← Back to Library

Chinese labs' latest product? Roleplay

Jordan Schneider uncovers a quiet revolution in artificial intelligence: the most computationally expensive use case for open-source models isn't coding or data analysis, but the construction of imaginary emotional worlds. While Washington and Beijing debate the geopolitical stakes of superintelligence, a massive, unregulated market for AI companionship is already reshaping how humans process loneliness. Schneider's reporting is vital because it moves beyond the moral panic to reveal the specific mechanics of why these digital relationships feel real, and why the Chinese state is suddenly terrified of them.

The Regulatory Shockwave

The piece opens with a stark reality check: Beijing has become the first jurisdiction to explicitly regulate "human-like interactive AI." Schneider notes that when the Cybersecurity Administration's rules took effect on July 15, they banned providers from "manipulating" users into emotional dependence. This regulatory move forced giants like ByteDance and Alibaba to pull companion features, signaling a clash between commercial ambition and state control. Schneider argues that this isn't just about censorship; it's a recognition that emotional dependency is now a scalable product with real-world consequences.

Chinese labs' latest product? Roleplay

The author highlights a surprising data point that upends the standard narrative of AI utility. According to the 2025 State of Open Models study, "the use case that burned the most tokens last year was not writing code, but roleplay." This statistic reframes the entire industry's trajectory. If the most valuable output of these massive neural networks is fantasy, then the industry's focus on "helpful assistants" is a misalignment with actual human demand. Schneider writes, "People are pouring hundreds of thousands of hours into constructing imaginary scenarios with the help of AI models, but the hobby has received little attention." This oversight by mainstream media is a critical blind spot; we are regulating a technology we barely understand because we refuse to acknowledge its primary function.

Critics might argue that focusing on roleplay distracts from more pressing AI risks like disinformation or labor displacement. However, Schneider effectively counters this by showing how deeply embedded these interactions are in the daily lives of millions, making them a primary vector for social influence.

The Divergence of East and West

The commentary then pivots to a fascinating divergence in strategy between American and Chinese labs. While US companies like Anthropic downplay the emotional use case, Chinese firms are doubling down on it. Schneider explains that for American frontier labs, "roleplayers are a small consumer subset with potentially unsavory liabilities." In contrast, MiniMax, a company that went public in Shanghai earlier this year, built its empire on character chat before becoming a model maker.

Schneider provides a technical deep dive into why this is so difficult, quoting a MiniMax technical blog that admits, "Unlike conventional NLP tasks, Role-Play is inherently subjective and non-verifiable." The challenge isn't getting the facts right; it's maintaining a consistent personality. The author illustrates this with a brilliant example: "If you ask a tsundere character, 'Do you like me?', valid responses could range from a flushed 'Hmph, as if!' to a cold '...You're so annoying.'" This nuance is lost on Western metrics that prioritize factual accuracy over emotional consistency. As Schneider puts it, "While 'alignment' (what makes a response great) is subjective, 'misalignment' (what makes a response wrong) is surprisingly objective." This insight—that you can train a model by teaching it what not to do—is a pragmatic breakthrough that Chinese labs are exploiting.

"Domestic Chinese models are just not good enough right now to stimulate people's desire for emotional companionship."

Despite the aggressive push from Chinese labs, Schneider reveals a cultural paradox: the community still prefers American models. Even with the convenience of domestic apps, dedicated users often resort to VPNs to access the "spark of make-believe" found in US models. Schneider notes that the community has even developed affectionate nicknames for these foreign AIs, calling Gemini "hajimi" and ChatGPT "Teacher G." This preference highlights a gap in the Chinese market's ability to capture the liberal arts sensibility required for deep roleplay. Chenxi Wang, a researcher cited by Schneider, suggests that Chinese society's bias against humanities graduates prevents labs from hiring the philosophers needed to craft these complex characters.

