Over 400,000 AI-related child sexual abuse material reports in 2025 – against 4,700 in 2023
A new analysis from Childlight shows that NCMEC, the US central reporting hub for abuse material, received over 400,000 reports of AI-generated abuse material in 2025, against roughly 4,700 in 2023. The day after, OpenAI launched a free classification tool – but the figures are report volumes, not documented prevalence, and the methods remain independently untested.
What the new analysis shows
Childlight – Global Child Safety Institute, a research centre at the University of Edinburgh, published an analysis in mid-November 2026 showing that reports of AI-related child sexual abuse material (CSAM) rose globally by more than 8,400 percent from 2023 to 2025. In 2023, the US organisation National Center for Missing & Exploited Children (NCMEC) received around 4,700 such reports. In 2025, the number was over 400,000 (The Guardian, Mashable).
The analysis is built on reports that technology companies themselves send to NCMEC, not on direct measurements of how much material is actually circulating. It includes deepfakes made with so-called "nudify apps" – tools that undress images of real children.
Childlight urges authorities to treat child sexual exploitation as a preventable public health crisis. The institute estimates that over 300 million children worldwide are affected each year by at least one form of technology-facilitated sexual exploitation or abuse (The Guardian).
The trajectory, year by year – and why the figures must be read carefully
The yearly figures come from different sources and should be kept apart. The Guardian and Mashable cite the 2023 figure (around 4,700) and the 2025 figure (over 400,000). The Verge fills in the intervening year: NCMEC's CyberTipline received 67,000 AI-related reports in 2024, an "exponential" rise continuing to 400,000 in 2025 (The Verge).
The most important caveat comes from Childlight itself. Professor Debi Fry, the institute's global data director, says that "while improved detection and reporting may explain some of the increase, the findings underscore why authorities and technology companies must act quickly to understand and prevent these new forms of abuse and rapidly remove material" (The Guardian). In other words: part of the growth may be due to more platforms detecting and reporting the material, not necessarily to the volume of abuse material having grown in step with the report figures.
A further methodological limitation: this analysis's regional scope (East Asia and the Pacific, Latin America and the Caribbean, North America) differs from Childlight's earlier studies, which covered Western Europe. That limits the ability to compare trends over time.
Where 400,000 reports becomes an infrastructure problem
The CyberTipline is the US central intake point: technology companies are legally required to report detected abuse material there, and NCMEC forwards it to law enforcement. When the volume of AI-related reports rises from 67,000 to 400,000 in a single year, it hits a workflow built on manual review. Researchers have warned that the inflow could make NCMEC's workload "unmanageable", according to The Verge (The Verge).
A broader picture of the scale comes from the Internet Watch Foundation (IWF), a British NGO that monitors abuse material globally. In 2026, the IWF has already assessed 6,310 AI-generated images as meeting the legal definition of child sexual abuse material – 40 percent above the full-year 2025 total, which was more than 4,500 (The Guardian). The IWF and NCMEC figures are measured differently – the IWF assessed material itself, NCMEC logs submitted reports – but both point in the same direction.
OpenAI's response: a classifier that looks at the images, not the hashes
The day after the Childlight analysis was published, OpenAI launched a CSAM classification tool, available free via API to vetted organisations (Fox News, The Verge).
Technically, the tool differs from established solutions in one essential respect. Traditional tools compare hash values – digital fingerprints – of already identified files. That works for known material, but never sees anything newly produced. OpenAI's classifier instead analyses visual characteristics of possible abuse material, and can therefore flag suspicious content that has never been identified before (Fox News). The tool produces scores that human reviewers use in their work, and according to OpenAI it is the first broadly available tool in its class to include a dedicated signal for whether an image may be AI-generated (Mashable).
That AI signal is the point against the backdrop of the Childlight data: while much of the report flood stems from synthetic material made from real children's images, hash-based systems cannot distinguish new synthetic variations from known material.
What OpenAI claims – and what remains to be proven
The performance figures are the company's own, not independently verified. OpenAI says the classifier flagged around 48 percent fewer images – that is, fewer false positives – while finding roughly as many abusive images as existing tools (The Hill). The company also states that images the tool flags are about twice as likely to contain abuse (OpenAI via The Hill and Mashable).
The press has pointed to two important caveats. Fox News writes that the tool needs independent testing (Fox News), and The Hill notes that human review is still required – the classifier scores and prioritises, but does not judge (The Hill). This is AIMag's analysis: the central limitation is not solved by the new signal – an AI-generated image based on a real child still points to a child who may need protection, and the assessment of that rests with humans and law enforcement.
Open questions
Three questions remain unanswered after this week. First, the gap between report volume and true prevalence: Debi Fry's own caveat about improved detection and reporting means the true growth in AI-generated material cannot be read directly from the NCMEC figures. Second, the detection race: it remains to be seen whether detection tools can keep pace with how the generation tools develop, and OpenAI's classifier has not yet been independently tested. Third, the regulatory framework: Childlight urges governments to act and treat the exploitation as preventable, but there is no concrete regulatory response in the available source material.
All figures in this story come via secondary press coverage; the primary reports from Childlight, IWF, NCMEC and OpenAI have not been directly reviewed by AIMag.

