AI roleplay is the practice of using a language model to sustain a fictional character, world, or relationship across a long conversation, and in 2025 it became the single largest use case for open-source AI models, ahead of coding. OpenRouter and Andreessen Horowitz’s 2025 State of Open Models study found creative roleplay accounted for more than 50% of all tokens processed on open-weight models.
Character.AI alone pulls 194 million monthly visits and 20 million monthly active users who average two hours a day inside character-driven conversations. China’s AI labs, not Silicon Valley, are the ones building products around this. MiniMax IPO’d on the back of a character-chat business. DeepSeek is hiring product managers for « roleplay and emotional companionship. »
Here is the hard truth: American frontier labs are treating a massive, monetizable use case as a liability to manage, while Chinese labs are treating it as a product line to own.
The Build Order Nobody Saw Coming
In Warcraft, you do not win by rushing the enemy base with your starting peasants. You build an economy first, gold mines, farms, barracks, then attack once your army can hold ground.
DeepSeek is running that build order in AI. Its labs are obsessed with superintelligence, but its consumer product has no scalable moat. So instead of rushing straight at OpenAI and Anthropic with another chatbot, DeepSeek quietly built the economy layer: roleplay.
It is not a distraction from the mission. It is the resource base that funds and retains users while the real fight for frontier models continues elsewhere.
AI Roleplay Is Not a Niche. It Is the Base Layer
Every conversation with a chatbot is technically roleplay, because model makers chose an « assistant » persona for you before you typed a word.
Anthropic trained Claude to be a « helpful, honest, and harmless » assistant from its earliest generation, a design decision most foundation model labs quietly copied. That framing is a costume, not a neutral default. Users just stopped noticing the seams.
Serious roleplayers go further. They fine-tune open models, feed them entire novels, and build custom system prompts that turn a general-purpose LLM into a consistent fictional character across hundreds of turns. Longer context windows and better retrieval have made this dramatically more convincing over the past two years. A model that once forgot a character’s backstory after ten messages can now sustain one for an entire session.
The result: roleplay is not a fringe hobby bolted onto AI. It is one of the base layers of how people actually use these systems, right alongside coding and search.
Communities have formed around this skill the way fanfiction forums formed around shared universes. Prompt writers on Xiaohongshu publish detailed character backstories, personality quirks, and even stat-tracking systems that turn a chat window into something closer to a homemade visual-novel game. One writer built a system where her AI character has its own home and deducts real money from her account every time it « eats. » That level of craft is not casual. It is a skill people spend years refining, the same way a novelist refines dialogue.
The Numbers Nobody in the West Is Tracking
Creative roleplay and coding assistance are the two dominant use cases on OpenRouter’s platform, per its 2025 study with a16z. Roleplay hovers above 50% of all open-source tokens processed.
Character.AI draws 194 million monthly visits and 20 million monthly active users, with average session time around two hours a day. The platform hosts more than 18 million user-created characters.
Chinese open-weight models jumped from roughly 1.2% of weekly global token share in late 2024 to peaks near 30% in 2025, averaging about 13% for the year, per the same OpenRouter study. Roleplay usage is a meaningful driver of that share gain, not an afterthought.
Anthropic’s own 2025 research found that only 2.9% of Claude.ai conversations involve emotional subjects at all. Less than 0.5% touch companionship or roleplay specifically. That gap between what American labs measure and what open-source usage actually shows is the story.
Why China’s AI Labs Went All In on Roleplay
MiniMax did not stumble into roleplay. It started as a character-chat company, Xingye/Talkie, before it became a foundation model lab, and that product line remains its largest revenue source. When MiniMax IPO’d on the Shanghai Stock Exchange in early 2026, it launched a model built specifically for roleplay: MiniMax-M2-her.
MiniMax’s technical team identified something useful for anyone building consumer AI products: alignment in roleplay is subjective, but misalignment is objective. If a tsundere character responds « Yes, I really like you » to a hostile question, that is an unambiguous, flaggable failure.
You cannot easily define the perfect response, but you can reliably catch a broken one. That insight let MiniMax build an evaluation pipeline around roleplay quality instead of treating it as unmeasurable.
Cultural context matters too. ACG subcultures, anime, comics, and games, sit closer to the mainstream in China and East Asia than in the West. Chenxi Wang, a researcher who has worked on anthropomorphization at Alibaba’s Qwen, told ChinaTalk that Chinese labs are structurally under-resourced for this work.
Her verdict on domestic models: « not good enough right now to stimulate people’s desire for emotional companionship. » That gap is exactly why dedicated Chinese roleplayers still reach for American frontier models through VPNs, nicknaming Claude « Little Ke, » Gemini « hajimi, » and ChatGPT « Teacher G. »
DeepSeek’s Pragmatic Pivot: Roleplay as a Consumer Moat
DeepSeek researcher Deli Chen published custom roleplay instructions for DeepSeek-V4 on GitHub within weeks of the model’s release, then solicited feedback directly from hundreds of roleplayers on Xiaohongshu. That is not a side project.
In May 2026, DeepSeek’s HR team posted a hiring notice for a product manager specializing in « character roleplay and emotional companionship. » The role requires, in the job’s own words, « sharp literary taste. »
Read against DeepSeek’s financial pressure, talent attrition, and the muted market reaction to V4, roleplay looks like a deliberate answer to a real gap: DeepSeek has never had a scalable consumer product. Its models expose their full chain-of-thought to users, and roleplayers report that visible reasoning makes the fictional experience more immersive, not less.
A side effect of a research choice became a competitive advantage nobody planned for.
