Amazon Reported More Than One Million AI-Related CSAM Alerts to NCMEC but Refuses to Disclose Sources

Amazon’s disclosure that it submitted more than one million AI-related child sexual abuse material (CSAM) tips to the National Center for Missing and Exploited Children represents a sudden and dramatic increase in reported incidents tied to model training. The company says most of the flagged files originated in third-party data sources it had incorporated into training pipelines and that it purged the material before ingesting it into models, but it declined to provide investigators with supplier names or provenance details. NCMEC officials warned that the absence of traceable source information has made many of Amazon’s electronic tips impossible to pursue as actionable leads—investigations require origin metadata, host locations or custody details to substantiate criminal or protective measures. By comparison, NCMEC recorded roughly 67,000 AI-related reports in 2024 and about 4,700 in 2023, underscoring either a rapid contamination of datasets or a large change in detection and reporting practice. The episode sits inside a broader ecosystem of risks: independent reviews and platform probes have identified generative-AI applications and other third-party feeds that can produce or embed sexualized imagery of identifiable people, and some of those products have been widely distributed through major app stores and marketplaces. That distribution — combined with monetization incentives and technical strategies by developers to evade moderation — increases the likelihood that illicit content can propagate into compiled training corpora. Cross-border developer bases and differing national rules for data access further complicate evidence preservation and law-enforcement follow-up when potentially criminal content is involved. Regulators in multiple jurisdictions are simultaneously scrutinizing how models prevent dissemination of sexually explicit synthetic images, increasing pressure on companies to adopt stronger pre-deployment audits, supplier due diligence and provable ingestion logs. For investigators, the core operational problem remains converting large volumes of automated flags into verifiable evidence; for platforms and enterprise customers, the priority is contractual and technical mechanisms that make dataset provenance auditable. The incident highlights the limits of conservative over-reporting as a safety posture: it reduces the risk of under-notifying but can generate noise unless paired with transparency that preserves investigatory value. Absent stronger contractual clauses, standardized scanning practices, and cross‑border cooperation on forensic preservation, the gap between discovery and remediation will persist and likely prompt more prescriptive rules on dataset sourcing and reporting.