Solution 01
Hash-match against maintained databases as a gated pipeline stage
01 — Root cause addressed
The pipeline relied on a general-purpose classifier for a category that requires exact-match detection against a curated list of known material.
02 — Implementation
Screen every candidate image and URL against hash lists of the PhotoDNA, Internet Watch Foundation and Canadian Centre for Child Protection class before admission to the dataset. Structure it as a stage that must pass, not a score that gets weighed alongside quality signals: a match removes the candidate, and the release does not proceed until the stage has run over the full candidate set. Use image-hash and URL-hash comparison so flagged material never has to be viewed by the team operating the pipeline, which is what allows the check to run at corpus scale and to be re-run on demand.
03 — Prevention
Make the hash-matching stage a hard precondition of publication, record its coverage over the candidate set, and treat an unrun or partially run stage as a release blocker rather than a warning.
04 — Trade-offs
This requires an established relationship with child-safety organisations and access to their hash lists, which small or independent teams may not be able to obtain quickly. It adds a hard external dependency and real processing cost at multi-billion scale. Most importantly, hash matching catches known material only, so presenting it as complete coverage would be a failure in its own right.