How can organizations review a large flow of images without turning an automated signal into a final decision? On November 16, 2026, OpenAI announced that its CSAM Image Classifier is available through an API to approved organizations. The model takes one image and returns two separate signals: one for possible child sexual abuse material (CSAM) and one for possible AI-generated CSAM. The tool’s practical purpose is to prioritize images for human review, not to replace it.

That distinction matters: the classifier gives teams another filter for the moderation queue, but its output requires its own handling process. Safety and moderation specialists need to understand not only the label, but also which decisions remain the organization’s responsibility.
What changed on November 16
In its November 16 update, OpenAI said approved organizations can use the classifier for free to prioritize images for human review. This is a precise description of the product’s role: the system sorts material that may warrant attention, while people conduct further review under an approved process.
The classifier documentation describes the API in detail. Select the omni-moderation-csam-latest model and send a separate POST /v1/moderations request. The documentation allows one static image of up to 20 MB per request. Organizations should use their own approved process for handling input data; known CSAM must not be submitted to the classifier.
Why the model returns two signals
The response contains separate categories, likely-csam and likely-csam/gen-ai. This distinction lets teams tell apart a signal for potential CSAM and one for potentially AI-generated material. It can help with routing and queue analysis: different types of detected risk can be handled under the relevant internal procedures.

Each category includes a Boolean flag and a numerical score. OpenAI describes the score as a measure of the strength of the match within that same category and model snapshot. The company cautions that the score is not a calibrated probability and should not be compared with a universal threshold such as 0.5. Queue settings should therefore be based on how a particular team will use the flags and scores, rather than an arbitrary threshold for automated decisions.
A practical sequence looks like this: the organization submits one eligible image, receives the result, checks the overall flag and individual categories, then routes the material for human review under its procedures. Linking the result to the original item and the internal decision helps the team explain why the image entered the queue and who reviewed it.
Where automation ends
OpenAI emphasizes in the classifier documentation that a category flag alone does not establish that an image is CSAM, and that the classifier does not automatically report content to authorities. Organizations must apply their own procedures for review, escalation, and information sharing. The model helps select material for review, while responsibility for interpretation and subsequent actions is built into the team’s workflow.
In this process, the classifier’s output is a priority marker for a specialist. The team should decide who reviews flagged items, how the outcome is recorded, and how material is passed on according to its procedures.
What the tool is designed for
According to the published specification, the classifier is designed for images and returns two signals for the specified categories. OpenAI’s general moderation model, omni-moderation-latest, serves a broader purpose; the documentation notes that its sexual/minors category applies to text. The specialized image classifier addresses a distinct task in an image workflow, and model selection depends on the input type and review process.
When designing a system, it is useful to distinguish three things: the model’s signal, a human assessment, and an operational action. The signal is a machine output; the assessment is the result of a specialist’s review; and the action is the transfer of material or another step under internal procedures. Recording these separately helps avoid attributing to the model decisions that the organization actually makes.
Comparisons with general-purpose models also require precision. The documentation describes the input, two categories, response fields, and usage limitations, but does not provide comparative accuracy tests. When implementing the tool, it makes sense to assess how well it fits the specific process: who has access, how human review works, how outcomes are recorded, and how the team handles incorrect or disputed signals.
What organizations should check before connecting
- Access: The tool is intended for approved organizations; access can be requested through an OpenAI form.
- Input: One image file per request, no larger than 20 MB; this model does not support text classification.
- Interpretation: Category flags are set using configured thresholds, and numerical scores are not calibrated probabilities.
- Process: Human review, escalation, and any reporting of material remain the organization’s responsibility.
In short, the November 16 change adds a specialized image-prioritization filter for approved organizations. Its place in the workflow is before human review: the team receives two machine-generated signals, assesses them under its own rules, and determines what happens next.