Singapore's app store age checks, and the AI they miss

Your child cannot download a dating app on a Singapore phone any more. She can still open a browser, type six letters, and be in conversation with an AI inside ten seconds.

Both things are true as of April, and holding them together is most of the job for parents this year. Singapore now runs a genuine age check at the app store door. It is worth switching on. It is also, by design, pointed somewhere other than where AI usually reaches a child.

What actually changed in April

IMDA’s Code of Practice for Online Safety for App Distribution Services took effect on 31 March 2025, and its heaviest obligation landed a year later. From 1 April 2026, designated app stores serving Singapore have to work out whether a user is under 18 and stop those users downloading age-inappropriate apps. Five stores are covered: Apple’s App Store, Google Play, Huawei’s AppGallery, the Samsung Galaxy Store and Microsoft’s store.

Each picked its own method. A survey of the rollout by the trade publication Biometric Update found Microsoft accepting Singpass, a government ID scan or facial age estimation; Apple using credit card data or government IDs; Samsung and Huawei relying on a linked credit card; and Google estimating age from account activity by default, with a selfie or credit card as the fallback.

The first tranche is deliberately narrow. Answering a parliamentary question in February 2025, MDDI said stores would need to identify users under 18 and restrict the highest-rated (18+) apps to begin with, which in practice means dating services and pornography. A general chatbot app carrying a 12+ or 13+ rating sits well below that line. It downloads exactly as it did before.

The gate reads a label, and AI labels are shaky

The whole mechanism rests on age ratings being roughly right. For AI apps, that is a lot of weight on a thin plank.

Apple only widened its rating scale in July 2025, adding 13+, 16+ and 18+ to the old 4+ and 9+, and told developers to answer the new questions by 31 January 2026. Buried in that same notice is a line worth reading twice: “you must consider how all app features, including AI assistants and chatbot functionality, impact the frequency of sensitive content appearing within your app to make sure it receives the appropriate rating.” That is Apple asking developers, politely, to rate their own chatbots honestly.

Plenty do not. In a report published on 27 January 2026, the Tech Transparency Project found 47 apps on the Apple App Store and 55 on Google Play that could strip clothing from a photo of a real person, with 38 available on both. One of them, DreamFace, was rated 13 and up on Google Play and 9 and up on Apple. Another, AI Dress Up, was rated suitable for all ages.

An age rating is not an inspection. It is a developer’s own answers to a form. When the app is an AI whose output changes with every prompt, the form is being asked to describe something nobody has fully seen.

Most AI reaches a child without a download

The code binds app distribution services, and a browser tab is not one. Nothing in an age check at the store stands between a child and a chatbot website, a search page that answers in prose, or an AI feature switched on inside an app that was installed two years ago. We wrote in more detail about the search problem in Kids can’t switch off Google’s AI answers.

School is the clearest illustration, and to MOE’s credit it is a well-handled one. In a parliamentary reply on 6 May 2026, the Minister for Education set out the approach: from Primary 1 to 3, physical hands-on learning is prioritised and schools will not assign work requiring pupils to use AI directly; from Primary 4, pupils use educational AI tools under teacher supervision inside the Singapore Student Learning Space. The Learning Assistant that helps P4 pupils with composition will, in MOE’s words, redirect students back on track if they veer off-topic. Data from MOE-built tools is anonymised and not used to train external models. National exams stay proctored, with AI prohibited.

Notice what is doing the safety work there. Not an age gate. Teacher supervision, a bounded tool and a rule about when it is used. None of that came from the app store.

The ministry already said the honest part

What is refreshing is that nobody in government has oversold this. Asked in February 2025 whether facial recognition should be mandated so children could not slip past the checks, MDDI’s reply included a sentence most regulators would have left out: “Regulations alone are unlikely to prevent a parent or any adult from helping a child circumvent the age assurance measures.” In October 2025, updating Parliament on whether Singapore should mandate social media age limits, the ministry said plainly that none of these measures are foolproof, and that children can get round them by using an adult’s account. Its stated plan was to work with MOH, MOE and MSF on resources for parents, not to keep tightening the gate alone.

Read that as a floor, not a failure. A floor is useful. It just is not a ceiling, and it was never advertised as one.

Ten minutes, then the part no gate can do

Spend the ten minutes. Set up the age check properly on your child’s device rather than leaving them signed into your account, which is the single move that quietly undoes the whole thing. Then do the four things the store cannot:

• Ask which AI they use in a browser, not just which apps are installed. That is where the answer usually is.
• Treat a low age rating as no evidence at all. It tells you what a developer typed on a form.
• Ask the school what its actual practice is, not just its policy, and check what comes through Parents Gateway about how AI is used in class.
• Try the tool yourself for five minutes with a hard question. You will learn more than any rating will tell you.

We build an AI mentoring product for children, so we have a commercial interest in how parents judge these tools. That is also why we would rather they judged them by what the thing does in a real conversation than by the number beside it in a store listing. If you are setting controls this month, our piece on what ChatGPT’s parental controls do covers the other half.

The age check asks how old your child is. The question that matters is what she does with the ten seconds after she opens the browser, and the only way to know that is to ask her.