OpenAI Releases a CSAM Image Classifier and New ChatGPT Warnings

OpenAI has announced new tools to combat child sexual abuse material (CSAM), including an image classifier that detects such images, AI-generated ones among them, and new warnings inside ChatGPT. The company set out the work in a Safety newsroom post titled "Strengthening the fight against online child sexual abuse," published on November 16, 2026. OpenAI News
The classifier was developed with the National Center for Missing and Exploited Children (NCMEC). OpenAI says it will give select companies access through its API, the programming interface developers use to call a model from their own software, at no cost. It is also working with Tech Coalition and ROOST to share access more widely. TechCrunch
In early testing, the classifier flagged 48% fewer images while finding the same number of abusive images.
On the ChatGPT side, OpenAI is adding warning screens meant to keep users from reaching material that sexualizes children. When a user attempts a prompt seeking such material, ChatGPT will show an alert and direct the user toward specialized support. The product will also suspend the account for 24 hours as a warning. Repeated offences could lead to permanent removal of the ChatGPT account.
OpenAI is also publishing guidelines for developers who build on its tools, covering how to prevent CSAM, detect it, and help users report it. That part matters for platforms and application builders downstream of the API, since the controls they inherit will depend partly on how they implement those guidelines.
The broader context here is architectural. A classifier offered through an API moves detection logic out of a single provider's stack and into third-party pipelines, where operators set their own thresholds, retention rules, and human review processes. The sharing arrangement with Tech Coalition and ROOST points toward that wider distribution, and the parties involved will need to agree on how flagged content is handled and reported. Practitioners will want to know whether the 48% reduction holds up at production volume and across the image types and formats their platforms handle. The announcement does not yet answer those questions.
In my view, the more useful reading of this announcement is as a template for how detection capability might be shared rather than rebuilt by each platform. Whether that model works will depend on details still to come, including the access criteria for the select companies and the evidence gathered once the classifier runs at scale. If the sharing arrangements work as described, more services could adopt a detection layer without building their own from scratch, which matters for smaller operators with limited tooling.