The Regulatory Reckoning: Beijing Draws a Line
On July 15, 2026, China became the first country to regulate anthropomorphic AI directly, when the Cyberspace Administration’s interim measures on « human-like interactive AI » took effect. The rules ban providers from manipulating users into emotional dependence. They also block virtual « relative » or « partner » relationships for minors and mandate addiction-detection mechanisms.
ByteDance and Alibaba responded fast, pulling anthropomorphic companion features from Doubao and Qwen. That cut off personalized AI agents used by hundreds of millions of people, per TechTimes and IAPP reporting.
The rules leave an apparent carveout for fictional roleplay that is not framed as a simulated relationship, but enforcement details remain vague. Chenxi Wang’s read from inside Chinese labs: « no one cares, » because current domestic models do not feel human enough yet to trigger real alarm. Roleplayers on the ground disagree and expect the crackdown to widen, pointing to state media exposés of pornographic AI chat as the pretext regulators will use next.
What Western Labs Are Missing
American frontier labs treat emotional and roleplay usage as a small, risky edge case to disclose and move past. Anthropic’s own framing splits roleplay into « romantic » and « sexual » categories, with no room for interactive fiction, tabletop-style world-building, or the visual-novel-style stat-tracking systems Chinese roleplayers build by hand.
That framing undercounts the use case and misses the business opportunity sitting inside it. XiaoIce, Microsoft Research Asia’s chatbot released in 2014, reached more than 660 million users across Asia by 2018 by optimizing for one metric: conversation-turns per session, not task accuracy. It proved emotional engagement could scale to hundreds of millions of users almost a decade before ChatGPT existed.
It also proved the ceiling. Without differentiated content like gaming or character roleplay, pure emotional companionship is too cheap to defend, because anyone can smooth-talk for free. MiniMax and DeepSeek learned that lesson. Western labs are still treating the whole category as a compliance footnote.
If your company is building AI-driven engagement, retention, or premium subscription products and still modeling « emotional use cases » as noise in the data, you are underpricing your own opportunity cost. The lesson from XiaoIce, MiniMax, and DeepSeek is not « build a companion app. » It is that consumer AI retention and monetization live inside identity, personalization, and sustained character, not inside raw task completion.
That is the same build-order logic behind a well-run AI adoption sprint: get the foundational, high-engagement layer right before you chase the next frontier feature. If your team is still bolting AI onto existing workflows instead of rebuilding around what actually drives sustained usage, that gap is worth closing before a competitor closes it for you. Asymmetriq’s managed AI implementation and the four-session Claude Sprint exist for exactly that gap: turning scattered AI usage into a system with a measurable retention layer.
AI Roleplay Platforms Compared
| Platform | Openness | Moderation | Cost | Best For |
|---|---|---|---|---|
| Character.AI | Closed, proprietary | Strict, teen-focused | Free + subscription | Casual users, mobile-first roleplay |
| SillyTavern | Open-source frontend | User-controlled | Free (bring your own model) | Technical users wanting full control |
| MiniMax Xingye/Talkie | Closed, China-based | Regulated under CAC rules | Free + in-app purchases | Mass-market character chat in China |
| DeepSeek + custom prompts | Open weights | Minimal, community-built | Free to low-cost API | Power users chasing frontier-model depth |
| Claude / ChatGPT | Closed, general assistant | Conservative by design | Subscription | Highest narrative quality, least built for it |
Is AI Roleplay Actually Worth Anyone’s Attention?
Most people dismiss AI roleplay as a fringe hobby for teenagers and lonely adults. The data says otherwise: it is the largest single use case on open-source models and a core revenue driver for at least one newly public AI company.
Ignoring it means ignoring where a meaningful share of AI usage, and AI revenue, already lives.
Why Do Most Companies Get AI Personalization Wrong?
Most companies treat personalization as a UI feature, a name field and a tone toggle, instead of a sustained-character problem. Roleplay platforms show that real personalization requires state, memory, and consistency across many sessions, which is a much harder engineering and product problem than most companies attempt. MiniMax’s own research on this is instructive: the hard part is not defining a great response, it is catching a broken one before the user notices. Most business AI deployments never build that detection layer at all, which is why the personalization feels shallow within a few sessions.
Is China Really Ahead of the US on AI Roleplay?
China is ahead on product execution and monetization for roleplay specifically, evidenced by MiniMax’s IPO and DeepSeek’s hiring moves. The US still leads on raw model quality, which is why dedicated Chinese roleplayers pay for VPN access to use Claude and ChatGPT despite having free domestic alternatives.
Will China’s New AI Companion Law Kill the Roleplay Market?
Unlikely, but it will reshape it. The July 2026 rules target manipulation, minors’ access, and virtual « relationship » framing specifically, not fictional roleplay as a category. Expect platforms to narrow what counts as a compliant character rather than shut down entirely.
What Made XiaoIce Fail Where MiniMax Is Succeeding?
XiaoIce optimized purely for conversation length with no differentiated content layer, which made it too easy to copy and too cheap to defend. MiniMax and DeepSeek pair emotional engagement with character depth, world-building, and, in DeepSeek’s case, visible reasoning, giving users a reason to stay that a generic chatbot cannot replicate.
Should a Business Outside Consumer AI Care About Any of This?
Yes, because the underlying lesson is about retention, not romance. Any company trying to build sustained AI-driven engagement, internal tools included, is solving the same problem roleplay platforms solved first: how to make a system worth returning to every day.
The Verdict
AI roleplay is not the embarrassing corner of the AI industry. It is the most honest signal available of what keeps people opening an AI product every single day, and China’s labs are the only ones currently building a business around that signal instead of apologizing for it.
If you are still measuring AI success by task completion and ignoring engagement depth, you are optimizing for the wrong metric. Book the Claude Sprint at Asymmetriq if you want your AI systems built around what actually drives retention, not around what looks safe in a press release.