An 8,400% surge in AI child abuse reports? What the headline number really measures
Since November 16, one number has been making the rounds worldwide: reports of AI-related child abuse material to NCMEC, the US reporting center, supposedly rose 8,400 percent in two years. We took that number apart. It's cleaner than earlier headlines, but it doesn't tell you how many new AI images are being made. Other figures show the real problem much more clearly.

What this is about
"Reports of AI-related child abuse material up 8,400 percent in two years." The headline started circulating on November 16 via the Guardian and the UK news agency PA, and it has been reprinted around the world ever since. It comes from the new "Into the Light Index 2026" by Childlight, a child protection institute at the University of Edinburgh. The index was presented at a ministerial summit in the Philippines. The math behind it: NCMEC, the US clearinghouse for tips about child sexual exploitation, received around 4,700 AI-related reports in 2023. In 2025 it received more than 400,000.
We originally set out to show that this number was mostly inflated by Amazon. In 2025 the company filed more than a million suspected-abuse reports from its AI training data, almost all of them false positives. Our reporting proved that wrong: as far as NCMEC's data shows, the 400,000 is already the cleaned-up figure, without Amazon's false positives. Our thesis still stands, just for a different reason. The 400,000 mostly reflects how companies report and how NCMEC categorizes those reports. It says very little about how much new AI material is being created. The real problem exists all the same, and other numbers document it better.
The Amazon story: 1.1 million reports, 99.6 percent false positives
The backstory explains why you should be wary of AI statistics on child sexual abuse material (CSAM). In the first half of 2025, NCMEC recorded around 485,000 AI-related reports. Media outlets read that as an explosion of AI-generated images. In January 2026, Bloomberg revealed that about 380,000 of them came from Amazon. Amazon had scanned public web data it planned to use for training its own AI models, using hash matching. That technique computes a digital fingerprint for every image and compares it against a list of known abuse images. According to Bloomberg, Amazon said it "intentionally uses an over-inclusive threshold for scanning, which yields a high percentage of false positives." And not a single one of those reports involved AI-generated material.
Fallon McNulty, who runs NCMEC's CyberTipline, told Bloomberg the volume was "really an outlier." The reports weren't actionable, she said, because Amazon didn't disclose where the data came from. Stanford researcher Riana Pfefferkorn calculated in an open letter to NCMEC that at least 78 percent of the AI reports from the first half of 2025 contained no AI-generated material at all.
Amazon has since rechecked. In its transparency report and a letter to Senator Chuck Grassley, the company lists 1,120,171 suspected-abuse reports for 2025. Of those, 1,098,047 concerned images and videos from the public web. After human review, 4,376 confirmed cases remained, or 0.4 percent. Amazon retracted the remaining 1,093,671 reports as false positives.
What's actually inside the 400,000
The Childlight figure is cleaner than earlier headlines. One version of NCMEC's data page still listed "1.5 million AI-related reports" for 2025, "more than 1.1 million" of them from Amazon involving training data. The current page says "more than 400,000." Subtract 1.1 million from 1.5 million and you land almost exactly on that 400,000. The new breakdown lists only just over 12,000 training-data reports. Our take: Amazon's false positives have largely been removed from the 400,000. NCMEC never says so explicitly, though.
That still doesn't make the number proof of "400,000 AI abuse images." Here's how NCMEC breaks it down:
| Category per NCMEC (2025) | Reports | Share (rounded) |
|---|---|---|
| AI involvement present, but unclear how AI was used | over 200,000 | approx. 50% |
| AI used to edit an existing abuse file, without a text prompt | over 145,000 | approx. 36% |
| Attempt to generate AI material via file upload and text prompt | over 30,000 | approx. 8% |
| Abuse material found in training data | over 12,000 | approx. 3% |
| Users generating or possessing AI abuse material | over 7,000 | approx. 2% |
| Other, e.g. exploitation via chat | over 3,000 | approx. 1% |
So only about two percent of reports describe the scenario most people picture when they read the headline: someone using AI to generate abuse images. In half the cases, nobody knows what role AI played. Then there's the form itself. It has exactly one "Generative AI" checkbox, which the technical documentation describes as being for files that are "believed to be AI-generated." Pfefferkorn criticizes the fact that providers tick that box for very different reasons: for training-data finds, for uploads to chatbots, or for images that were actually generated. According to her January 30 update, NCMEC is only at the start of a multi-year modernization of its reporting interface.
Individual large reporters also skew the curve. According to its transparency reports, OpenAI sent exactly 75,027 reports to NCMEC in the first half of 2025, up from 947 in the first half of 2024. For the second half of 2025, OpenAI lists 107,817 reports. Together that's just under 183,000, nearly half of the 400,000. OpenAI says it reports every instance of abuse material, including uploads and mere requests. According to reports, part of the increase is because OpenAI added more review staff in late 2024 and rolled out more image upload features. Whether all of OpenAI's reports end up in the AI category isn't publicly broken down.
Finally, there's the baseline. NCMEC only started tracking AI-related reports in January 2023. Percentages calculated from such a small base in the first year of tracking balloon quickly. Childlight itself adds a caveat: data director Debi Fry said better detection and reporting could explain part of the increase. Most reprints leave that sentence out.
The more reliable numbers: what analysts actually review
If you want to know how much AI abuse material is in circulation, look at numbers backed by human review of the material:
- NCMEC itself: From January 2023 to December 2025, the center's staff classified more than 158,000 images and videos from reports as AI-generated. Over the same period, the reporting companies had flagged only just over 11,000 files that way. NCMEC also says more than 275 victims of AI abuse material have been identified since 2023. These are real children whose images were manipulated.
