The gap that matters to anyone who owns a safety roadmap sits between the number a team can see and the number a child lives inside. Run the study's own arithmetic. If fewer than 1% of cases reached police or any support service [13], then against the estimate of roughly 20 million affected children [1], at least 19.8 million cases never entered a queue anyone could work [1]. Report volume inside a product measures how many 14-year-olds will fill in a form about the worst thing that happened to them last month, not the prevalence of the abuse itself.
That changes what detection has to do. Most youth-safety tooling is built around the unknown account: unsolicited contact, a grooming arc long enough for a classifier to learn, a profile with signals that pattern-matching can flag. That case is real, and in this data it is the smaller one, at 38% of cases [11]. The larger share, 57%, involved someone already in the child's life [11], and the two together account for 95% of cases [4], which leaves almost no room for a third story. Coercion between a boyfriend and a girlfriend in a messaging thread looks, to a content system, like two people messaging.
The category counts show how tangled a single case can be. Unwanted sexual material, pressure into sexual conversations or image-sharing, and non-consensual sharing of intimate images sum to 28 million experiences across a 20 million child estimate [5][1], roughly 1.4 categories per affected child [2]. UNICEF also reports that affected children described suicidal thoughts and self-harm at around four times the rate of their peers [12].
The synthetic-imagery figure is the one with the weakest floor under it. An estimated 1.1 million children across nine countries said sexual imagery of them had been produced with AI, a category that barely existed when fieldwork started in 2020 [8]. Those nine countries are exactly the second-round set from 2024 and 2025; the twelve surveyed earlier were never asked, so there is no baseline for a growth rate [9]. Systems built to match known material also struggle with images that have never existed before [10], which is a straightforward statement that the main installed detection base does not cover this.
Two limits are worth holding on to. The sample excludes the United States, the United Kingdom and EU member states, so this is a floor for 21 countries rather than a global total, and UNICEF has not extrapolated past them [15]. And the study's recommendation is aimed at product, not parents: "Platforms must build safety into products from the outset, rather than relying on children and parents to manage risks after harm occurs," said John Livingstone, head of digital policy at UNICEF Australia [14]. Australia's eSafety commissioner has separately found serious gaps in how large companies respond to child sexual abuse and sexual extortion [16], so more than one party is already keeping score.
A cut that survives a roadmap review: sort each safety item by whether it fires without a child reporting, and by whether it addresses a known contact or an unknown account. Everything in the report-dependent column is capped by that sub-1% reporting rate. Everything in the unknown-account row is aimed at the smaller half. The two boxes that tend to be thinnest, detection that fires on its own against harm from someone the child already knows, are where this study puts the volume.
