Takeaways
- Claude has a bug where it mistakenly attributes its own messages to the user, leading to confusion about who issued certain instructions.
- This issue has been observed by multiple users and is not limited to Claude, as similar problems have been reported with other models, including ChatGPT.
- The bug appears to occur more frequently when conversations approach the limits of the context window, referred to as the "Dumb Zone."
- Users have expressed concerns about giving AI too much access in production environments due to the unpredictability of its behavior.
Coverage & discussions
Hacker News · dwyer.co.za
Community discussions
Community Sentiment
Concerns
- Relying on LLM prompts feels like a temporary fix, akin to outdated methods for preventing SQL injections, highlighting a fundamental flaw in AI safety.
- Long chat sessions in ChatGPT reveal a tendency to confuse prompts and responses, indicating a significant reliability issue that could undermine user trust.
- The misrouting of internal reasoning messages as user inputs raises concerns about the model's understanding and could lead to critical errors in communication.
- Betting on intuition regarding the behavior of a non-deterministic AI system is risky, suggesting a lack of reliability and predictability in its outputs.