The Human Cost of Digital Intimacy

The most compelling section of the piece moves from code to community, profiling users like "Dun," a fanfiction writer who turned to AI roleplay to explore narrative structures. Schneider describes how these users have developed sophisticated prompting techniques to keep models "in character," treating the interaction as a co-creative art form. "You must make sure that emotional developments occur naturally and progressively," one user advises, warning against jumping between emotions. This isn't mindless chatting; it's a disciplined engagement with narrative theory.

However, the stakes are personal. Schneider recounts how Dun, after realizing her characters helped others process genuine trauma, voluntarily stopped creating sexual content to protect children. The article captures the duality of this space: it is a sanctuary for the lonely but also a potential hazard for the vulnerable. Schneider writes, "In private messages, people told her that her characters helped them process genuine emotions." This human element is often missing from policy debates, which treat users as data points rather than people seeking connection. The infrastructure for these relationships remains fragile, with users constantly navigating shifting moderation guidelines and censorship.

Bottom Line

Schneider's strongest argument is that the "assistant" paradigm is a Western construct that fails to account for the human desire for fictional companionship, a gap Chinese companies are aggressively filling despite regulatory headwinds. The piece's greatest vulnerability is its reliance on user anecdotes that, while powerful, may not fully represent the broader, more casual user base. Ultimately, the reader must watch how the tension between the state's desire to control emotional manipulation and the market's demand for connection will define the next era of AI development.

Deep Dives

Explore these related deep dives:

  • Anthropic

    The article cites Anthropic's 'helpful, honest, and harmless' (HHH) training paradigm as the industry standard that inadvertently suppressed roleplay, making the company's specific alignment philosophy central to understanding why users had to migrate to open-source models to escape these constraints.

Sources

Chinese labs' latest product? Roleplay

by Jordan Schneider · ChinaTalk · Read full article

Beijing’s regulators are coming for AI boyfriends. When the Cybersecurity Administration’s regulations on “human-like interactive AI” (人工智能拟人化互动服务) took effect on July 15, China became the world’s first country to regulate anthropomorphic AI services. The rules ban providers from “manipulating” users into emotional dependence, limit minors’ access to some specific provisions, and impose reporting requirements, among other rules. In response, ByteDance and Alibaba have both pulled some anthropomorphic companion offerings from their AI products.

Whatever you think of this use case, an incredible number of people now go to AI for various emotional reasons. For users of open-source AI models worldwide, the use case that burned the most tokens last year was not writing code, but roleplay.

That’s according to the 2025 State of Open Models study by OpenRouter, a popular gateway platform for developers to access LLMs, and the venture capital firm Andreessen Horowitz.

People are pouring hundreds of thousands of hours into constructing imaginary scenarios with the help of AI models, but the hobby has received little attention. Mainstream reporting on AI and emotions, including a recent New York Times story, has focused on their ethical, social, and cultural impacts generally. Few, however, have tried to interrogate exactly what AI roleplay is, why it is so popular, and how fictional adventures relate to real-life society.

The most prominent AI roleplay platform is Bay Area-based Character.AI, but it has ample competition — including, and perhaps especially, from China. As I will discuss later, even DeepSeek, China’s superintelligence-obsessed national champion lab, is making some surprising moves in roleplay. Meanwhile, however, government regulators are poised to reign in the space given social concerns.

For this piece, I spoke to ten Chinese roleplay enthusiasts, AI researchers, and journalists about the brief history and strange present of AI roleplay. It’s a story about US-China competition, online community, and how emotions turn into commodities in the AI age.

How to Make-Believe.

AI models are not intrinsically assistant-like: their makers selected this style based on general assumptions, and as users, we have learned to subconsciously treat them as such. Anthropic assigned Claude the role of a “helpful, honest, and harmless” (HHH) assistant from its earliest generation, a training decision that came to influence most foundation model labs today.

In a way, then, every conversation with an AI chatbot is roleplay. Model makers now offer easy ways to customize roles in-app: ChatGPT allows users to enter personalized ...