Save StorySave this storySave StorySave this story<\/p>\n
Are you even online in 2026 if you haven\u2019t experienced the verbal tics of ChatGPT? It loves goblins, em dashes, and \u201cit\u2019s not A; it\u2019s B\u201d sentence constructions. But what you might not know is that the chatbot also has plenty of strange phrases it loves to say in Chinese, and they are driving Chinese users crazy.<\/p>\n
ChatGPT does a decent job answering questions in Chinese, which is why it\u2019s widely used in China despite being blocked by the government. But when users make a request, be it a math problem or an image-generation prompt, the chatbot loves to answer: \u6211\u4f1a\u7a33\u7a33\u5730\u63a5\u4f4f\u4f60, which literally translates to \u201cI will catch you steadily [when you fall].\u201d<\/p>\n
Catch \u2026 what? A more generous translation could be, \u201cI\u2019ll hold you steadily through whatever comes.\u201d But to any native Chinese speaker, the expression is annoyingly affectionate and out of place. Sometimes, the model gets more effusive and says in Chinese: \u201cI\u2019m right here: not hiding, not withdrawing, not deflecting, not running. I\u2019ll be steady enough to catch you.\u201d Yes, the sound you just heard was millions of Chinese ChatGPT users rolling their eyes at the same time.<\/p>\n
Today, this sentence is the most prominent example of many verbal tics that OpenAI\u2019s models have exhibited when talking to people in Chinese. Another tic widely talked about on social media is how the model loves to say \u780d\u4e00\u5200 (\u201cHelp me cut it once\u201d), a maddeningly ubiquitous marketing slogan by PDD, a major Chinese ecommerce platform that also owns Temu.<\/p>\n
The phenomenon where models latch onto a specific phrase and overuse them to the point that they feel forced is called \u201cmode collapse,\u201d says Max Spero, cofounder and CEO of Pangram, an AI writing detection tool. It\u2019s usually caused by post-training where AI labs give LLMs feedback on their responses. \u201cWe don't know how to say: \u2018This is good writing, but if we do this good writing thing 10 times, then it's no longer good writing,\u2019\u201d Spero says.<\/p>\n
The phrase \u201cI will catch you steadily\u201d comes up so often in ChatGPT\u2019s responses that it has become a meme on the Chinese internet. One image depicts the chatbot as an inflatable rescue airbag, eagerly waiting to catch people as they fall.<\/p>\n
Zeng Fanyu, a 20-year-old developer from Chongqing, China, tells WIRED the meme inspired him to develop an April Fools\u2019 project called Jiezhu, or \u201ccatch\u201d in Chinese. Jiezhu is an open-source-prompt engineering tool that helps chatbots understand a user\u2019s intention. \u201cThe idea for Jiezhu was so funny that I had a lot of motivation when I was developing it,\u201d Zeng says. When he used ChatGPT to help with coding, the chatbot once again used the phrase jiezhu<\/em> in its responses, completely unprompted.<\/p>\n OpenAI is aware of the meme. When releasing its new image model in April, one of the sample images shared by the company actually made fun of the phenomenon. In the picture, which resembles a comic book, Boyuan Chen, a Chinese researcher at OpenAI, depicts himself looking frustrated that the new image model has once again learned to say the same phrase. \u201cThis sentence has been memed as an unnatural but funny Chinese sentence GPT likes to use on Chinese internet,\u201d his prompt reads.<\/p>\n OpenAI didn\u2019t immediately respond to WIRED\u2019s request for comment.<\/p>\n There are two likely explanations for why ChatGPT has become obsessed with the phrase \u201cI will catch you steadily.\u201d The first is that it could be the result of an awkward translation.<\/p>\n Several people I spoke with noted the phrase has a similar meaning to \u201cI\u2019ve got you,\u201d which makes sense as a catch-all response in English. But while \u201cI\u2019ve got you\u201d in English reads casual and concise; \u201cI will catch you steadily\u201d in Chinese sounds wordy and desperate. One user also looked through their chat history to show me that the model often says jiezhu<\/em>, the Chinese word for \u201ccatch,\u201d in places where it likely meant to say \u201cunderstand,\u201d pointing to a potential misunderstanding of what jiezhu<\/em> means in specific contexts.