The AI Just Stopped Showing Its Work. What Do You Teach Now?

For three years the best advice anyone could give a parent about AI fitted in one sentence: teach your child to ask how it got there.

It worked because you could check. The model showed its steps, the steps were sometimes wrong in visible ways, and a child who read them learned something real about how the answer was made.

GPT-6 Astra, released on 4 September 2026, uses a reasoning technique OpenAI calls recurrent depth — reported as working in a way that obscures some or all of the model's chain of thought. Researchers have raised concerns about it on monitorability grounds.

If the reasoning is not on display, the advice needs rebuilding. Here is what survives, what does not, and what replaces it.

Last verified: 5 September 2026.

What actually changed

Be precise, because the coverage has not been.

What has not changed: the model still answers. It can still be asked to explain itself, and it will produce an explanation.

What has changed: that explanation is less reliably a record of how the answer was reached, and more a plausible account produced afterwards.

That distinction sounds academic. It is the whole thing. A child reading visible reasoning was auditing a process. A child reading a produced explanation is reading a second answer — which can be wrong in the same direction as the first, and sound just as good.

The advice that no longer carries the weight

"Ask it how it worked that out"

Still worth doing — but demote it. It now tests whether the model can produce a coherent-sounding account, which is not the same as whether the answer is right.

The tell: ask the same question in two fresh conversations and compare the two explanations. If they differ in shape, you are looking at reconstruction, not recall.

This is worth doing with your child once. It teaches more about how these systems work than any explanation you could give them, and it takes four minutes.

"Read the steps and find the mistake"

This was the best exercise available and it degrades when the steps are partly hidden. A child cannot find the error in reasoning they cannot see.

What replaces it is below, and it is arguably a better skill anyway.

The three that get stronger

Each of these was always good practice. Each now carries more of the load.

1 · Check outside the model

The single most important habit, and the one that was easiest to skip when the reasoning looked convincing on screen.

The practice: one answer a week, checked against something that is not an AI. A book, a school worksheet, a search that lands on a named source.

What the child learns is not distrust. It is that a claim has a source, and that finding the source is a normal thing to do — which is the actual definition of literacy in this decade.

2 · Ask what is missing, not whether it is right

"Is this right?" invites a yes. "What's missing?" requires comparing the answer to a mental model of a complete one.

This works whether or not the reasoning is visible, because it is not about the model's process — it is about the child's. That is precisely why it survives the change.

3 · Notice confidence that does not match difficulty

The most transferable skill on this page. Teach your child to notice when a machine is equally confident about the capital of France and about something genuinely contested.

A model that shows less reasoning gives fewer signals of uncertainty. So the signal to watch shifts from how it explained itself to whether the confidence fits the question.

Concretely, ask together: "Would a person who really knew this sound this sure?"

Two mistakes to avoid this month

Do not tell your child the AI is lying. It is not. It produces a plausible account, which is a different failure and needs a different response. A child told the machine lies stops using it in front of you, which costs you the only supervision that works.

Do not go back to banning. The reasoning becoming less visible is an argument for more checking, not less use. A child who stops using it does not become better at evaluating it — they just meet it later, alone, with no practice.

The one-sentence version, updated

Three years ago: ask how it got there.

Now: check somewhere else.

That is a smaller sentence and a bigger skill. It does not depend on any model's design choices, it works for the next model and the one after it, and it is the same skill that lets an adult read a headline and go looking for the study.

You now know what to teach instead. The harder part is the age-by-age version: a six-year-old cannot check a source, and a fourteen-year-old needs more than a habit — the practice looks genuinely different at each stage.