Independent newsroom

Tuesday, November 17, 2026

The File

Every claim traces to a primary source. Uncertainty is labeled. Corrections are public.

AI · Tuesday, November 17, 2026

OpenAI offers approved organizations a free tool to flag suspected child abuse images

The File desk · Nov. 17, 2026, 6:18 a.m. ET

Confirmed2 primary sourcesAdded 6:18 a.m. ET, Nov. 17

Status

Confirmed. The model card and the platform changelog are OpenAI's own pages, both opened November 17. The performance figures on the card are OpenAI's measurements. This story does not use the announcement blog, which returned 403.
  • Confirmed
    The card, published November 16, 2026, describes a no-cost classifier for approved organizations, the two public scores, and the limits OpenAI states.
  • Confirmed
    The changelog lists the model name omni-moderation-csam-latest on the moderations interface.
A check against known files can only catch what someone has already found. On Monday, November 16, 2026, OpenAI published a model card for a tool that scores images for possible child sexual abuse material, including images such a check can miss.1 Approved organizations can use it at no cost.1
The card calls the tool the OpenAI CSAM Image Classifier.1 CSAM is the card's shorthand for child sexual abuse material.
To see what is new, put the two methods side by side. Hash matching compares an image with a library of files that have already been identified. A classifier instead returns a risk score, which can flag an image the library has never stored, including material the card calls potentially novel or made with generative AI.1 Why does that gap matter now? The card says generative AI can raise the volume of that new material.1
The score is meant to help a platform route images to trained people for review.1 But notice what the card does not claim for it. It states the limit in one sentence: "A classifier score alone does not confirm that an image is CSAM, establish its legal status, or determine whether it must be reported."1 The tool is meant to sit beside hash matching, contextual abuse detection, human review, and legal reporting.1
OpenAI says it built the classifier with assistance from the National Center for Missing and Exploited Children, the nonprofit known as NCMEC.1 The design has two parts. The first is an internal CLIP image encoder. CLIP, the card says, learns visual patterns from pairs of images and text, and the encoder turns an image into a numerical summary.1 The second, a three-layer multilayer perceptron, turns that summary into scores.1 The public interface exposes two of those scores, likely-csam and likely-csam/gen-ai, and other subcategory signals stay inside the company.1
OpenAI's platform changelog for November 16, 2026 lists the model as omni-moderation-csam-latest. It says approved organizations can use the OpenAI CSAM Image Classifier at no cost to help prioritize images for human review.2 The model returns separate signals for potential child sexual abuse material and for potential AI-generated material.2
What has not been shown matters too. The card says the classifier has not been tested on video.1 Its online test used a sample of ChatGPT traffic, and the card says those results may not carry over to another platform.1 So how will the tool do on someone else's site? The card itself cautions that its own results may not answer that.
The intended users are trust-and-safety teams, non-governmental organizations that work to prevent child sexual abuse, and law enforcement.1
Why it matters (Estimate). Estimate. A trust-and-safety team can now add a score for pictures that hash matching, which looks only for known files, may not catch, including pictures a generative system made. The card says the score is not a legal finding and should not, by itself, decide whether an image must be reported. So a platform that uses the tool still needs trained reviewers. This is our reading of the card and the changelog.
What the classifier offers is a score, not a verdict. For any organization that wants it, access starts with an intake form.1 And that is where things stand.
How we know

Notes

  1. 1.OpenAI Deployment Safety Hub, "OpenAI CSAM Image Classifier Model Card," published November 16, 2026. Intake form: https://openai.com/form/csam-classifier-intake. The card says images already confirmed as CSAM should not be submitted to the service. Performance numbers on the card are OpenAI's own measurements and are not repeated here. Retrieved by this desk Nov. 17. Confirmed. Source 1. https://deploymentsafety.openai.com/csam-image-classifier
  2. 2.OpenAI platform changelog, November 16, 2026 feature entry for omni-moderation-csam-latest on v1/moderations. Retrieved by this desk Nov. 17. Confirmed. Source 2. https://platform.openai.com/docs/changelog

What is still unknown or disputed

  • How the tool performs outside the ChatGPT sample. The card says those results may not carry over.
  • What the announcement blog adds. A fetch of that page returned 403, so this story does not use it.
  • Video. The card says the classifier was not tested on video.

Primary sources

Every claim in this story is drawn from the documents below. If a fetch failed, that is recorded on the card.

  1. Source 1

    OpenAI CSAM Image Classifier Model Card

    OpenAI · November 16, 2026

    A classifier score alone does not confirm that an image is CSAM, establish its legal status, or determine whether it must be reported.

    https://deploymentsafety.openai.com/csam-image-classifier

  2. Source 2

    Platform changelog, November 16, 2026

    OpenAI · November 16, 2026

    Approved organizations can use the OpenAI CSAM Image Classifier at no cost to help prioritize images for human review. The model returns separate signals for potential child sexual abuse material (CSAM) and potential AI-generated CSAM.

    https://platform.openai.com/docs/changelog