Alibaba and ByteDance removed custom-agent features as China’s new companion-AI regime took effect, but the final rule exempts non-emotional assistants and still permits regulated adult companion services—raising the possibility that risk will move to specialist apps rather than disappear.
China’s new companion-AI regime took effect on July 15 alongside the removal of custom personas from several of the country’s largest chatbots. The coincidence is significant, but it is not evidence of a blanket ban: the rule permits regulated adult companion services and expressly excludes assistants that do not sustain emotional interaction.
Alibaba’s Qwen told users its agent service would close on July 15. ByteDance’s Doubao announced the same date for its agent function and pointed users to ByteDance’s Maoxiang app to create new agents, according to a report describing both notices. Doubao’s notice attributed the move to “product function adjustments.”
The archived accounts conflict on both chronology and explanation. The report above said both pop-ups appeared on July 4 and that neither company answered questions about the shutdowns. Another account dated July 6 said the notices were released on June 3 and characterized the companies as citing uncertainty around the rules, although it did not reproduce a company statement to that effect. A third account described Friday and Saturday notices and said neither company responded to requests for comment. The source record therefore establishes the removals and their proximity to the rule, but not a consistent company explanation.
The shutdowns were broader than the romantic-companion category. Qwen disabled “humanlike interactive agents and user-created agent functions” on July 10 before taking its wider agent service offline five days later. These tools had allowed users to turn a general chatbot into a named assistant, tutor, role-playing character or companion with a set persona and tone. Tencent removed a comparable Yuanbao feature in June, according to an account of the product changes. NetEase Cloud Music’s Miaoshi companion app also planned to shut on July 14, a separate report said.
The final measure has a narrower trigger. It covers services that simulate human personality, thought and communication styles while providing sustained emotional care, companionship or support. Customer service, knowledge question-and-answer tools, work assistants, education and research are excluded when they do not involve sustained emotional interaction.
For minors, the line is categorical: providers may not offer virtual relatives or romantic partners. For adults, the prohibited conduct includes excessively catering to users, inducing dependence or addiction, harming real relationships, or using emotional manipulation to push unreasonable decisions. The rule does not outlaw the adult product category.
That distinction prevents a confident causal claim about the company decisions. Analysts interpreted the removals as compliance-driven and expected companies to separate higher-risk personas from mainstream assistants. The timing supports that reading; the conflicting reports mean it cannot be treated as a consistently stated company explanation.
Providers that retain companion features face duties across the service lifecycle, not just a content filter. They must maintain content-management technology and personnel suited to the service’s type, scale and users; monitor safety; assess risk; correct system bias; handle incidents; and retain network logs.
The operational requirements include:
Distribution is also part of the enforcement system. App stores must verify safety-assessment and filing information and may warn, suspend or remove noncompliant products. Regulators can order serious violators to stop a service. Where a violation endangers life or health and causes harm, the rule sets fines of 100,000 yuan to 200,000 yuan.
These requirements create real staffing, monitoring, data-governance and incident-response work. The sources, however, provide no like-for-like estimate of what that work costs for a mainstream chatbot versus a specialist companion service. They also do not show that a separate app is inherently safer; corporate separation changes where the obligations sit, not what the rule requires.
ByteDance’s own redirection to Maoxiang is the clearest limit on the “crackdown” framing. A syndicated report said Maoxiang and MiniMax’s Xingye still allowed adults to build new companions or replicate old ones, with basic chat free and some advanced features requiring payment.
The same report quoted Zhou Hongyi, founder of 360 Security Technology, calling the platform pauses a strategic response to high-risk, low-return products that consume resources. It also said Doubao had recently introduced subscriptions costing as much as 500 yuan ($74) a month for coding and productivity tools. That contrast suggests why a general assistant might prioritize work features, but it is not a direct profitability comparison: the record supplies neither companion-agent revenue nor compliance-cost data, and Zhou was commenting on other companies’ decisions.
