Built for images no list contains

Hash matching compares an image against previously identified material: turn the image into a fingerprint and check it against catalogs of what has already been found. A visual classifier can assess images absent from that reference set, which is the gap OpenAI's classifier is built for — material that may never have been seen before, scored so human reviewers can decide what to look at first.

Two scores are exposed through the API: likely-csam, and likely-csam/gen-ai for material that may be AI-generated. The two are not mutually exclusive; an image can score highly on both. OpenAI calls it the first classifier with broad availability to carry a specific AI-generated signal. That is the company's claim, and "broad" needs a qualifier: access is free, but only for vetted organizations that apply through an intake form, and neither the announcement nor the model card publishes the approval criteria.

The system was trained, evaluated and designed with assistance from NCMEC, which runs the CyberTipline reporting hub. Early access partners named in the post include fal, invideo AI and SafetyKit.

What the 48% actually measures

In OpenAI's comparison on user-uploaded images, the new classifier flagged an estimated 48% fewer images while finding roughly the same number of abusive ones, and flagged images were about twice as likely to contain abuse. Read carefully, that is an efficiency result: fewer images flagged at roughly unchanged detection in that comparison. The post does not disclose the size of that sample or when the evaluation ran, so the figure cannot be checked or rerun by anyone outside the company.

The model card is more specific, and less flattering. At the default thresholds (0.996324 for likely-csam, 0.90 for the gen-ai score), the held-out evaluation reported precision of 0.991 and recall of 0.802 for likely-csam — 5,490 of 6,849 positive examples — and recall of 0.864 for the gen-ai score. The training set behind it held 505,000 positive and 913,000 negative examples. The card then states that precision was substantially lower in an online evaluation drawn from ChatGPT traffic, and notes the held-out set contained a much higher proportion of abuse material than live traffic does.

The subgroup numbers deserve the same attention as the headline. Recall was lowest for images tagged as showing infants, at 0.493 for the main score. For likely-csam, recall was 61.27% for examples tagged Black, against 84.48% to 90.32% for the other reported race tags, and higher for female-tagged than male-tagged examples (0.893 against 0.603). The card also notes the classifier has not been tested on video frames, that results on OpenAI traffic do not guarantee performance on another service, and that images submitted to the classifier API are processed without retention and are not used for training or evaluation.

None of that makes the tool a bad release. It does mean the offline figures did not transfer unchanged to the reported online evaluation, and any organization adopting it should test it on its own traffic and its own hardest categories before trusting the thresholds.

The ChatGPT side: a pause, not just a refusal

The second half of the announcement changes what happens inside ChatGPT. Per OpenAI, a person who tries to find material that sexualizes children now gets a warning that explains the harm, including when the material is AI-generated. The flagged conversation is removed and access is paused for 24 hours, while specialist support stays available. Repeated violations can bring longer restrictions, account closure or reports to authorities.

OpenAI says it built the intervention with prevention organizations and researchers, including Protect Children, the Lucy Faithfull Foundation, Dr Caoilte Ó Ciardha at the University of Kent, and Joel Scanlan at the University of Tasmania; the independent evaluation named in the post covers a trial of deterrent messages, not the classifier or the full intervention. The company also published CSAM guidance for developers and said it will supply child-safety law enforcement units with AI investigation tools and training, with access managed by program administrators.

What to watch

Two things will decide whether this announcement matters beyond OpenAI's own products. The first is distribution: how many organizations are approved, how quickly, and under what criteria. Access criteria and uptake will help establish its reach. The second is whether OpenAI publishes online-traffic results with the same candor as this first model card. The card is the strongest part of the release. It gives adopters the subgroup gaps and the live-traffic warning up front, instead of leaving them to be found later.

Sources

  1. [1] OpenAI, "Strengthening the fight against online child sexual abuse with new tools and support", 16 November 2026 (announcement read via TechCrunch and PPC Land; listed without a URL because the desk did not confirm the openai.com address)
  2. [2] OpenAI CSAM Image Classifier model card, Deployment Safety Hub, 16 November 2026 (read first-hand by ALFW)Read source
  3. [3] TechCrunch, Ivan Mehta, "OpenAI releases a new image classifier and ChatGPT warnings to combat CSAM", 16 November 2026Read source
  4. [4] PPC Land, "ChatGPT users who seek child abuse material face 24-hour suspensions", 16 November 2026Read source