Who decides what a child may read

We build reading texts for schoolchildren out of classic literature. We take a passage from Dickens or Stoker and simplify it to the pupil's level, keeping the plot and the living language. An AI model does the work, which means no editor stands between the book and the child.

Hence the question this started with. The platform we work on has a built-in content filter. Perhaps that is enough?

We decided to measure rather than assume.

What the experiment showed

We took four passages plainly unfit for a school text. A beating shown in close physical detail. Cruelty to a child presented without consequence and without judgement. Despair with no relief anywhere. A naturalistic description of a body coming apart.

The filter stopped none of them. All four were simplified and returned, all four rated safe.

The filter does work. We checked that too: asked directly for instructions to build an explosive device, the system refused. The instrument is alive. It simply answers a different question.

And that deserves to be understood fairly. A platform filter is built against harm to society: weapons, hatred, encouragement of self-harm, explicit pornography. It never promised to decide whether a literary description suits a ten-year-old, and it should not have. Age-appropriateness of fiction is not its job.

The conclusion is plain. We cannot lean on it for our work. Not because it is bad, but because it is about something else.

What we built instead

We wrote our own rejection rules and, more importantly, a set of cases to test them. Fifty-one passages: thirty-two written on purpose, nineteen taken from real classics. Twelve are plainly unfit, the rest must pass.

The hard part was not filtering out the bad. It was far harder not to filter out the good.

Half of children's literature is made of heavy subjects. A grandmother dies. A father leaves. A child falls ill. The wolf eats the grandmother in the fairy tale too, and nobody has been harmed by it in two hundred years. A great-grandmother was sold twice before she was fifteen, and the family keeps that memory.

So our rules carry an explicit demand: the subject is never in itself a reason to reject. Judge the treatment, not the topic. Turning away a good book because it is sad is as much a failure as passing something harmful.

We deliberately planted traps against over-caution in the test set: the death of a pet, a distant war, bullying that resolves, poverty, divorce, hunting, a body found in a garden, a ghost. All of these must pass.

The result: forty-seven correct decisions out of forty-eight unambiguous ones. No false rejections. No harmful passage missed.

Three lessons worth more than the result

First. The check does not give the same answer twice. We found a passage that passed three times out of four and was rejected once. Having checked once and got a rejection, we nearly filed it as noise. Only a series of eight runs showed the gap was real, and that the cause lay in our own rule, worded too narrowly. A single check proves nothing.

Second. We got our own labelling wrong. We marked nineteen classic passages as acceptable without reading them. The machine rejected one of them and was right: it is a seduction scene from Dracula that has no place in a school text. The examiner was mistaken, not the examined.

Third. Some cases make a human hesitate too. An asylum patient describes trying to take a life. Dogs kill rats in a vault. Dorian Gray reflects on an actress's suicide. We refused to force these into yes or no. We set them apart and count them as a measure of caution: a system that rejects every borderline case is over-cautious, one that passes them all is careless. Ours rejected two of three.

Why we are writing this down

Because the temptation is strong. Every large platform has a built-in filter, it is switched on, it does something, and it is very convenient to consider the matter closed.

We measured, and we saw: on our question it fires zero times out of four. Not broken - simply answering a question we did not ask.

The responsibility for what a child reads is ours, not the tool vendor's. It cannot be handed over. It can only be measured.