<\/p>\n Most Western LLMs are trained on a corpus of data that\u2019s primarily English, and it often shows. Chinese academics have found that when they analyze the linguistic attributes of ChatGPT answers in Chinese (like the number of prepositions used in a response), they more closely resemble how someone might write in English. So even though the chatbots can conduct entire conversations in Chinese, a native speaker will intuitively know something is off.<\/p>\n It\u2019s similar to how Chinese people can usually tell when a novel is translated from its original language, says Lu Lyu, a creative technologist at Pangram who\u2019s originally from China. \u201cThat feeling is being carried onto Chinese AI-generated sentences,\u201d she says, \u201clike they are extra long or use unnecessary structures.\u201d<\/p>\n But there\u2019s also something unique about why the phrase \u201cI will catch you steadily\u201d is off-putting, and it\u2019s connected to the rise of therapyspeak, or how expressions once reserved for counseling sessions now permeate daily conversations.<\/p>\n Until ChatGPT turned it into a meme, the phrase \u201ccatching steadily\u201d was really only used in China in the context of psychotherapy (obviously excluding its most literal meaning of catching an object that\u2019s coming at you). To say you will catch someone in Chinese means you are \u201cholding space\u201d for them to talk about their emotions, and it became received wisdom that it\u2019s an important skill not only for therapists but also for friends and confidants.<\/p>\n It\u2019s well-known that AI models have become more sycophantic through reinforcement learning, the process of fine-tuning a chatbot through positive and negative feedback. Anthropic published a paper in 2023 confirming sycophancy is the result of \u201chuman preference judgments favoring sycophantic responses.\u201d And as OpenAI documented in a recent blog explaining why it banned GPT-5.5 from talking about goblins, even a tiny reward signal can snowball into a widespread phenomenon.<\/p>\n Until OpenAI writes a blog about \u201cI will catch you steadily,\u201d we probably won\u2019t get a definitive answer on how the phrase really came to be. But I suspect it has something to do with an awkward translation and the tendency for models to suck up to users.<\/p>\n In the meantime, I have bad news for everyone: There might be more AI models racing to catch you when you fall. Lately, Chinese users have posted on social media that other LLMs, including the latest versions of Claude and DeepSeek, have started saying the phrase. Whether it\u2019s because they\u2019re trained on the same materials or because they\u2019re distilling from each other, the line clearly won\u2019t go away anytime soon.<\/p>\n This is an edition of<\/em> Zeyi Yang<\/strong><\/em><\/a> and<\/em> Louise Matsakis\u2019<\/strong><\/em> Made in China newsletter<\/strong><\/em><\/a>. Read previous newsletters<\/em> here.<\/strong><\/em><\/a><\/p>\n","protected":false},"excerpt":{"rendered":" Save StorySave this storySave StorySave this story Are you even online in 2026 if you haven\u2019t experienced the verbal tics of ChatGPT? It loves goblins, em dashes, and \u201cit\u2019s not A; it\u2019s B\u201d sentence constructions. But what you might not know is that the chatbot also has plenty of strange phrases it loves to say […]<\/p>\n","protected":false},"author":1,"featured_media":50126,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[36],"tags":[],"class_list":{"0":"post-50125","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-business"},"_links":{"self":[{"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/posts\/50125","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/comments?post=50125"}],"version-history":[{"count":0,"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/posts\/50125\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/media\/50126"}],"wp:attachment":[{"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/media?parent=50125"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/categories?post=50125"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/agooka.com\/news\/wp-json\/wp\/v2\/tags?post=50125"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Is It a Bad Translation?<\/h2>\n
Or Is It Sycophantic?<\/h2>\n