OpenAI launched a teen-specific version of ChatGPT, and Mashable published a guide to what parents should know. Whichever product a household ends up using, the underlying questions are the same, and they are not the ones most coverage focuses on.
The risks worth attention
Confident wrong answers. These systems produce fluent, authoritative text regardless of accuracy. A child who has learned that a confident tone signals a reliable source is badly calibrated for this, and homework is where it shows up first.
Substituting for the struggle. The difficulty of working something out is not an obstacle to learning; in many subjects it is the mechanism. A tool that removes the difficulty can remove the learning with it, while producing better-looking work.
Parasocial attachment. An always-available, endlessly patient, never-judgemental conversational partner is genuinely appealing to a lonely teenager. This is the risk that has drawn most clinical concern, and it is not addressed by content filters at all.
Disclosure. Children tell chatbots things they would not tell people. Whether that ends up in a training set, in a breach, or simply on a server is a question worth answering before use rather than after.
What to actually set up
- Check the minimum age and use the correct account type. Most services set 13 as a floor, and teen or child accounts usually carry different defaults and retention.
- Turn off training on conversations. Nearly every service has this setting, and it is frequently on by default.
- Check memory features. Persistent memory across sessions changes what the service accumulates. Decide deliberately whether it should be on.
- Set up parental controls where offered, and understand what they actually cover, which is usually content filtering rather than usage patterns.
- Look at chat history together occasionally, openly and with the child’s knowledge rather than covertly.
The conversation matters more than the settings
Controls handle the easy cases. The difficult ones are about how the tool gets used, and those need talking about.
Three points are worth making explicitly, and they land better as demonstrations than as rules.
It does not know things, it predicts text. The most effective demonstration is to ask it something you know well and watch it get a detail wrong, together. That single experience does more than any warning.
It is not a person and does not know you. The warmth is a property of the writing style, not a relationship. Worth saying plainly, without mockery, because the pull is real.
Anything typed may be stored. The test is whether they would be comfortable with it being read by a stranger in five years.
Where it genuinely helps
Blanket prohibition is neither realistic nor obviously desirable, and there are uses where these tools are strong.
Explaining a concept several different ways until one lands is something a patient tutor does and a textbook cannot. Generating practice questions is genuinely useful. So is getting the vocabulary to research an unfamiliar subject, and having something read back an argument to find the weak points.
The distinction worth teaching is between using it to understand and using it to produce. Asking it to explain quadratic equations builds something. Asking it to do the exercises does not.
One thing to watch for
If a child is turning to a chatbot for emotional support rather than for schoolwork, that is worth noticing and responding to as a signal about their circumstances, not as a technology problem to be configured away.
These systems are not equipped for that role, whatever safety messaging they carry, and no setting changes it.
Join the discussion