What ChatGPT's new parental controls do, and what they don't

OpenAI made two announcements about teenagers this month, and most parents will have seen neither. On 13 July it expanded the safety notifications attached to ChatGPT’s parental controls, and three days later it published a short manifesto titled “Why teens deserve access to safe AI”. Together they are the clearest picture yet of how the world’s most-used chatbot thinks parents should be involved.

That matters here more than most places. A study presented at NUS in May, using nationally representative data from 2,985 Singapore children in the SG-Leads cohort, found that more than half of eight-year-olds here have already used AI tools, rising to more than nine in ten by age 13. The most popular app among them is ChatGPT.

So the control panel OpenAI has been quietly building is now relevant to a very large number of Singapore households. It is worth 15 minutes of your evening. It is also worth understanding what it deliberately refuses to do.

What changed this month

The parental controls themselves launched in September 2025. A parent sends an invite from their own ChatGPT account, the teen accepts, and the parent gets a settings page for that teen’s account. July’s update added two things.

First, wider safety alerts. Parents were already notified if OpenAI’s systems, checked by a small team of trained human reviewers, saw signs a teen might be considering self-harm. Now they will also be told if their teen’s account is banned for violent threats or activity. OpenAI stresses these alerts are meant to be narrow: fictional writing, gaming chat, news, politics and ordinary anger are not supposed to trigger anything.

Second, parents can now switch on Study Mode from the parental controls page. Study Mode, built with teachers and learning scientists, pushes ChatGPT towards guiding questions and step-by-step explanations rather than simply handing over the answer. When a parent enables it, every new chat the teen starts opens in that mode. If you have read our piece on when AI helps kids learn, you will recognise why this is the setting we care most about: hints teach, answers replace.

The same week, OpenAI said teens who spend long stretches in ChatGPT will now get more frequent reminders to take a break, and repeated that it is rolling out an age-prediction system so that suspected under-18s get the teen experience automatically, whatever their account claims.

What you can actually set

Once accounts are linked, the parent’s page offers a short list of switches. Quiet hours, so ChatGPT simply cannot be used at 1am on a school night. Voice mode on or off. Image generation on or off. Whether ChatGPT can use saved memories about your child. And, on by default for linked teens, a “reduce sensitive content” setting that adds extra guardrails against graphic material, viral challenges, and sexual, romantic or violent roleplay.

One default is worth noticing: for a linked teen account, “improve the model for everyone” is off. Your child’s conversations are not used to train OpenAI’s models unless you turn that on. Quietly, that settles a question about children’s data that regulators elsewhere have spent years arguing over.

What you deliberately cannot see

Here is the part that surprises most parents: linking accounts does not let you read your teen’s conversations. OpenAI’s help pages are blunt about it. The controls manage settings and send narrow safety alerts, and that is all. If a teen unlinks, the parent is told it happened, but never shown what was said.

We think that design is right, and we say that as an interested party (Mentus AI builds an AI mentor for children, so we watch this space professionally). A teenager who knows every word is being read simply moves the real conversation somewhere else. Alerts for genuine danger, privacy for everything else, is the balance that keeps a teen actually using the account you can see.

The honest limits

OpenAI’s own announcement admits the guardrails are “not foolproof and can be bypassed if someone is intentionally trying to get around them”. Study Mode is a default, not a lock: new chats start in it, and nothing stops a determined teen switching it off mid-chat. And the whole system assumes a teen with their own account who accepts your invite.

Which brings us back to that Singapore study. The eight-year-olds it describes are not the linked-teen users these controls were designed for, and however they are getting to ChatGPT, none of these protections travels with them. MOE is introducing educational AI tools under supervision from Primary 4, but at home the youngest users are exactly the ones the new controls cannot reach.

So do the 15 minutes if your teen has an account: link, set quiet hours, turn on Study Mode. Then do the thing OpenAI itself recommends alongside its own product, which is to talk with your child about how they actually use it. The settings are the easy part. They always were.