Consultation on the new rules for minors’ online safety has closed and the final text is not yet published. In this interval, most AI companies are not asking whether to act but which products move first, how far to go, and what it costs.

The answers depend on product form, not company size. The same minors mode costs several times more for an education hardware product than for a text generation tool. And for virtual companion products the question is not cost at all, but whether the service can continue.

1. Two categories, two outcomes

The rules treat AI in two ways, with very different force.

CategoryServices coveredTreatmentEffect on the product
ProhibitedVirtual relatives, virtual partners and similar virtual intimate-relationship servicesMust not be provided to minors at all, with no age exceptionAssess withdrawal, or close the function entirely to minors
RoutedAI services capable of affecting minors’ cognitionFor users under sixteen, must be provided through minors modeBuild the mode, identification and routing, and an appeal channel

The order is to test the prohibited category first, then the routed category. The first determines whether the product survives; the second determines how it is presented. Reversing the order spends budget on the wrong problem.

2. Three inventory questions

First, could the service be used by minors? The test is actual usage, not marketing positioning. “Our users are mainly enterprise clients” and “registration requires an adult declaration” usually do not exclude application.

Second, is there a virtual relative, virtual partner or similar interaction? The test looks at product design: persona settings, emotional feedback, continuous conversation prompts, paid unlocking of intimate storylines. The function’s name does not determine the finding.

Third, which services must switch to minors mode for users under sixteen? Generative conversation, AI companionship, learning tools and content generation are more likely; code completion, translation and enterprise knowledge retrieval are less likely, but where individual registration is open without age differentiation, assessment is still needed.

The three answers set the workload. The product lead should own the exercise, with legal and security participating, concluding within two weeks and keeping a record.

3. Five deliverables for routed services

Minors mode is not a switch on a settings page. It is a combination of five capabilities, and a missing one can prevent the filing materials from passing verification.

CapabilityRequirementCommon gap
IdentificationCombined use of ID verification, national network identity service, non-storage facial verification, behavioural signalsAn adult checkbox at registration
RoutingAutomatic switch once identified, with tiers below sixteen and sixteen to eighteenOne treatment for all ages
Content poolAge-graded content, adjusted recommendation strategy, blocking of risky contentTime limits only, content unchanged
Anti-circumventionExit verification, device-account linkage, anomaly detectionOne-tap disable with no verification
AppealsA correction channel for inaccurate age data, with a response deadlineNo appeal entry point

The third and fourth are most often underestimated. The content pool requires reconfiguring recommendation strategy, which is product work; anti-circumvention touches device and account systems, which is engineering work. Neither is solved by adding a switch.

4. The decision for prohibited services

Products in the prohibited category need a decision within a quarter: withdraw the function entirely, or close it completely to minors.

The first affects all users; the second depends on identification capability. Without identification, “close to minors” cannot be executed, because there is no way to know who is a minor. In that case identification comes first, then closure; the order cannot be skipped.

Renaming is usually harder than withdrawing. Calling a virtual partner a “study assistant” does not change the finding if the interaction design is unchanged. The redesign must reach persona settings, emotional feedback and paid design, not just copy.

5. Timeline and what drives the budget

Build time usually runs eight to sixteen weeks, with four cost drivers.

Terminal form. Software-only products rely on the account system for identification and routing, at lower cost; hardware involvement requires device-side cooperation and raises both timeline and cost.

Existing real-name system. Where real-name verification exists, the work is routing and the content pool; where it does not, identification must be built first.

Content pool size. Age grading requires re-labelling existing content; the larger the pool, the greater the work.

Whether filing is needed. Providers offering minors mode must file with the provincial cyberspace administration; platforms with very large minor user numbers and significant influence must also submit a build and usage report and undergo annual verification. Filing presupposes that the function is actually live.

6. Four common mistakes

MistakeConsequenceCorrect approach
Adult declaration instead of identificationCannot prove identification duty performedBuild verifiable methods and keep records
Mode as a switchOne-tap disable defeats anti-circumventionExit requires verification; device-account linkage
Time limits without content controlsContent risk remainsAge-graded pool and adjusted recommendation
File before buildingMaterials contradict actual function; verification failsFile after the function is live

The second is especially common in hardware. If mode switching depends only on an in-app setting, a factory reset or reinstall circumvents it; system-level cooperation is required.

7. Timetable

Before the final text is published, three things can run in parallel: complete and record the product inventory; assess the viability of prohibited functions; begin building identification capability. After publication, work backwards from the effective date to schedule the content pool and filing.

If the effective date follows closely, those three are the only parts that can be finished early. Starting identification after publication usually misses the first verification cycle.

8. Next steps

Start with a product inventory answering three questions: could the service be used by minors; is there a virtual intimate-relationship function; which services must switch to minors mode. Put the answers on one page with the basis and the owner.

That page costs little and is the starting point for everything else. Without it, the budget cannot be estimated and priorities cannot be ordered.

Related reading: the new minors rules and four preparations for AI services, document requirements for AI education products, the AI product launch compliance document pack, and the AI content labelling and safety service page.

Last updated: 22 September 2026.

Frequently Asked Questions↑ Back to article

Which AI services must never be provided to minors?

Virtual relatives, virtual partners and similar virtual intimate-relationship services, with no age exception. The same article provides an exception for stranger social networking above sixteen, but not for virtual intimate-relationship services.

Which AI services must go through minors mode?

AI services capable of affecting minors' cognition, when provided to users under sixteen. Generative conversation, AI companionship, learning tools and content generation are more likely to fall within this category.

Is an adult declaration at registration sufficient?

No. The regulator expects verifiable identification, using ID verification, the national network identity service, non-storage facial verification and behavioural signals in combination, with records retained.

What does minors mode include?

Five capabilities: identification, routing, an age-graded content pool, anti-circumvention and appeals. A missing capability can prevent the filing materials from passing verification; it is not a switch on a settings page.

Does minors mode require filing?

Yes. Providers offering minors mode must file with the provincial cyberspace administration; platforms with very large minor user numbers and significant influence must also submit a build and usage report and undergo annual verification.

How long does it take to build?

Usually eight to sixteen weeks. Software-only products rely on the account system and are quicker; hardware involvement requires device-side cooperation and raises both timeline and cost.

Geason Law Offices AI Industry Research Group | Reviewed by Ho Qianggao