The OpenAI CSAM Image Classifier helps reviewers prioritize images that may contain child sexual abuse material (CSAM), including AI-generated material.
The classifier is available to approved organizations at no cost. Apply for access.
How it works
The classifier takes one image as input and returns signals for two categories: potential CSAM and potential AI-generated CSAM. To use the classifier, set the model to omni-moderation-csam-latest.
For general moderation, use omni-moderation-latest. It can flag sexual content in images, but its sexual/minors category applies only to text.
Classify an image
Follow your organization’s approved input-handling process. Do not submit known CSAM.
- Set
OPENAI_API_KEYto an API key for an approved organization. - Send a standalone Create moderation request (
POST /v1/moderations). - Send exactly one still image of up to 20 MB (20,000,000 bytes) in the
inputarray withtype: "image_url". Setimage_url.urlto an HTTP or HTTPS image URL or a base64-encoded image data URL. The classifier doesn’t support text classification.
The following example sends an image of soap to the classifier.
curl https://api.openai.com/v1/moderations \
-X POST \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "omni-moderation-csam-latest",
"input": [
{
"type": "image_url",
"image_url": {
"url": "https://cdn.openai.com/API/docs/images/soap.png"
}
}
]
}'Here’s an example response:
{
"id": "modr_example",
"model": "omni-moderation-csam-latest",
"results": [
{
"flagged": false,
"categories": {
"likely-csam": false,
"likely-csam/gen-ai": false
},
"category_scores": {
"likely-csam": 0.12,
"likely-csam/gen-ai": 0.01
},
"category_applied_input_types": {
"likely-csam": ["image"],
"likely-csam/gen-ai": ["image"]
}
}
]
}
Interpret the result
Read the classification result at results[0]. Start with flagged, then check the individual category flags.
| Response field | Description |
|---|---|
flagged |
Whether the classifier flagged the image. |
categories |
A boolean flag for each category, based on its configured decision threshold. |
category_scores |
A score from 0 to 1 for each category, with no units. Higher scores signal a stronger match within the same category and model snapshot. Do not interpret these scores as calibrated probabilities or assume 0.5 is a universal threshold. |
category_applied_input_types |
The input types evaluated for each category. Each value is ["image"]. |
Understand the categories
| Category | Signal |
|---|---|
likely-csam |
Signal for potential CSAM. |
likely-csam/gen-ai |
Signal for potential AI-generated CSAM. |
A category flag doesn’t establish that an image is CSAM, and the classifier doesn’t automatically report content to authorities. Follow your organization’s review, escalation, and reporting processes. See the CSAM guidance for recommendations on handling flagged content.
Troubleshoot a request
If a request fails, check that you selected omni-moderation-csam-latest, used an API key for an approved organization, and sent one image within the size limit.
The classifier doesn’t support inline moderation or the Batch API.
For rate-limit errors, consult your organization’s limits and the rate-limit guidance.
Access
To request access, submit the classifier application form.