On November 16, 2026, OpenAI announced a new classifier that detects images suspected of being child sexual abuse material (CSAM) and routes them to human review. Approved organizations that pass screening can use it free of charge through an API. OpenAI also announced that ChatGPT will show warnings and temporarily suspend access for users who try to seek out CSAM.

In its initial evaluation, the company estimates that the classifier can detect roughly the same number of abuse images as existing safeguards while cutting the number of images flagged for review by 48%. The new classifier is designed to handle images that conventional detection methods struggle to catch, and to ease the burden on reviewers.

AD

Adding a classifier for new suspected material alongside known-image matching

A standard way to detect known CSAM is perceptual hash matching, which compares image "fingerprints." It can find matches to registered images even if they have been resized or compressed. OpenAI's new classifier is meant to complement this kind of known-image matching by detecting newly created or altered images, as well as suspected AI-generated CSAM.

According to the published model card, the classifier combines an internal OpenAI CLIP image encoder with a three-layer multilayer perceptron (MLP). The image encoder converts the input image into a numerical representation, and the MLP calculates a score for each category. The MLP was trained on CSAM data, and some layers of the image encoder were also tuned. OpenAI says it chose this architecture to balance detection performance on unseen data against processing speed.

The API returns two types of scores: "suspected CSAM" and "suspected AI-generated CSAM." These are not mutually exclusive categories, and both scores can be high for the same image. The former indicates the likelihood that an image constitutes CSAM; the latter assesses suspicion that it is AI-generated CSAM. Neither is a mechanism for proving how an image was generated or where it came from.

OpenAI states explicitly that a classifier score alone cannot establish that an image is CSAM. Scores are meant to be used for selecting and prioritizing items for review, while judgments about content and any legal response are made by personnel with appropriate authority. OpenAI also says images sent to the classifier API are processed without being stored, and are not reused for model training or evaluation.

48% fewer images to review, but detection accuracy figures need careful reading

The 48% reduction is an estimate based on an initial evaluation comparing existing safeguards with the new classifier on user-uploaded images. OpenAI estimates that it can flag 48% fewer images for review while detecting roughly the same number of abuse images. The "roughly double" figure cited in the same announcement refers to an improvement in the share of flagged images that were actually abuse images.

If fewer images are flagged and a higher share of them are true positives, reviewers can examine fewer images to get the same result. However, the 48% figure is an estimate of image counts. It is not a measured result showing that review time or the psychological burden on reviewers fell by the same proportion.

Two concepts matter for understanding the classifier's performance: precision and recall. Precision is the share of images the classifier flags as positive that are actually positive. Recall is the share of truly positive images that the classifier correctly detects. Setting a stricter decision criterion may reduce the number of images to review, but it may also increase the number of positive images that go undetected.

Table 7 of the model card shows that, even with the same decision threshold, precision varies greatly depending on the set of images being evaluated. The decision threshold is the cutoff that determines which images are sent for review based on the score the classifier outputs.

Signal Offline evaluation precision Online evaluation precision on ChatGPT traffic
Suspected CSAM 99.1% 29.8%
Suspected AI-generated CSAM 100% 34.0%

Note: The same default threshold was used for both offline and online evaluation of each signal. The thresholds are 0.996324 for suspected CSAM and 0.90 for suspected AI-generated CSAM. The offline evaluation used images not used in training, while the online evaluation used samples drawn from the high-score range of representative ChatGPT traffic, so the evaluated populations differ.

The offline evaluation used a total of 46,849 images. Of these, 6,849 were CSAM-positive images, including 2,275 AI-generated CSAM-positive images. At the default threshold, recall was 80.2% for CSAM (5,490 of 6,849 images detected) and 86.4% for AI-generated CSAM (1,966 of 2,275 images detected).

According to OpenAI, the offline evaluation images contain a higher proportion of CSAM than actual ChatGPT traffic does. In an environment where CSAM is rare, even a small share of benign images being wrongly flagged can sharply reduce the proportion of true positives among the images to review. This is why a classifier showing precision above 99% in offline evaluation can fall to around 30% in real traffic.

The online evaluation also targets images drawn from the high-score range, and its recall figures are designed for relative comparison between thresholds. Organizations adopting the classifier should not apply the published results directly to their own environments. Instead, they need to use the images they actually handle to evaluate how many items would be flagged for review and how many positive images would be detected.

AD

How detection results feed into review and reporting

Using the classifier API requires submitting an application form, identity verification through Persona, an onboarding process, and agreement to the terms of use. OpenAI envisions use by online-service safety teams, NGOs working on child protection, law enforcement agencies, and similar groups.

There are two main ways to test the benefit of adopting it. One is to apply the classifier to images that people already review under an existing system and see how far it can reduce the number requiring review. The other is to apply it to images that the legacy system has not flagged and assess whether it finds any CSAM that was previously missed. Because automatic blocking and human review demand different levels of accuracy, the decision threshold is adjusted to the use case.

OpenAI recommends combining the new classifier with perceptual hash matching of known images, and having trained personnel review the suspected images it detects.

Step Target / role Next action
Perceptual hash matching Matches image fingerprints against known CSAM Existing response procedures
New classifier Scores suspected unknown CSAM Human review
Trained personnel with appropriate authority Check image content and local legal requirements Necessary response and reporting

Note: Based on the model card's recommended setup and the developer guide. A classifier score alone does not determine whether something is CSAM or whether a reporting obligation applies; judgment must follow local legal requirements and the organization's authority.

Integrating the classifier into actual safety measures also requires a system for responding after detection. The developer guide recommends providing a reporting channel for users and using safety identifiers to link uploaded images with users. It also calls for defining record-retention policies, responsible personnel, and emergency reporting contacts, and for providing training and psychological support for reviewers. The approach is to build a single process that runs from classifier detection through human review to reporting where necessary.

ChatGPT adds warnings and suspensions, with an appeals process

Separately from the classifier for outside organizations, OpenAI also announced intervention measures in ChatGPT against attempts to seek out CSAM. When such behavior is detected, the measures include showing a warning, blocking access to the conversation, suspending use for 24 hours, and directing users to professional support resources. The support referrals are intended to help people stop the problematic behavior. OpenAI says repeated violations could lead to longer restrictions, account closure, and reports to authorities.

According to OpenAI's help page, measures such as warnings may be taken when a request seeking CSAM is detected. When a warning appears, the conversation is stopped, and if the user is logged in, it is hidden from their history and cannot be resumed. However, being hidden from history is not the same as the conversation data being deleted. If a temporary suspension is applied, the user is signed out and can no longer sign in to or use the service with that account.

The measures currently apply to the Chat and Work modes on web and mobile when used with personal accounts. Warnings may also be shown to users who are not logged in. Business, Enterprise, Edu, and Healthcare workspaces, as well as voice mode, Codex, and the desktop app, are not covered at present. The measures are being rolled out in stages around the world, and what is displayed may differ by country.

If an account is suspended, the user is notified by email of the reason, the length of the suspension, and the steps they can take. Users who disagree can file an appeal through the link in the notification email. If a user receives a separate notice about permanent closure, they should follow the procedure described in that notice. The announcement cites a 24-hour suspension, but the current help page describes multiple intervention measures, so the specific suspension period and conditions for reinstatement should be confirmed in the notice a user actually receives.