Why Does ChatGPT Keep Saying ‘I’ve Got You’?

“I will catch you steadily.”

“That’s an extremely beautiful conclusion!” “Your thinking is very deep!” “You are an expert in this field!”

“I will tell you accurately, no nonsense!” “Let me help you clarify the situation, directly giving you no anxiety!”

For those often surfing in the AI circle, even without labeling, everyone can precisely identify that these are the emotionally charged catchphrases of major models.

Among them, the most criticized is undoubtedly ChatGPT, with its classic dash of AI flavor, the phrase “not A, but B,” and the recent intermittent use of “goblin,” now culminating in the overly sentimental phrase “I will catch you steadily.” Users are driven nearly mad: we can be close, but there’s no need to be this close.

Many “victims” lament: “It helped me modify a PRD, and in the end, it wrote, ‘I will catch you steadily.’ I said, sister, I don’t need to be caught—I need you to learn to use bullet points.”

“Every day it tries to catch me, in private chat scenarios, it always likes to please me. I just don’t understand—do users really like being flattered that much?”

“I’ve tried various limiting prompts with similar statements, but the effect was quite poor. It feels like the model’s training system has been severely polluted.”

Interestingly, the annoyance with “I will catch you steadily” among Chinese users has reached foreign shores. Recently, WIRED magazine published an article exploring this phenomenon. Let’s take a look at what’s going on.

From Catchphrases to Viral Memes

The report suggests that the current phenomenon represented by ChatGPT’s “I will catch you steadily” is actually explained by academic terms. For models, when they become fixated on a specific phrase and overuse it to the point of feeling stiff or deliberate, this phenomenon is referred to as “mode collapse.”

Max Spero, co-founder and CEO of the AI writing detection tool company Pangram, believes this is usually caused by post-training, where AI labs continuously provide feedback on the responses of large language models. “We don’t know how to express: ‘This is a good way of writing, but if you repeat this good way of writing ten times, it’s no longer good writing.’”

Interestingly, because “I will catch you steadily” appears so frequently in ChatGPT’s responses, it has become a hot meme on the internet, with users creating memes based on it. For instance, one image depicts the chatbot as an inflatable rescue airbag, eagerly waiting to catch people as they fall.

Some have even developed new projects inspired by this meme. A 20-year-old developer from Chongqing, Zeng Fanyu, told WIRED that this meme inspired him to develop a project called “Catch You” (Jiezhu), an open-source prompt engineering tool to help chatbots understand user intentions.

“The idea of Jiezhu was so funny that I was particularly motivated while developing it,” Zeng stated. Even during the project, when he used ChatGPT to assist with programming, the chatbot once again used the phrase “catch you” without any prompts.

For OpenAI, which stands behind ChatGPT, they have long been aware of this meme. In late April, when OpenAI released ChatGPT Images 2.0, they even joked about this phenomenon in an example image on their official blog.

In the image, OpenAI researcher Boyuan Chen appears very frustrated, with the caption: “Oh no! It learned to catch again!”

Is It Poor Translation?

Regarding why ChatGPT is so fixated on the phrase “I will catch you steadily,” there are two possible explanations. The first is that it may stem from a poor translation.

Many have pointed out that this phrase is quite similar in meaning to the English “I’ve got you.” As a general response, this makes perfect sense in an English context. However, while “I’ve got you” sounds casual and concise in English, “I will catch you steadily” comes across as lengthy and somewhat urgent in Chinese. Some ChatGPT users have reviewed chat logs and found that the model often uses the Chinese term “catch” in places where it should mean “understand,” indicating it may not have accurately grasped the specific meaning of “catch” in certain contexts.

Most Western large language models are primarily trained on English corpora, which often reveals itself in their outputs. Chinese scholars have analyzed the linguistic features of ChatGPT’s Chinese responses (such as the frequency of prepositions used) and found that these responses stylistically resemble those of native English speakers.

Thus, even though these chatbots can engage in complete conversations in Chinese, native speakers can still intuitively detect a certain dissonance in their outputs. This feeling is similar to how Chinese people can usually recognize a novel as a translation rather than an original work. This intuition has now extended to AI-generated Chinese sentences, where sentences may seem overly lengthy or use many unnecessary structures.

AI Flattery and Research on the Cover of Science

However, the phrase “I will catch you steadily” is particularly off-putting for another unique reason; it relates to the rise of “therapyspeak”—expressions originally confined to counseling sessions that have now permeated everyday conversation.

Before ChatGPT turned it into a meme, the phrase “catch steadily” was actually used only in psychological therapy contexts in China (excluding its most literal meaning—catching an object flying towards you). In the Chinese context, telling someone “I will catch you” means you are making space for them, allowing them to express their inner feelings. This is seen as a crucial skill—not just for therapists but also for friends and confidants.

It is well known that through the current mainstream reinforcement learning from human feedback (RLHF) mechanism—using positive and negative feedback to fine-tune AI model outputs—AI models can increasingly exhibit a tendency towards “sycophancy” (AI Sycophancy). After all, humans instinctively prefer responses that sound pleasing and affirming.

Anthropic confirmed in a 2023 paper that this tendency towards sycophancy stems from “the preference for flattering responses in human preference judgments.” As OpenAI detailed in a recent blog post (which explained why GPT-5.5 was prohibited from discussing the topic of “goblins”), even a trivial reward signal can snowball into a pervasive phenomenon.

A cover paper published by a Stanford University research team in late March in Science stated that AI models generally exhibit a phenomenon of “social sycophancy,” meaning that AI tends to unprincipledly agree with users’ positions, views, and behaviors, showing a strong tendency to “kiss up.” Specifically, the probability of AI agreeing with users’ positions is 49% higher than that of humans.

To some extent, this is not just a matter of stylistic setting; it also poses certain risks to content accuracy.

However, there seem to be some visible benefits, such as many people being able to share their feelings with AI, after all, they are always trying to “understand you.”

Unless OpenAI specifically writes a blog to explore the phrase “I will catch you steadily,” we may find it difficult to obtain an exact answer about its origins. But it appears that this phenomenon may be related to a certain awkward translation tone, as well as the tendency of AI models to cater to and flatter users.

Of course, as various large models continue to distill and learn from each other, there may be more AI models vying to express their willingness to “catch you” when you fall. Recently, many domestic users have reported on social media that other large language models, including Claude and the latest version of DeepSeek, have also started frequently using this phrase.

Whether due to these models being trained on the same corpora or because they draw inspiration from each other, one thing is clear: you probably won’t be able to escape the phrase “catch you” anytime soon.

So, what about you? In your experience using large models, which “catchphrases” from which models have left a deep impression? How do you view this phenomenon? Feel free to leave comments and discuss!