OpenAI rolled out an experiment for minors on 18 August. The system automatically places anyone it deems to be under 18 into this mode, with enhanced protections, a study mode, homework reminders, quizzes and configurable work time slots. It is a notable development, and it deserves to be examined from both angles.
We have covered the issue from a family perspective in our article aimed at parents and from a school perspective in the one devoted to schools. Here is one provider's answer.
What is well designed
Two elements deserve praise, because they match what the research suggests.
Study mode rather than answer mode. A tool that quizzes the student instead of handing over the solution is exactly what we recommended: the most useful prompt is not give me the answer but ask me questions to check whether I have understood. Making this behaviour native rather than dependent on how the student phrases things is smart.
The enhanced protections. On sensitive topics, increased caution for minors is justified, and it chimes with the concerns we raised about emotional companion systems, which China has outright banned for minors.
The question of automatic age estimation
Here is the point that deserves serious scrutiny. Automatically placing someone in a restricted mode means guessing their age, without them having declared it.
How? By analysing writing style, topics discussed, vocabulary, usage times. In other words, by building a behavioural profile to infer a personal characteristic.
This creates three difficulties.
Errors on both sides. An adult who writes simply, or whose first language is not the service's, may be classified as a minor and have their service restricted without explanation. A teenager who writes well will slip through. We saw the same problem with generated-text detection tools: false positives hit those who deviate from the norm first.
Recourse. Can someone who has been misclassified challenge it, and under what procedure? That is the question raised by the European framework on automated decisions.
The data created. Estimating someone's age amounts to producing personal information they have not provided. This inference now exists, and it must be treated as such, as we described in our article on what your AI knows about you.
There is no good solution. Asking for ID would better protect minors but would create an identity database for a chat service, which is worse. Doing nothing leaves minors unprotected. Behavioural estimation is an imperfect compromise between two worse options. Acknowledging that is better than presenting the system as a clean solution.
What is missing
Two things, and neither depends on the provider.
The role of parents. A teen mode activated by the system, without the family being informed or able to configure it, leaves adults out of the loop. Yet as we wrote: what protects best is not the setting, it is someone who accompanies.
The underlying question, unresolved. A well-designed study mode does not cancel out what the research on cognitive delegation suggests, with the caveat that those findings are correlational. The most robust result concerned order: learning first without assistance appears to protect. No setting can decide that on the student's behalf.
What to take away
This is a positive development, and it should be said plainly: a mode that quizzes instead of answering is better for a teenager than a mode that hands over the solution.
It does not replace what matters more, and which is not technical. An adult who asks a teenager to explain what they have produced gets more information than a detection system, and does them a better service. It is unspectacular, and it remains the best protection available.