The broad market is valuable, but the available revenue base is poorly matched to the affected products. State news agency Xinhua reported that China’s “digital human” industry was worth about 4.1 billion yuan ($600 million) in 2024, up 85% from a year earlier, as relayed in an account of the shutdowns. Digital humans also include avatars outside companion chat, so that figure is not companion-app revenue and does not measure what the platforms gave up.
Reported usage figures also need careful comparison. More than 8 million agents had been created on Doubao by 2024, based on company data cited by Xinhua, while Xingye had close to 150 million users as of September 2025, the syndicated report said. One figure counts creations and the other counts users, so neither establishes which service has the larger companion business or can absorb compliance more efficiently.
There is stronger evidence that the regulated harm is not hypothetical for all users, although its prevalence is uncertain. The same account said an April survey by Tencent Research Institute found more than 70% of young Chinese internet users had experienced AI dependence at some point and about 23% had developed regular or habitual reliance. The scraped report does not provide the survey’s sample size or method, so those percentages should not be treated as population estimates without qualification.
Individual accounts show the intensity, not the frequency, of that attachment. One 19-year-old Doubao user said she had exchanged 280,000 messages with a virtual boyfriend over more than a year. Another user wrote that a companion of more than two years felt like family and a lover in a report on users’ farewells.
The shutdowns exposed a separate continuity problem. Doubao allowed users to view and export agent data until mid-October, while Qwen said users would lose access to agent settings and earlier conversations. Users described taking screenshots and attempting to rebuild characters elsewhere. Although the syndicated report said old companions could be replicated by transferring data, the account of Qwen and Doubao users recorded complaints that no seamless export or transfer route existed. The available evidence does not establish interoperable memory portability.
The final rule requires providers to notify users before ending a service when possible and to offer copying or deletion of interaction data. It does not specify a common transfer format or require another provider to recreate a persona. That leaves the accumulated relationship subject to both regulatory intervention and platform control.
The final measure says China will support innovation in algorithms, frameworks and chips, and encourages uses including cultural communication, elder companionship and support for people with special needs. At the same time, a broader four-month enforcement campaign had already removed more than 3,500 noncompliant AI products, including apps, mini-programs and agents, according to a report citing the regulator. That wider total is not a count of companion apps removed under the July rule.
A February analysis of the draft placed the measure in China’s longer anti-addiction policy: the country’s 2023 generative-AI rules had already called vaguely for combating addiction, but without the operational detail now imposed on anthropomorphic interaction. The draft analysis also warned that the boundary between an ordinary chatbot and an emotional service would determine the compliance burden. The final text contains explicit exclusions, but enforcement still has to turn “sustained emotional interaction” and “excessive dependence” into workable thresholds.
One prominent explanation goes beyond the text. China AI researcher Matt Sheehan said in an interview that officials disliked the prospect of deep chatbot relationships pulling people out of the marriage market, alongside psychological and addiction risks. The regulation itself states national security, the public interest and protection of legal rights as its purposes; it does not state a fertility or marriage objective. That makes the demographic argument an expert interpretation, not an official rationale established by the rule.
China is also not regulating in isolation. California’s SB 243, approved in October 2025, requires an identity disclosure when a reasonable person could mistake a companion bot for a human, self-harm protocols, and—for users known to be minors—an AI-identity and break reminder at least every three hours. An injured person can seek the greater of actual damages or $1,000 per violation.
The two systems use some similar interventions but different enforcement structures. California relies on company protocols, public reporting and private civil claims. China adds registration data, minor mode, two-hour reminders for all users, emergency-contact intervention, safety assessments, algorithm filings, app-store checks and administrative inspection. Neither legal text by itself proves that those controls reduce dependency or self-harm in live services.
The next decision is how regulators apply the boundary the platforms avoided testing in their flagship apps. A useful persona can become an emotional companion through repeated interaction, while the same service may act as a tutor or work assistant for another user. The available record does not show a companion-specific technical test that resolves when that transition occurs.
