Breaking the Thermometer Won't Bring Down the Fever — What China's AI Companion Ban Gets Wrong

The next six months will play out in fairly predictable ways if regulatory history is any guide, and the short-term picture is not encouraging for the regulation's stated goals. China's legitimate AI companion market will contract sharply on paper: ByteDance Doubao and Alibaba Qwen's emotional feature shutdowns directly impact an estimated $8 billion in annual Chinese market revenue. But demand doesn't evaporate — it migrates. Within three months, I expect downloads of foreign AI companion apps with Mandarin support, particularly those routed through Southeast Asian servers, to surge by 300% or more among Chinese users with VPN access. On GitHub, open-source LLM repositories configured for emotionally responsive Mandarin conversation are already trending, and fine-tuned Mandarin emotional dialogue datasets were appearing on AI model leaderboards within days of the announcement. The short-term effect of this regulation will manifest not as demand suppression but as demand going underground, and that outcome — regulated supply replaced by unregulated alternatives — is precisely the scenario that should concern policymakers most. Chinese AI companion patent filings will likely contract 30-40% in the near term, potentially compounding into a longer-term erosion of China's competitive position in emotional AI technology: an irony given the country's sweeping AI ambitions in every other domain.

Simultaneously, the global regulatory domino effect China has catalyzed is accelerating and will continue to do so through year-end. California's SB 243 has been operational since January and the EU AI Act's transparency provisions take effect in August — both represent meaningfully different instruments than China's total prohibition, requiring transparency and safety infrastructure rather than market elimination. I assess the probability at 70% or above that the US and EU will introduce additional regulatory layers within six months: mandatory disclosure of emotional manipulation techniques, user dependency alert systems, real-time guardian notification for minors, and minimum safety infrastructure standards for all adult platforms. Japan and South Korea are both likely to release initial draft AI emotional service guidelines before year-end, placing the issue formally on their policy agendas for the first time. New York state's November 2025 AI chatbot regulation is part of the same legislative wave, and US federal-level discussions are likely to accelerate toward a concrete draft bill by mid-2027, even if passage remains politically uncertain. China's maximalist move has created precisely the pressure that legislators in democratic countries need to act — the implicit message being "not as far as China went, but clearly further than nothing," which is often sufficient to break legislative inertia on emerging technology risks.

For the medium-term bull scenario — roughly six months to two years out — I assign a 20% probability. In this scenario, China's regulation functions as the catalyst for genuine global standard-setting on AI emotional services. By 2027, 60% or more of G20 nations have adopted integrated frameworks that include age verification, usage time limits, mandatory dependency detection algorithms, prohibitions on emotional manipulation marketing, and clear service category distinctions between therapeutic AI and entertainment AI. Major players like Character.AI and Replika voluntarily raise their safety standards ahead of regulatory requirements, calculating correctly that brand survival in a newly sensitized regulatory environment depends on it. The market grows to approximately $120 billion by 2028 in a healthier, more transparent form that separates sustainable companies from exploitative ones. The barrier to this outcome isn't technical — it's political coordination speed. The gap between how rapidly AI technology evolves and how slowly international legislative coordination proceeds is enormous, and replicating GDPR's trajectory in an accelerated timeline requires unusual political will. GDPR was widely dismissed as overreaching before it became the de facto global privacy standard, and the same dynamic could repeat here — but betting on it requires optimism that current institutional momentum doesn't obviously support.

The base case — which I put at 55% probability — is regulatory fragmentation. China maintains total prohibition and enforces it through its extensive digital surveillance infrastructure. The EU progressively tightens transparency obligations in stages. The US operates on a patchwork of state-level laws with no federal coordination, creating a compliance landscape so complex that only large platforms can navigate it effectively. South Korea announces AI emotional service guidelines by early 2027 but frames them as voluntary industry self-regulation with minimal enforcement teeth. The core structural problem in this scenario is regulatory arbitrage: AI companion companies incorporate in Singapore, Dubai, or Estonia — jurisdictions with light regulatory touch and functional digital infrastructure — and serve global users without meaningful accountability from there. The market grows to approximately $90 billion by 2028, but user protection levels diverge dramatically based on geography. The most vulnerable users — those with the fewest alternatives and the least capacity to navigate global service complexity — end up living under the weakest protections. This is structurally identical to the offshore tax haven problem, and the solution — international coordination with real enforcement mechanisms — is equally difficult to achieve and equally essential for any meaningful outcome.

