Most CSAM detection still rests on hash matching, which only catches images someone has already seen and logged. OpenAI’s new classifier goes after the ones nobody has seen yet, and that is where the real gap sits. The company announced it this week, along with a free access program for vetted organizations through the OpenAI API.
OpenAI says it is the first broadly available CSAM classifier with a specific signal for whether content may be AI generated. That matters because generative tools, including image and video apps, are now a source of the problem as well as a target. The model was trained and evaluated with help from NCMEC, which runs the CyberTipline.
What the numbers say
OpenAI compared the classifier with its existing protections on user-uploaded images. The claims are specific:
- 48% fewer images flagged, with roughly the same number of abusive images found.
- Flagged images were about twice as likely to contain abuse.
- fal reviewers confirmed 72% of high-confidence flags as novel CSAM.
- invideo AI saw one in four flagged cases lead to a ban.
- SafetyKit measured 90% detection on drawings and animations that sexualize children.
SafetyKit also reported false positives below 1% on legal adult content. Those are partner-reported figures, and the comparison baseline is OpenAI’s own prior tool, not an independent benchmark. Take them as encouraging, not settled.
Why smaller platforms should care
The practical win is for small trust and safety teams. Fewer false positives means less reviewer time spent on exhausting queues, and a scored output lets teams triage. Early partners like fal, invideo AI, and HeyGen are exactly the fast-growing generative startups that ship image and video features before they have a big moderation staff.
OpenAI is not going it alone. It is distributing access through the Tech Coalition and ROOST, joining the Tech Coalition’s Lantern signal-sharing program, and publicly backing Thorn’s Safer. That is the competitive context. Thorn’s Safer is the established option for detection, and hash-sharing systems remain the backbone. OpenAI is positioning this as a complement for novel content, not a replacement.
What is still unclear
OpenAI also mentioned interventions that interrupt harmful behavior and point users to help, developer guidance, and support for law enforcement investigators. The announcement offers little technical detail on those. Nor does it say how vetting works, or how many organizations will get in.
Still, this is more than a press-friendly pledge. A free, tested classifier that handles AI-generated and illustrated material fills a real hole. If independent evaluations hold up the partner numbers, it could become a default layer for any platform that accepts user uploads. Developers building image or video products should apply for access and test it against their own data.




