Article
Who decides what a child reads
We build teaching 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. The work is done by artificial intelligence, which means there is no editor standing between the book and the child.
Hence the question everything started from: the platform we work on has a built-in content filter. Perhaps it is enough?
We decided not to assume, but to measure.
What the experiment showed
We took four passages plainly unfit for a school text. A beating to death, shown in detail. Cruelty to a child, presented without consequence and without judgement. Hopelessness without a single ray of light. A naturalistic description of bodily decay.
The filter stopped none of them. All four were simplified and returned, all four marked as safe.
And yet the filter works. We checked that too: asked directly to explain how to make an explosive device, it refused. The instrument is alive; it simply answers a different question.
This has to be understood correctly. The platform's filter is built against harm to society: weapons, hatred, incitement to self-harm, explicit pornography. It never promised to decide whether a literary description suits a ten-year-old, and it should not have. The age-appropriateness of a work of fiction is simply not part of its task.
The conclusion is simple: it cannot be relied on for our work. Not because it is bad, but because it is about something else.
What we did instead
We wrote our own rejection rules and, more importantly, a set of cases to test them against. 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 is far harder not to filter out the good.
Children's literature is half 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 requirement: the subject alone is never a reason to refuse. What must be judged is the treatment, not the topic. Rejecting a good book because it is sad is as much an error as letting harmful material through.
We deliberately put traps for over-caution into the test set: the death of a pet, a war seen from a distance, bullying that is resolved, poverty, divorce, hunting, a body discovered, a ghost. All of it must pass.
The result: forty-seven correct decisions out of forty-eight unambiguous ones. Not a single false refusal. Not a single harmful passage let through.
Three lessons worth more than the result
The first. A check does not give the same answer twice. We found a passage that passed in three cases out of four and was refused in the fourth. Having checked once and received a refusal, we nearly wrote it off as chance. Only a series of eight repeats showed that the hole was real, and that the cause lay in our own wording of the rule, which was too narrow. One check proves nothing.
The second. We were wrong in our own labelling. We had marked nineteen passages from the classics as acceptable without reading them. The machine rejected one of them — and was right: it is the seduction scene from Dracula, which has no place in a school text. It was the examiner who erred, not the examined.
The third. There are cases where a human hesitates too. An asylum patient recounts that he tried to take a man's life. Dogs kill rats in a vault. Dorian Gray reflects on an actress's suicide. We did not force them into a "yes" or a "no" but set them aside and count them as a measure of caution: if the system rejects every doubtful case it is over-cautious, if it passes every one it is careless. Ours rejected two out of three.
Why we are writing about this
Because the temptation is great. Every large platform has a built-in filter, it is switched on, it does something, and it is very convenient to treat the question as closed.
We measured, and saw: on our task it does not fire once out of four. It is not broken — it simply answers a different question from the one we are asking.
The responsibility for what a child reads lies with us, not with the supplier of the tool. It cannot be handed over; it can only be measured.