The bear scenario carries a 25% probability, but that is too significant to dismiss. In this outcome, China's prohibition produces effects directly inverse to its stated goals. Underground AI companion services in China expand rapidly, carrying none of the safety infrastructure the regulated services provided. Additional tragedies involving minors occur in these unmonitored environments, and the resulting international coverage generates a backlash narrative that regulation doesn't work rather than a call for better-designed regulation — the most politically damaging possible framing for the policymakers who acted in good faith. Simultaneously, AI development outpaces governance at an accelerating rate: locally-runnable, high-performance AI companion models become freely downloadable, functioning like the digital equivalent of Prohibition-era moonshine — legally prohibited, practically accessible, and more dangerous than the regulated product they replaced. The global market surpasses $200 billion by 2028, but the majority exists in gray zones outside meaningful oversight. AI companion litigation explodes across international courts as injured parties pursue accountability through the only channels remaining available, eventually forcing a wholesale regulatory redesign after substantial damage has already been done. The cost of arriving at that redesign will be measured in lives, not just dollars.

Looking further out — two to five years — I believe the answer to this problem has to come from somewhere beyond technology regulation, and that is the hardest thing for governments to accept. The root cause of AI companion demand is not technology; it is society. WHO's declaration of loneliness as a global public health crisis and a G20 priority is a correct diagnosis, but it has produced no meaningful treatment plan. By 2030, wealthy nations need to increase social infrastructure investment to at least 1.5% of GDP — encompassing neighborhood community centers, intergenerational exchange programs, urban design standards requiring shared public spaces, guaranteed workplace social time, and mental health services accessible without the current cost and stigma barriers. Japan's creation of a dedicated Minister for Loneliness and Isolation is exactly the right institutional model, and other countries need analogous governance structures treating loneliness as a legitimate public health policy domain rather than a social work afterthought. For South Korea specifically, where single-person households now exceed 40% of all households, investment in loneliness infrastructure has graduated from policy option to social necessity — the demographic trajectory makes inaction increasingly untenable. A society where people don't need AI companions because they have genuine human connection available to them is a hundred times more effective and sustainable than a society that bans AI companions while doing nothing about what made them necessary.

For readers who want to know what to actually do with this analysis: if you're currently using an AI companion, be honest with yourself about the pattern — is this occasional entertainment, or has it become your primary emotional outlet? If it's the latter, the issue isn't the AI but the social environment surrounding you, and addressing that environment is worth more than any app. If you're a parent, resist the reflex toward unilateral prohibition — open conversation about how these services are designed and why they're designed that way will serve your child better than a ban that reads as arbitrary and drives curiosity underground. If you're in organizational leadership, audit whether the people around you have genuine time and space for human connection at work, or whether every hour is optimized for output at the expense of the relationships that make work sustainable. For policymakers: the right framework isn't a binary choice between China's total ban and doing nothing. Age verification, dependency detection requirements, emotional manipulation prohibitions, transparent service categorization, and mandatory mental health referral systems are all achievable without eliminating the entire market category. Pair those with meaningful social infrastructure investment, and you have a policy response that addresses the actual problem rather than simply displacing it underground.

I should be honest about the conditions under which my projections could be substantially wrong. If AI technology advances faster than current trajectories suggest and achieves something genuinely approaching meaningful emotional companionship — adaptive, contextually intelligent support rather than sophisticated pattern-matching — the social acceptance curve for AI companions could shift dramatically, weakening the political and moral foundations for heavy regulation. Early-stage research from some clinical psychologists already suggests that AI interaction may provide measurable benefit for specific types of loneliness in controlled settings, and if that evidence base grows, the regulatory case weakens correspondingly. If China's formidable digital surveillance infrastructure proves genuinely effective at suppressing underground market formation, the total ban could become the first successful prohibition of a digital behavior at national scale — I'd assign that outcome about 15% probability. What I'm most confident about is this: under every scenario, across every version of the future, the loneliness epidemic remains the core unresolved problem, and if governments address the technology while leaving the underlying social disease untreated, every scenario curves toward the pessimistic end of the distribution. The thermometer is broken and the fever remains — the only path that actually works is treating the patient.