Measure the review queue at the threshold you would actually use.
OpenAI announced on November 16 that vetted organizations can request no-cost API access to a classifier for potential child sexual abuse material in images, including newly generated material. The model returns a risk score to help prioritize review. In one comparison using user-uploaded images, OpenAI estimates that the tool could flag 48% fewer images while finding roughly as many abusive images as its existing tool. Its model card describes different precision on held-out images and samples of ChatGPT traffic. OpenAI release · Model card
Estimate the queue at your chosen threshold
At that reported reduction, 1,000 flags in a baseline queue would become 520. That arithmetic gives reviewers a staffing example. It does not forecast another platform’s queue. A rollout decision also needs a platform-specific estimate of incremental detections, false flags, and reviewer time at the chosen threshold. The model card says its still-image evaluation does not establish performance on video. MustHave’s CyberTipline reporting explainer



