News

OpenAI releases CSAM image classifier, adds child-safety interventions

OpenAI released a restricted-access CSAM image classifier, added ChatGPT account interventions and published developer guidance; investigator training remains planned.

D
Nov 16, 2026 · 3 min read

OpenAI released an image classifier for child sexual abuse material, or CSAM, through its API on November 16. The launch also includes new ChatGPT abuse interventions, developer guidance and support for child-safety investigators. The classifier is not an unrestricted public API: it is available at no cost only to approved organizations after application, verification, onboarding and acceptance of applicable terms.

For vetted organizations, the classifier provides a new risk signal for triaging images, with trained human review still required for classifier-based reports. In OpenAI’s evaluation of images uploaded by users, the company said the classifier flagged 48% fewer images than its existing tool while identifying roughly the same number of abusive images. It also said flagged images were about twice as likely to contain abuse. OpenAI did not disclose the sample size or a complete public methodology for the comparison, and the result has not been independently established.

The API model, omni-moderation-csam-latest, accepts one still image per moderation request and returns separate likely-csam and likely-csam/gen-ai signals. According to the API documentation, it does not support text, inline moderation or the Batch API. A score does not confirm that an image is CSAM, prove that it was AI-generated, make a legal determination or automatically report it to authorities.

The classifier model card says the system combines an internal CLIP image encoder with a three-layer multilayer perceptron, a small neural network trained to classify the resulting image representation. OpenAI says submitted images are processed without retention and are not used for training or evaluation. The company developed the system with assistance from the National Center for Missing and Exploited Children, which operates the CyberTipline.

OpenAI also reported results from a held-out evaluation of 46,849 images. At its default thresholds, the company measured 0.991 precision and 0.802 recall for likely-csam, and 1.000 precision and 0.864 recall for likely-csam/gen-ai. Precision measures how often flagged material matched the evaluation label; recall measures how much labeled material the classifier found. OpenAI cautioned that performance can differ outside ChatGPT traffic, that related images may remain across evaluation splits despite perceptual deduplication, and that the system can produce false positives and false negatives. The model card calls for a shared independent benchmark.

Results also varied across OpenAI’s demographic tags. Recall for examples tagged Black was 61.3%, compared with 84.5% to 90.3% for the other reported race tags. Recall was also lower for male-tagged examples than for female-tagged examples. OpenAI called those comparisons directional because a language model assigned the tags and some groups had small sample counts.

In ChatGPT, OpenAI said an attempt to find abusive material now brings an explanation and support resources, removal of the flagged conversation and a 24-hour access pause. Repeated violations may lead to longer restrictions, account closure or reports to authorities. The changes add another intervention layer alongside OpenAI’s recent ChatGPT violence safeguards and tighter teen weapons limits.

The company also published developer guidance on prevention, perceptual-hash matching, classifiers for previously unseen material, user-level attribution, human review, response playbooks and registration for reporting. OpenAI describes the guidance as a starting point, not legal advice or a universal standard of care.

For authorized child-safety law-enforcement units, OpenAI described a first tool that can review long chat logs and legal-process returns, identify potential connections and build source-linked timelines. Program administrators manage access, but OpenAI did not name the tool, identify participating agencies or say how many units currently have access. The initial investigator training and field guide are still planned; later training at regional, national and international conferences remains a proposal.

More news