- Internet Watch Foundation (IWF): The UK organization has analysts assess every image under UK law. In the first half of 2026, they found 6,310 illegal AI images, 40 percent more than in all of 2025 (4,512). 98 percent depict girls, 79 percent show children aged 7 to 13, and 190 images show toddlers under two. 88 percent fall into Category C, the least severe tier, covering things like sexualized posing and nudity. The trend in video is even starker: the IWF found 13 AI videos in 2024 and 3,443 in 2025. 65 percent of them fall into the most severe Category A. The IWF also acknowledges that its numbers depend on where analysts actively search and what gets reported.
These numbers are smaller than the headline, but they count reviewed material. And they're rising sharply too.
The real problem: nudify apps and Grok
Court filings show how AI abuse works in practice. Elon Musk's AI company xAI faces several lawsuits in federal court in the Northern District of California. One class action was filed on March 16, 2026, and expanded in July to include additional plaintiffs and Stability AI as a defendant. In one case, a stepfather allegedly used Grok to generate more than 7,000 images from a photo of his then eleven-year-old stepdaughter. A second class action followed on August 26. It accuses xAI of training Grok on known abuse material. The Center for Countering Digital Hate estimates that Grok generated more than three million sexualized images between December 29, 2025, and January 8, 2026, more than 23,000 of which likely depicted children. None of the cases has reached a verdict yet.
Then there are "nudify" apps, which digitally undress people in photos. The City of San Francisco has been suing operators of such sites since August 2024. By February 2026, according to the trade publication Indicator, there had been two settlements and 11 sites had been shut down. One operator paid a $100,000 penalty. In July 2026, the city asked Apple and Google to remove 13 such apps from their stores. In 2025, Meta sued the developer of the CrushAI app in Hong Kong for running ads on Facebook and Instagram at scale despite a ban.
And in Germany?
We couldn't find a comparable German figure. The BKA's 2025 federal situation report, published in July 2026, counts 41,677 cases of child sexual abuse material, 2.7 percent fewer than the previous year. It warns in general terms that AI makes it possible to "alter real content or artificially generate new content without in-depth expertise." It doesn't give a separate AI metric. NCMEC sent the BKA more than 100,000 criminally relevant tips in 2025; how many involve AI remains unknown. jugendschutz.net's 2025 annual report, with 15,099 violation cases, doesn't break out an AI share either. Kerstin Claus, the German federal government's Independent Commissioner for Child Sexual Abuse Issues, said at the BKA presentation, in a different context:
"What isn't a crime isn't counted, and what isn't counted isn't fought."
Our assessment
The 8,400 percent isn't fake news, but it is a misleading number. It counts reports, not images, offenses, or victims. It depends on how many companies report, how thoroughly they scan, and how they tick a single, ambiguous checkbox. The 2023 baseline comes from the first year NCMEC tracked AI cases at all. By our math, Amazon's false positives are no longer in the Childlight figure. But only about 7,000 of the 400,000 reports describe generating or possessing AI abuse material.
Our take: other numbers are a better basis for justifying legislation. Those include the IWF's analyst-reviewed finds (6,310 images in six months, videos up from 13 to 3,443), NCMEC's own classification of more than 158,000 AI files, and more than 275 identified victims. Add the concrete cases from the lawsuits against xAI and the nudify operators. Together they show a growing, real problem that hits girls hardest. It doesn't need an inflated percentage. If you're basing policy on it, you should push for precise reporting forms that record training-data finds, uploads, and generated material separately. Germany would first need to start counting at all.
Sources
- Irish News (PA): Reports of AI-generated child sex abuse images soar 8,400% irishnews.com
- UA.News: Zusammenfassung des Guardian-Berichts ua.news
- Startup Fortune: AI child abuse reports have jumped 8,400% startupfortune.com
- explainx.ai: Childlight 8,400% AI CSAM Reports explainx.ai
- Childlight: Into the Light childlight.org
- NCMEC: Generative AI (Aufschlüsselung 2025) ncmec.org
- NCMEC: CyberTipline Data (aktuelle Fassung) missingkids.org
- NCMEC: CyberTipline Data (Fassung mit 1,5 Mio.) missingkids.org
- Amazon: CSAM Transparency Report 2025 aboutamazon.com
- Stanford CIS: Letter to NCMEC about AI CSAM report statistics (Riana Pfefferkorn) cyberlaw.stanford.edu
- Engadget: Amazon discovered a 'high volume' of CSAM in its AI training data engadget.com
- The Star (Bloomberg): Amazon found 'high volume' of child sex abuse material in AI training data thestar.com.my
- Let's Data Science: OpenAI Sends 75,027 CyberTipline Reports To NCMEC letsdatascience.com
- OpenAI: Trust & transparency openai.com
- IWF: More AI-generated child sexual abuse images found in the first six months of 2026 than all of 2025 iwf.org.uk
- IWF Annual Data & Insights Report 2025: AI-generated CSAM iwf.org.uk
- CyberScoop: Deepfake CSAM lawsuit against xAI, Grok expands cyberscoop.com
- AI Lawsuit Tracker: xAI ailawsuittracker.com
- ITHome (Bloomberg-Bericht zu Grok-Klagen) ithome.com
- Indicator: San Francisco's city attorney sued the operators of 22 AI nudifiers indicator.media
- TechCrunch: Meta sues AI nudify app Crush AI techcrunch.com
- BKA: Pressemitteilung Bundeslagebild Sexualdelikte 2025 bka.de
- jugendschutz.net: Vorstellung des Jahresberichts 2025 jugendschutz.net
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.