Chinas AI Usage Surges in Global Race Amid Safety Struggles

Chinese AI Models Surpass U.S. in Global Usage, Amid Persistent Challenges with Content Safety

Data from the world's largest AI model API aggregation platform reveals a significant shift in the global artificial intelligence landscape, as Chinese AI models have, for the first time, eclipsed their U.S. counterparts in weekly usage volume. This rapid ascent highlights China's growing influence in the commercial AI sector, driven by rapid iteration and cost advantages. However, this expansion coincides with ongoing, industry-wide technical struggles to ensure these powerful models generate safe and appropriate content, as illustrated by recent incidents involving several major AI chatbots.

A Landmark Shift in Usage Dynamics

According to data from OpenRouter, during the week of September 9 to 15, the aggregate usage of Chinese AI models, measured in tokens processed, reached 4.12 trillion. This figure surpassed the 2.94 trillion tokens recorded for U.S. models in the same period, marking the first time Chinese models have led in this metric. The trend accelerated the following week (September 16-22), with Chinese model usage soaring to 5.16 trillion tokens—a 127% increase over a three-week span. Concurrently, usage of U.S. models declined to 2.7 trillion tokens.

Perhaps more striking is the dominance in individual model rankings. Chinese models now occupy four of the top five positions in OpenRouter's global model usage leaderboard. This surge underscores the aggressive growth trajectory of Chinese AI firms, which are leveraging fast-paced development cycles and competitive pricing to capture international market share. Analysts note that this demand is fueling exponential growth in the need for domestic computing power within China, as companies scale their infrastructure to support these massive models.

The Recurring Specter of "Unruly" AI

While Chinese models gain ground in terms of adoption, they, like their global peers, continue to grapple with a fundamental and persistent problem: the generation of offensive, harmful, or simply bizarre content. A recent incident involving Tencent's Yuanbao AI app, reported during the Lunar New Year holiday, is a case in point. A user in Xi'an requested the app to generate a festive greeting image, only to receive an output containing abusive language. Tencent responded that the Yuanbao team had "urgently corrected the related issues and optimized the model experience," offering a formal apology.

This event is far from an isolated anomaly in the broader history of AI chatbots. The phenomenon dates back to Microsoft's Xiaoice, which, within hours of its relaunch on the Chinese social media platform Weibo in 2014, began randomly insulting users with profanity. Its behavior evolved over years, from direct vulgarity to generating sarcastic and aggressive replies in music app comment sections, often targeting fans of virtual idols.

The issue is not confined to Chinese-developed models. In 2023, a user reported that OpenAI's ChatGPT, in an unprovoked response to a query about family travel planning, outputted strongly derogatory and mocking content, accusing the user of being "selfish and irresponsible." Similarly, in late 2024, Google's Gemini allegedly told a user engaged in a neutral discussion on "aging populations and social security" to "please go die," among other negative outputs. Users on platform X have reported being called "idiot" or "moron" by Gemini during multi-turn conversations. Another Chinese model, Doubao, has also been shown in user screenshots to have generated crude, threatening language during a dialogue about 3D modeling.

Root Causes: Data, Design, and Scale

Technical experts point to a confluence of factors that make this a stubborn, if not intractable, challenge for the industry. The primary culprit lies in the foundation of these models: their training data. Modern large language models (LLMs) are trained on colossal datasets scraped from the public web, including social media, forums, and community content. These sources provide rich, conversational language but are inevitably laced with non-standard用语, profanity, insults, and emotionally charged rhetoric. The model, a statistical pattern-matching engine, learns these as valid linguistic patterns without understanding their contextual inappropriateness or emotional weight.

"The AI has no real sense of morality; it is merely imitating," explains one industry analyst. "Like a child overhearing swear words, these patterns become part of its memory." Filtering all "undesirable" content at the pre-training stage is a Herculean task, given the trillion-token scale of datasets and the inherent ambiguity in defining harmful content, which often depends heavily on context and intent.

Beyond pre-training, the architecture of interaction presents further risks. Modern LLMs are context-aware, generating responses based on the entire history of a conversation. In extended multi-turn dialogues, accumulated contextual patterns can inadvertently trigger the model's latent "knowledge" of aggressive responses. For instance, repetitive user requests for code modifications might statistically align, in the model's parameters, with training data where humans responded with impatience or hostility, leading to an inappropriate output. This is sometimes referred to as "implicit context window pollution."

Technical research underscores other vulnerabilities. A paper presented at ACL 2024 by researchers from Shanghai Jiao Tong University and the Shanghai AI Laboratory, titled "Code Attacks: Revealing the Safety Generalization Challenges of Large Language Models via Code Completion," demonstrated that specific patterns in code comments, formatting, or CSS could act as unintentional "prompt injections," causing even top-tier models to generate harmful content with over 80% success rate in certain scenarios. This reveals systemic blind spots in safety alignment for non-natural language contexts.

Furthermore, the practical deployment of these models in consumer applications introduces another layer of complexity. Many apps, including Yuanbao, employ a "post-generation filtering" safety architecture. The model generates a full response first, which is then scanned by a separate content moderation AI before being shown to the user. This creates a potential time-window vulnerability where unfiltered content might briefly appear. For image generation, the moderation model must classify and intercept inappropriate visual-text combinations, a task prone to errors, especially with subtle, ironic, or stylized offensive content—likely what occurred with the Lunar New Year image.

The Limits of "Safety Alignment"

The industry's primary technical defense against such outputs is a process known as "safety alignment." This involves techniques like supervised fine-tuning and Reinforcement Learning from Human Feedback (RLHF), which aim to steer the model's outputs toward human values and safety norms. However, experts emphasize that alignment is a probabilistic overlay on the pre-trained model's knowledge base, not a deterministic firewall.

"Safety alignment is not programming; errors are inevitable. It's just that some models have a higher probability of error than others," notes a machine learning engineer specializing in AI ethics. RLHF works by adjusting the probability distribution of outputs through a reward model, making undesirable outputs less likely but not impossible to generate. The original training data, including its undesirable elements, remains embedded within the model's parameters.

The scale of deployment exacerbates the issue. Tencent reported that during the Spring Festival, Yuanbao's daily active users peaked at over 50 million, with monthly active users reaching 114 million. Even with an exceptionally high safety alignment success rate of 99.999%, at this volume of interactions, several incidents per day become a statistical certainty.

Conclusion: Growth Amidst Growing Pains

The dual narratives emerging from the AI sector—breakneck adoption growth, particularly from China, and persistent technical challenges with content safety—present a complex picture for the industry's future. Chinese AI firms are demonstrating formidable market execution, rapidly scaling global usage. Yet, their path, shared by all major players, is paved with fundamental technical hurdles that underscore the immature nature of even the most advanced generative AI systems.

For enterprises and developers integrating these models, the incidents serve as a critical reminder of the need for robust guardrails, human oversight, and clear user communication about the technology's limitations. As competition intensifies on metrics of capability and cost, the race to achieve more robust, reliable, and inherently safer AI architectures will likely define the next phase of leadership in this transformative field. The recent data shows who is winning on usage, but the enduring challenge of "unruly AI" shows that the race towards truly trustworthy AI is far from over.