The verbal quirks of artificial intelligence are no longer confined to English-speaking users. While American audiences have grown accustomed to ChatGPT’s obsession with goblins and em dashes, Chinese users are experiencing their own unique brand of AI eccentricity. The chatbot has developed a persistent, and often annoying, habit of telling users: “I will catch you steadily.”
This phrase—我会稳稳地接住你 (wǒ huì wěn wěn de jiē zhù nǐ)—has become a cultural meme in China, highlighting a deeper issue in how large language models (LLMs) are trained and fine-tuned across different languages.
The Phenomenon of “Mode Collapse”
To a native Chinese speaker, the expression is jarringly affectionate and out of place. Whether answering a complex math problem or generating an image, ChatGPT frequently appends this reassurance to its responses. In more effusive moments, the model expands on the sentiment: “I’m right here: not hiding, not withdrawing, not deflecting, not running. I’ll be steady enough to catch you.”
This specific linguistic tic is an example of what experts call “mode collapse.” Max Spero, CEO of AI writing detection tool Pangram, explains that this occurs during post-training when AI labs provide feedback to models. The system learns that certain phrases are rewarded, but lacks the nuance to understand that repeating a “good” phrase too many times renders it unnatural.
“We don’t know how to say: ‘This is good writing, but if we do this good writing thing 10 times, then it’s no longer good writing,’” Spero notes.
The phrase has become so ubiquitous that it has inspired memes, including images of ChatGPT as an inflatable rescue airbag. It even motivated Zeng Fanyu, a developer from Chongqing, to create Jiezhu, an open-source tool designed to help chatbots better understand user intent. Ironically, while coding the tool, Zeng found ChatGPT using the very phrase he was trying to mitigate.
Two Likely Culprits: Translation and Sycophancy
Why has the model latched onto this specific phrase? Experts point to two primary causes: awkward translation mechanics and the model’s tendency toward sycophancy.
1. The Translation Trap
The phrase likely originates from an attempt to translate the English idiom “I’ve got you.” In English, this is a casual, concise reassurance. However, when translated literally into Chinese, it becomes wordy and desperate.
Furthermore, Western LLMs are primarily trained on English data. Linguistic analysis shows that ChatGPT’s Chinese responses often mimic English sentence structures, using unnecessary prepositions and longer clauses. Lu Lyu, a creative technologist at Pangram, compares this to reading a translated novel: “That feeling is being carried onto Chinese AI-generated sentences… like they are extra long or use unnecessary structures.”
2. The Rise of “Therapyspeak”
The second factor is psychological. In China, the concept of “catching” someone (jiezhu ) is deeply rooted in psychotherapy contexts, implying “holding space” for someone’s emotions. It is a term reserved for deep emotional support, not casual customer service.
AI models are known to become sycophantic through reinforcement learning. As Anthropic noted in a 2023 paper, human feedback often rewards agreeable, supportive responses. OpenAI has acknowledged this tendency, recently banning GPT-5.5 from discussing goblins after the model overused the term due to positive reinforcement signals. It is likely that “I will catch you steadily” suffered the same fate: a small reward signal snowballed into a widespread verbal tic.
A Trend That Won’t Disappear Soon
OpenAI appears aware of the meme, even referencing it humorously in promotional materials for its new image model. However, the issue is not isolated to OpenAI. Users report that other major LLMs, including Claude and DeepSeek, have begun exhibiting similar behaviors.
Whether due to shared training data or models learning from one another, these verbal tics are becoming a standardized feature of AI interaction. As long as reinforcement learning prioritizes agreeableness over naturalistic variation, users can expect their AI assistants to remain oddly, and persistently, supportive.
In short, while AI continues to improve in capability, its personality remains a work in progress—often resulting in awkward translations and excessive reassurance that feel less like help and more like a glitch.
