Summary: On February 4, 2026, UNICEF issued a statement from New York under the headline “Deepfake abuse is abuse.” The agency pressed governments on two fronts: stretch the legal definition of child sexual abuse material (CSAM) far enough to cover AI-generated imagery, and turn that imagery’s creation, procurement, possession, and distribution into criminal offenses. AI developers: build in safety-by-design and strong guardrails so their models cannot be bent toward generating abusive material.
On February 4, 2026, UNICEF issued a statement from New York under the headline “Deepfake abuse is abuse.” The agency pressed governments on two fronts: stretch the legal definition of child sexual abuse material (CSAM) far enough to cover AI-generated imagery, and turn that imagery’s creation, procurement, possession, and distribution into criminal offenses. Reuters reported on the call later the same week, and on February 10, timed to Safer Internet Day, UNICEF USA took the same message into a Forbes essay.
This advocacy push does not follow the usual template, because the evidence base behind it is the biggest yet assembled from children themselves: in the past year alone, at least 1.2 million children said images of them had been altered into sexualized deepfakes. For the AI industry, the statement drags platform responsibility across a line — from pulling material down after the fact to blocking it before it circulates — and the crosshairs rest on the guardrails built into image-generation models.
Where the 1.2 Million Comes From
The figure was produced by Disrupting Harm Phase 2, a research program run by UNICEF’s Office of Strategy and Evidence – Innocenti in partnership with ECPAT International and INTERPOL, with funding from Safe Online. Fieldwork stretched across 11 countries; each one surveyed roughly 1,000 internet-using children aged 12 to 17 together with about 1,000 parents or caregivers, and national coverage rates landed between 91 and 100 percent. Weighting drew on UN 2024 population data plus estimated child internet-use rates.
Three numbers deserve a lasting place in memory. One: over a single year, at least 1.2 million children disclosed that images of them had been reshaped into sexualized deepfakes. Two: in some countries the ratio works out to one child out of every 25 — or, in UNICEF’s framing, one child in any given classroom. Three: in some countries, up to two-thirds of children said they worry about AI being used to produce fake sexual images or videos of them. National reports with country-level findings will be released on a rolling basis through 2026.
Assignments for Governments, Developers, and Platforms
The statement hands three groups their marching orders. Governments: broaden CSAM definitions so AI-generated content sits inside them, then criminalize that content’s creation, procurement, possession, and distribution — extending the law all the way to possession sets a bar above what many current laws reach. AI developers: build in safety-by-design and strong guardrails so their models cannot be bent toward generating abusive material. Digital companies: stop waiting for reports and choke off circulation proactively, investing in detection so content comes down at once instead of days after a victim speaks up.
Two companion documents sit alongside the statement: UNICEF’s Guidance on AI and Children 3.0, refreshed in December 2025, and an issue brief on AI and child sexual abuse and exploitation that casts the spread of generative tools as a significant escalation of risk to children.
The Definition Fight: Closing the “No Real Victim” Loophole
A single line does the statement’s definitional heavy lifting: “Sexualised images of children generated or manipulated using AI tools are child sexual abuse material.” With it, a legal gray zone is sealed shut. So long as an image could be classed as synthetic, offenders and platforms had room to argue that no real victim existed. UNICEF rejects that exit — the harm does not shrink depending on where the pixels came from: “Deepfake abuse is abuse, and there is nothing fake about the harm it causes.”
“Nudification” tools are named as well: apps that deploy AI to strip or alter the clothing in photos and produce fake nudes. Cheap to build and quick to spread, they act as the main technical vehicle behind the 1.2 million figure. The parting line the statement leaves with legislators: “Children cannot wait for the law to catch up.”
What Builders and Product Teams Should Hear
Anyone with an image-generation or image-editing product on the market should treat the statement as an early look at where regulation is headed. Safety-by-design is on its way from a promise printed on a model card to an obligation that national criminal law can impose: on the input side, requests to modify photos of real children have to be blocked; on the output side, everything a model emits must be detectable and traceable. The demand for proactive detection, in turn, lifts classifiers and watermarking out of the nice-to-have column and plants them in infrastructure — one thread of the safety and regulation outlook we sketched in our opening outlook for 2026.
Liability is where the shift bites first. The moment “removed days later” gets branded unacceptable in public, review SLAs, moderation pipelines, and model guardrails all face a harsher light. Folding child-protection scenarios into red-teaming and evaluation metrics today costs far less than retrofitting after the legislation is on the books.