Why is ChatGPT Speaking Strange Chinese?

As artificial intelligence systems expand across languages and markets, researchers are increasingly finding that fluency does not always equal naturalness, and nowhere is that tension more visible than in Chinese, where ChatGPT’s conversational habits are beginning to look less like human speech and more like a strangely uniform, over-rehearsed script that has caught the attention of users, linguists and AI developers alike.

ChatGPT is producing unusual conversational patterns in Chinese, with users and researchers noting that its phrasing can sound unnatural, repetitive and at times oddly poetic.

A report by Wired examined how OpenAI’s ChatGPT handles Chinese, the world’s most spoken language by native speakers according to the Language School at Middlebury College.

One recurring expression highlighted in the report is “我会稳稳地接住你,” which translates literally to “I will catch you steadily.” The phrase is often used in emotional contexts, signaling reassurance or willingness to engage with someone’s feelings. Wired journalist Zeyi Yang notes that a more figurative translation could be “I’ll hold you steadily through whatever comes,” though many Chinese speakers reportedly find the phrasing unnatural or irritating in everyday conversation.

In other cases, ChatGPT has been observed using “砍一刀,” which can mean “help me cut it once” or “slash the price.” The phrase is associated with aggressive promotional language used by Chinese e-commerce platforms such as Pinduoduo, and appears in chatbot responses in ways that feel like copied advertising speech rather than natural dialogue, according to Wired.

These quirks have become widely discussed among Chinese internet users, where ChatGPT is sometimes portrayed in memes as a large inflatable airbag designed to catch people as they fall, reflecting its repeated “I will catch you steadily” phrasing.

Experts suggest the behavior may be linked to a phenomenon known as “mode collapse,” which affects large language models during training. The idea is that human data annotators who refine AI outputs may unintentionally favor familiar or culturally dominant expressions, while less familiar phrasing is filtered out or underrepresented.

Once a model is trained, correcting or removing these patterns becomes difficult. Developers can reinforce certain responses as preferred, but controlling the balance between frequency, variation, and contextual appropriateness remains a challenge.

“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,” Max Spero, cofounder and chief executive of AI-writing detector Pangram, told Wired.

The issue highlights a broader challenge in artificial intelligence development, where improving fluency in one context can unintentionally produce unnatural or repetitive behavior in another language environment.

Despite these quirks, Chinese remains one of the most important testing grounds for global AI systems, given its scale, complexity and diversity of usage patterns.

Stay ahead of the stories shaping our world. Subscribe to Impact Newswire for timely, curated insights on global tech, business, and innovation all in one place.

Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.

Got a story to share? Pitch it to us at info@impactnews-wire.com and reach the right audience worldwide