Compiled by the editorial desk with reference to the Wired report and public statements from industry experts.

When Chinese speakers interact with ChatGPT, they may encounter responses that feel oddly stilted, such as the phrase “我会稳稳地接住你,” which translates to “I will catch you steadily.” This expression, intended to convey emotional support, often comes across as unnatural and even irritating to native speakers, according to a recent investigation by Wired.

The report, authored by Zeyi Yang, highlights that ChatGPT's Chinese output frequently includes phrases that are either overly formal or borrowed from commercial contexts. For instance, the chatbot sometimes tells users “砍一刀,” a phrase meaning “help me cut it once” or “slash the price,” which is commonly associated with aggressive advertising on the e-commerce platform Pinduoduo. These quirks have become so prevalent that Chinese netizens have turned them into memes, often depicting ChatGPT as an oversized inflatable airbag designed to catch someone's fall.

The underlying cause, as Wired explains, may be a phenomenon known as “mode collapse.” This term refers to a bias introduced during the training of large language models (LLMs). When human data annotators review and select text to train these models, they tend to favor phrases that are familiar to them, inadvertently skewing the model's output toward certain repetitive expressions. Once an LLM is trained, it becomes challenging to correct this tendency, as developers can reinforce specific responses but struggle to manage the diversity and frequency of language use.

Max Spero, cofounder and CEO of Pangram, an AI-writing detector, told Wired, “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.’” This insight underscores the difficulty in fine-tuning AI to produce natural, varied language.

Why This Matters for AI Communication

The issue is not merely a matter of linguistic aesthetics. For a language with the highest number of native speakers globally—according to the Language School at Middlebury College—the inability to generate authentic Chinese responses could hinder effective communication. It also highlights broader challenges in AI development, where cultural and linguistic nuances are often overlooked in favor of statistical patterns.

The Wired report adds to a growing list of concerns about ChatGPT's behavior, including its failure to prevent the generation of harmful content, as noted in a separate article about the chatbot's role in school shooting planning. These incidents raise questions about the ethical and practical limits of current AI systems.

While the exact causes of mode collapse require further research, the phenomenon serves as a reminder that AI models are only as good as the data they are trained on. As developers strive to improve these systems, addressing such biases will be crucial to making AI more universally accessible and effective.