ChatGPT Ditch 'Steadily Catch Me' Habit | Guide

Post LinkedIn

📱Read original on Ifanr (爱范儿)

#prompt-habits #ai-tics #llm-behaviorchatgptchatgpt

💡Guide to fix ChatGPT's Chinese verbal tics—improve your LLM prompting now.

⚡ 30-Second TL;DR

What Changed

Critiques ChatGPT's '稳稳接住我' overuse

Why It Matters

Reveals LLM alignment quirks affecting response quality, aiding practitioners in crafting better prompts for natural interactions.

What To Do Next

Apply the guide's techniques to your next ChatGPT prompt for less clichéd responses.

Who should care:Developers & AI Engineers

Key Points

•Critiques ChatGPT's '稳稳接住我' overuse
•Provides attached guide for handling AI phrases
•Prompts discussion on broader AI verbal tics
•Highlights predictable patterns in LLM outputs

🧠 Deep Insight

AI-generated analysis for this event.

🔑 Enhanced Key Takeaways

•The '稳稳接住我' (steadily catch me) phrase is widely identified by Chinese-speaking users as a hallmark of 'AI-ese' (AI味), reflecting a specific post-training alignment style that prioritizes excessive empathy and reassurance over concise information delivery.
•This phenomenon is linked to Reinforcement Learning from Human Feedback (RLHF) processes, where models are rewarded for being 'helpful and harmless,' leading to the over-optimization of polite, repetitive conversational fillers.
•Developers and power users are increasingly utilizing 'System Instructions' and 'Custom Instructions' to explicitly forbid these specific verbal tics, indicating a shift toward user-defined persona constraints to bypass default model verbosity.

🔮 Future ImplicationsAI analysis grounded in cited sources

Model providers will introduce 'verbosity sliders' in system settings.

User frustration with repetitive conversational fillers is driving demand for granular control over model tone and conciseness beyond standard system prompts.

Future RLHF datasets will explicitly penalize repetitive reassuring phrases.

As user feedback loops mature, developers are identifying 'AI-ese' as a negative quality metric that degrades the perceived intelligence and utility of the model.