Google flags 50,000 likely AI-generated child abuse files in 30 days

The new labels are intended to help investigators prioritize cases involving children in immediate danger, but they do not determine whether the harm shown is real or synthetic

Google has introduced labels intended to help investigators identify files likely to contain child sexual abuse material created wholly or partly using AI.

Google says a new system used across Google and YouTube tagged roughly 50,000 files as likely AI -generated child sexual abuse material during a 30-day period.

The labels are being added to reports sent to the National Center for Missing & Exploited Children (NCMEC), giving analysts and law enforcement an early indication of whether material may have been created wholly or partly using AI.

That distinction has become harder to make at speed. Reports to NCMEC involving AI-generated content exceeded 182,000 in 2025, according to Google. Analysts must then assess whether each file is authentic, synthetic or a combination of the two before deciding how urgently it should be investigated.

The system does not make that decision for them. Google says the tag provides additional context and should not be treated as a final determination of whether a file depicts real-life or AI-generated harm.

A label for triage, not a verdict

The 50,000 tagged files represented approximately 6% of all files analyzed by the system during the 30-day period.

Files and reports are not the same thing, as one report can contain multiple files. Google and YouTube sent 53,000 reports to NCMEC during the period, with at least one likely AI-generated file appearing in approximately 11% of them.

For investigators, the purpose is to identify where attention may be needed most urgently. John Shehan, Senior Vice President at NCMEC, says:

“Being able to distinguish likely AI-generated content gives analysts and law enforcement valuable context, helping them prioritize cases involving real children who may be in immediate danger.”

The tag could therefore help with triage, but it does not replace human assessment. It also does not establish whether a real child was harmed in the creation of material that appears synthetic.

Google extends tools for law enforcement

The classifier is one part of a wider update to Google’s work with law enforcement agencies.

A pilot with Costa Rica’s Organismo de Investigación Judicial is expected to reduce the time required for forensic analysis by between 80% and 90%, while also limiting investigators’ exposure to child sexual abuse material. That figure is a projected reduction, not the measured result of a completed pilot.

In the United States, the Wyoming Internet Crimes Against Children Task Force is using Google’s Content Safety API and open-source models to sort incoming CyberTips and digital forensic evidence.

Google is also automatically translating NCMEC CyberTips into local languages in certain jurisdictions. In Thailand, the company says translated reports helped authorities identify 47 victims from 43 reports in 2026.

Automated systems produced 99.57% of Google’s reports

Google has also updated its Global Transparency Report to separate content found through automated detection from material reported manually.

According to the company, proactive systems using hash matching and machine learning accounted for 99.57% of all content it reported to NCMEC during the second half of 2025. That percentage covers Google’s wider detection work. It should not be read as an accuracy rate for the new AI-generated content classifier.

Google’s Child Safety Toolkit, which includes technology intended to help organizations prioritize abusive content for review, remains available free of charge to qualifying partner organizations.