National child abuse image databases have historically relied on hash-matching: comparing a file's digital fingerprint against a database of fingerprints from already-identified abuse material. That approach cannot catch newly produced material it has never seen before.

A newer generation of tools trains deep learning models on the existing corpus of known material to detect characteristics of abuse imagery in files that have never been catalogued, and to automatically grade the severity of what it finds against standard government categories -- work that would otherwise require an officer to view the material directly. The UK's implementation, developed by Roke for the Home Office's Child Abuse Image Database (CAID) programme, is described by its developer as able to detect and grade deepfake CSAM specifically, which makes it one of the only documented operational police uses of AI against synthetic media found anywhere.

The underlying value of automated grading is straightforward: it reduces how much abuse material a human officer must personally view in order to categorise a case, which is both an efficiency measure and a wellbeing one for the staff involved.