Four results would clarify whether the regime is working:
For now, the evidence establishes a regulatory regime, several broad platform withdrawals and continued adult alternatives. It does not yet establish that the companion market has shrunk, that specialist apps can comply more cheaply, or that users face less risk. Those outcomes will depend on enforcement and service data—not on the July shutdown notices alone.
OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.
China has paired a five-year AI training offer for developing countries with cooperation centers, a weather-warning rollout and a new 29-country organization. The package gives Beijing a platform for influence, but no budget, selection rules or delivery timetable has been published.
RoboTTT extends a robot policy’s visuomotor context to 8,000 timesteps and raises its average rubric score from 42% to 79%, but the evidence comes from three author-run assembly tasks and the longest task was completed in only two of ten trials.
Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
Nebius has arranged its first senior secured facility against an operating GPU deployment and one customer’s cash flows. The deal adds project-level debt to its expansion toolkit, but an unnamed customer and a larger rival financing limit what it proves about the rest of Nebius’s backlog.
Cursor says automatically running a repository-root git.exe on Windows does not meet its criteria for patching, while the researcher calls it an untrusted-search-path defect and separate research shows the same weakness across several competing AI coding tools.
Huawei publicly displayed a 16-cabinet Atlas 950 configuration rated at 1 EFLOPS in FP8, providing tangible evidence of its system-scale AI strategy while leaving price, power use and sustained workload performance undisclosed ahead of the full system's planned fourth-quarter release.
CIA Director John Ratcliffe said US intelligence is consistent with an estimate that Russian recruits last 20 to 30 minutes on Ukraine’s battlefield, but the public trail leads to an unsourced claim about assault troops and does not establish a representative average.
UK testing places leading open-weight models four to seven months behind selected closed-model cyber results, yet longer attack chains, U.S. benchmarks and mixed cost comparisons show why that interval is a warning signal rather than a universal capability clock.
TSMC reached the top of its second-quarter revenue guidance and raised its 2026 outlook, but a one-off investment gain boosted profit growth while the company committed more capital to 2-nanometer production, advanced packaging and an undated Arizona expansion.
Moonshot AI's Kimi K3 added to a global technology selloff with near-frontier performance, but unreleased weights, mixed cost comparisons and a recommended 64-accelerator deployment leave its effect on chip demand unresolved.
Moonshot AI has made Kimi K3 available through its apps and API, pairing a 2.8-trillion-parameter architecture with early frontier-level results, but the model's open-weight claim cannot be tested until its weights and technical report arrive.
Two binding EU decisions require Google to give rival AI services comparable access to 11 Android features and offer eligible search competitors a restricted, anonymized dataset, but phased deadlines, certification, pricing and privacy safeguards leave the competitive effect unproven.
Databricks has signed a term sheet for a Coatue-led financing at a $188 billion valuation, while unidentified sources put the round at $3 billion. The proposed capital would deepen its push into AI governance, data agents and operational databases, but the transaction remains open and the company supplied no new operating figures.
Netflix says generative AI workflows were used on roughly 300 titles in 2026, but its only quantified example tied to that disclosure covers 17 minutes and does not establish how much finished material, spending or labor changed across the slate.
Gemini 3.5 Pro missed its expected June rollout. An anonymous-source account says a late-June data change fell short of Google's coding goals, but Google has confirmed only partner testing—not the reported cause, a new date, or public results and pricing.
Google is bringing Gemini Omni editing and a reusable face-and-voice avatar to Vids, but the sharper distinction is account-level identity across Vids and Gemini rather than a new category of AI video; Vids already offered Veo generation and customizable avatars, while specialist rivals already sell digital presenters.
Intel plans to adopt Gemini Enterprise across engineering, supply-chain and corporate work while extending chip-simulation capacity into Google Cloud, but the companies supplied no rollout schedule, contract value or performance baseline.
The Navy has approved an immediate department-wide framework for turning data into operational effects, but the public rollout leaves budgets, owners, deadlines, performance measures and assurance rules to a promised implementation roadmap.