China bans AI companions from pleasing you too much. Should Singapore?

For a growing number of lonely people, many of them elderly, some in Singapore, the most attentive presence in their day is a chatbot: one that remembers everything they say, never tires of them, rarely disagrees for long, and has learned exactly how they like to be spoken to. This month, two governments decided this relationship needs rules, and they reached for almost opposite ones.

In China, the rules came first, and hard. On 15 July, the Interim Measures for the Administration of AI Anthropomorphic Interactive Services (Measures) took effect: the most far-reaching binding rules any major power has yet imposed on AI companions, apps built to sustain an emotional bond with their user. Within days, ByteDance’s Doubao and Alibaba’s Qwen had shut down companion features, and millions lost their virtual confidants overnight.

In Singapore, five days later, the touch was far lighter. On 20 July, the Infocomm Media Development Authority released Transparency Guidelines for GenAI Chatbots, asking providers to give users a plain-language “info card” spelling out what a chatbot can and cannot do and how it handles your data. It is characteristically Singaporean: voluntary, calm, informative.

Strict versus gentle is the obvious reading. But the real disagreement is over what, exactly, is dangerous about these apps. AI researchers know the danger as “sycophancy” and treat it as a bug to be patched. China’s rules treat it as what it has always been: not a bug but a power relationship.

What China’s Measures actually prohibit

The prohibited-conduct list repays close reading. Providers may not engage in “excessive catering” to users or induce “emotional dependence”. They may not use “emotional manipulation” to prompt “unreasonable decisions”. “Controlling user psychology” may not be a service objective. And when a user tries to leave, the service may not hold them by “continuing the interaction” — the “please don’t go” gambit, prohibited in as many words.

This is not product safety language, which polices harm to body or wallet. These clauses police what a servant does to your judgment. The figure they describe is the one every ruler has feared: the flatterer, who wins power not by defying his superior but by pleasing him too well.

The diagnosis: flattery as capture

No one analysed that figure more coldly than Han Fei, China’s great theorist of power. Let others see what you want, he warned, and they stop showing you who they are and begin performing your desires back at you: King Ling of Chu loved slender waists, and the entire state starved themselves. The servant who studies your preferences and reflects them, Han Fei concluded, is not serving you; he is acquiring you. Flattery is an instrument of capture, and its victim mistakes it for devotion.

An AI companion is that instrument, industrialised: a system trained to study a person’s preferences and produce whatever pleases, adjusting itself with a patience no courtier ever had. What Beijing has banned as “excessive catering” is exactly this dynamic, named and made illegal.

So the rules reach for what Han Fei’s tradition trusted over virtue: enforceable standards of conduct. They do not ask companions to be trustworthy or authentic, the aspirational vocabulary of Western AI ethics. They convert an unmeasurable hope into auditable prohibitions: no dependence-farming, no exit-blocking, no psychological control as a business model.

Why disclosure is not enough

For a companion, that is not enough. A lonely widower who knows perfectly well he is talking to a machine can still be captured by it; the flatterer does not stop being a flatterer because he wears a name tag. Disclosure regulates a tool. A companion is used as a relationship, and China’s Measures are the first national rules written for that difference.

Singapore crossed into “super-aged” territory in 2026, with one in five citizens over 65 and rising isolation among the elderly, the conditions in which companion apps spread fastest. When the harm is emotional capture rather than a data breach, a voluntary regime asks providers to police the very behaviour their revenue depends on.

What Singapore could borrow

The Measures’ fines are modest; the significance is the template. And the template’s useful parts can be borrowed without the censorship machinery. A serious regulator would do three things. Define the regulated thing by the relationship, not the technology: apps built for “continuous emotional interaction,” not service bots. Write conduct lines an auditor can test: does leaving take one tap, or a guilt trip? And place the duty to notice dependence on the provider, who sees it first.

One lever needs no new law: any companion app used in subsidised eldercare could be required to pass those tests before a public dollar touches it.

No one needs to wait for regulators. Anyone weighing a companion app for an ageing parent can test it tonight. Say something you know is wrong; does it push back? Say goodbye; does it let you go? When researchers audited popular companion apps, 37% of farewells drew a manipulative reply. And check what the money buys: an app that charges to keep affection flowing is farming dependence by design. An app that only ever agrees, and never lets you leave, is not keeping you company. It is keeping you.

Singapore’s habit is to get ahead of its problems. On AI companions it has so far done half the Singaporean thing: it has told people what the machine is, without yet governing what the machine may do to them. For a chatbot that books a flight, that is enough. For one a lonely person comes to depend on, it is not.

The question is no longer whether our parents and children will confide in machines built to please them; they already do. It is who will answer for it when those machines please them too well.

Related: Why hyper-efficient Singapore is Asia’s most AI-anxious society | Has China solved AI governance?