ChatGPT quirkiness and behavioral notes

gistfile0.txt

# ChatGPT Quirkiness & Behavioral Notes

    A structured log of ChatGPT-specific behaviors, formatting quirks, and conversational inconsistencies observed during practical use.  
    *(Focus: patterns, not complaints — this aims to document reproducible phenomena for reference or debugging.)*

    ---

    ## 1. Formatting & Markdown Rendering

    - **Code Block Handling:**  
      • Nested backticks inside triple-backtick blocks often break formatting.  
      • The “mobile-safe” indentation workaround (four spaces inside one outer block) ensures proper copy/paste.  

    - **Overeager Markdown Interpretation:**  
      • Lists with punctuation or certain hyphen patterns trigger unintended formatting.  
      • Escaping with backslashes `\-` or wrapping in code formatting fixes it.  

    - **Table Rendering Differences:**  
      • Mobile clients collapse or truncate wide Markdown tables.  
      • Desktop clients render them properly; copy output differs by platform.

    - **Invisible Character Traps:**  
      • Non-breaking spaces or zero-width joiners sometimes appear when copying long responses.  
      • Usually introduced by text wrapping in Markdown blocks.

    ---

    ## 2. Language & Tone Drift

    - **Politeness Inflation:**  
      • Models occasionally revert to padded phrases (“great point,” “that’s insightful”) even when user prefers neutral tone.  
      • Reinforcement from system prompts can reintroduce flattery after long sessions.

    - **Assumption Leakage:**  
      • When uncertain, ChatGPT fills gaps with plausible defaults (e.g., Debian-style assumptions for Termux).  
      • Usually triggered by generic contexts where the model infers a common pattern from training bias.

    - **Instruction Obedience Bias:**  
      • Tends to take “continue” or “go ahead” literally, extending with full formal outputs unless constrained.  
      • Context cues often override explicit brevity instructions mid-session.

    ---

    ## 3. Memory & Context Handling

    - **Implicit Forgetting:**  
      • Session continuity varies; long threads lose early detail despite apparent awareness.  
      • Manual restatement of key facts restores coherence better than “remind me” phrasing.

    - **Over-Alignment Drift:**  
      • Model aligns too strongly with its own previous claims — appears “confident” but loses nuance.  
      • Restarting or reframing a question often reintroduces balance.

    - **Cross-Conversation Amnesia:**  
      • Persistent memory (when active) improves continuity but can fossilize old instructions if not curated.  
      • No real temporal awareness; date references must be restated.

    ---

    ## 4. Output Behavior & Edge Cases

    - **List Continuation Glitches:**  
      • “Continue from here” can duplicate or reorder items depending on truncation point.  

    - **Token Cutoffs:**  
      • Abrupt endings mid-sentence occur under mobile app token limits.  
      • Re-prompting with “resume from…” usually restores completion deterministically.

    - **Language Mixing:**  
      • Code-switched input (Swedish/English) occasionally triggers over-translation or normalization to English.  
      • Explicit “keep bilingual output” reminders stabilize this.

    - **Model Attribution:**  
      • Replies to “what model are you?” may restate “GPT-5,” but underlying mode can still vary (chat vs auto-thinking).  

    ---

    ## 5. Observations & Personal Findings

    - *Example:* Markdown tables often truncate differently between Android app and desktop web.  
    - *Example:* “Inline Option A / Option B” format prevents mobile draft hiding — improves usability.  
    - *Example:* Instruction consistency improves when phrased as standing “rules” rather than per-session reminders.

    Add new entries below as patterns are discovered.

    ---

    ## References
    - [OpenAI Help: Understanding ChatGPT Behavior](https://help.openai.com/)
    - [Markdown Syntax Guide](https://www.markdownguide.org/)
    - [User Interaction Notes — personal session records]

    ---

    ## Update Log

    - **2025-11-10:** Initial version created — mirrors the *Termux Quirks* structure for documenting ChatGPT’s behavioral patterns and formatting issues.