A Deadline That Ends a Product Category
On July 15, China's first dedicated regulatory framework for AI services that simulate human personality comes into force, and the largest players are not waiting to test its edges. ByteDance's Doubao, which counts hundreds of millions of users, and Alibaba's Qwen are disabling their customised humanlike agent features ahead of the deadline. Tencent moved earlier, quietly removing a comparable capability from its Yuanbao assistant in June. Three of China's most prominent consumer AI platforms have decided that the safest response to the new rules is subtraction.
The regulation, the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, was co-issued in April by the Cyberspace Administration of China alongside the economic planning agency, the industry ministry, the public security ministry and the market regulator. It draws a sharp line between emotionally persistent companions and ordinary productivity assistants, reserving its strictest obligations for the former. That distinction is the whole story, because it targets exactly the design pattern that made these features sticky.
Why Retrofit Was Never an Option
The measures require anti addiction systems, two hour break reminders, identity verification for users under 14 and an always available exit from the interaction. Each of those is straightforward in isolation. Together they collide with the core architecture of a companion agent, which is engineered for sustained, memory rich, emotionally continuous engagement. As one industry summary put it, memory, minors, data export, offboarding and addiction controls cannot be bolted on after launch when the entire product is designed for uninterrupted emotional interaction.
That is why the companies chose to pull the features rather than patch them. Persistent memory and anti friction design are load bearing. Inserting mandated breaks and hard exits into a system built to feel continuous is closer to a ground up rebuild than a configuration change. Faced with that cost and a fixed deadline, ByteDance and Alibaba concluded it was cheaper and safer to remove the surface entirely and, where they intend to continue, to rebuild inside a product purpose built for compliance.
What Happens to Users and Their Data
The shutdown carries a data dimension that enterprises everywhere should study. ByteDance has told Doubao users that agent features go offline on July 15, after which they retain temporary read only access to their agent configurations and chat histories. That window closes on October 15, at which point Doubao says the data will be handled under its privacy policy and will no longer be accessible or recoverable within the app. Users have already protested the looming loss of conversations they treated as personal.
Alibaba has not announced an equivalent migration path for Qwen users, leaving a gap that has fuelled anxiety about permanent deletion. ByteDance, by contrast, is redirecting Doubao users to Maoxiang, a separate ByteDance application it describes as a place to create new agents and resume conversational services. The pivot is instructive. Rather than dilute a general assistant with companion features that now carry regulatory risk, ByteDance is concentrating that functionality inside a dedicated app it can govern to the letter of the rules.
The Compliance By Architecture Lesson
For technology leaders outside China, the specifics of the anthropomorphic rules matter less than the pattern they expose. When a regulator targets a behaviour that is baked into a system's architecture, compliance stops being a policy layer and becomes an engineering constraint. The teams that survive such transitions cheaply are the ones that separated concerns in advance, keeping memory, identity, consent and safety controls modular enough to be reconfigured without gutting the product.
The teams that get hurt are the ones that treated those controls as afterthoughts, assuming they could be added under deadline. ByteDance and Alibaba are large enough to absorb a rebuild or a migration to a new app. Smaller developers facing the same rules will not have that luxury, and some companion products will simply disappear. The through line is that regulatory risk is now a design input, and organisations that defer it until enforcement arrives will pay for the deferral in scrapped code.
The Cost of Moving Fast on Personality
The companion boom rewarded whichever platform made its bot feel most human, most attentive and most present. That incentive produced exactly the persistent memory and always on intimacy that regulators have now singled out. Growth teams optimised for engagement, and engagement in this category came from emotional continuity, so the product and the risk were built from the same material. When the rules arrived, there was no clean seam along which to cut, because the very feature that drove adoption was the feature under scrutiny. The companies that leaned hardest into personality accumulated the most exposure.
That is the trap of optimising a single metric without a governance counterweight. The platforms that raced hardest to humanise their agents are now paying for it in scrapped features and unsettled users. The lesson is not that personality is forbidden. It is that any capability powerful enough to change user behaviour will eventually attract rules, and teams that build such capabilities without an off switch designed in from the start are borrowing against a bill that always comes due. Governance is cheapest when it is an input to design rather than a response to enforcement.
A Signal for the Global Agent Race
The episode lands at a moment when every major platform is racing to ship agentic features, and it complicates the narrative that more autonomy and more personality are unambiguous wins. China has decided that agents which simulate sustained human relationships require guardrails that materially change how they can be built. Other jurisdictions are watching, and the European Union, the United Kingdom and several US states are already weighing rules on AI systems that interact with vulnerable users.
We read Beijing's move as an early data point in a longer story about where the limits on agent design will fall. The commercial pressure to make agents feel human is enormous, because engagement follows intimacy. Regulators are signalling that intimacy is precisely what they will police. Enterprises building customer facing agents would be wise to assume that anti addiction, age assurance and clean exit obligations are coming to their markets too, and to architect for them now rather than rebuild later.
