OpenAI Opens a Free CSAM Image Classifier to Vetted Organizations on the Moderation API

OpenAI's CSAM image classifier is free to approved organizations on v1/moderations. What it returns, how access works, measured limits, reporting duties.

OpenAI released its CSAM Image Classifier on November 16. The model scores still images for potential child sexual abuse material (CSAM), runs as omni-moderation-csam-latest on the v1/moderations endpoint, and costs nothing for organizations OpenAI approves. It returns two signals, one for potential CSAM and one for potential AI-generated CSAM, and the changelog describes its job as helping teams prioritize images for human review. API changelog · Announcement

The announcement’s headline figure is OpenAI’s own estimate: measured against its existing tool, the classifier would flag 48% fewer images while finding roughly the same number of abusive ones, and a flagged image was about twice as likely to contain abuse. The model card has the numbers a review team needs. At its default threshold the likely-csam signal caught 80.2% of the CSAM images in a held-out test set, and on sampled ChatGPT traffic 29.8% of its flags were true positives. Announcement · Model card

Every performance number below is OpenAI’s or a partner’s, and we found no independent benchmark. S5 Labs has not run the classifier, which is gated, and nothing here is legal advice.

What a request looks like

The classifier takes a standalone POST /v1/moderations request with exactly one still image of up to 20 MB, passed as an HTTP or HTTPS URL or a base64 data URL. It does not classify text. The result at results[0] holds a flagged boolean, a boolean per category in categories and a score from 0 to 1 per category in category_scores. The categories are likely-csam and likely-csam/gen-ai. Classifier guide

It follows different rules from the general moderation model:

omni-moderation-csam-latestomni-moderation-latest
Access and priceApproved organizations, no costFree to use
InputOne still image per request, up to 20 MBText and images, images up to 20 MB
Image CSAM signallikely-csam, likely-csam/gen-aiNone (sexual/minors covers text only)
Batch APINot supportedSupported
Published rate limitsNone found250 requests per minute (Free) to 5,000 (Grow)

Classifier guide · Moderation guide · Model page

The guide warns against reading the scores as calibrated probabilities or treating 0.5 as a universal threshold, and says a higher score signals a stronger match within the same category and model snapshot. The category booleans follow a decision threshold that the guide calls configured but does not give. The model card lists defaults of 0.996324 for likely-csam and 0.90 for likely-csam/gen-ai, but no page we read says the API’s flagged field uses them. Classifier guide · Model card

Per the guide, a flag does not establish that an image is CSAM, and the classifier reports nothing to authorities. The card adds that one image can score highly on both signals and that the AI-generated score does not verify where an image came from. It also states that images sent to the classifier API are processed without retention and are not used for training or evaluation, a claim we cannot check. Classifier guide · Model card

The classifier cannot run as inline moderation in the Responses or Chat Completions APIs, so an image-generation pipeline has to send each output in its own request. OpenAI’s moderation guide says never to submit known CSAM. Moderation guide

Who can get access

To get access, an organization submits OpenAI’s intake form, completes Persona verification, goes through program onboarding and accepts the program’s terms. The model card names the intended users as trust and safety teams, child-protection NGOs and law enforcement at appropriately authorized organizations. CSAM guidance · Model card

ROOST, which makes the self-hosted review console Coop that OpenAI integrated with at launch, describes a further condition: approval requires that an organization can review flagged content and report it to the National Center for Missing & Exploited Children (NCMEC). That is ROOST’s description, and no OpenAI page we read states it. In the pages we read, we found no written eligibility criteria, approval timeline, statement on availability outside the US, list of supported image formats, or rate-limit and latency figures for this model. ROOST

What the model card measured

Results at the default thresholds:

SignalDefault thresholdHeld-out recallHeld-out precisionPrecision on sampled ChatGPT traffic
likely-csam0.9963240.8020.9910.298
likely-csam/gen-ai0.900.86410.340

Model card

A recall of 0.802 means about one in five of the 6,849 held-out CSAM images scored below the default threshold. For live traffic the card says “we do not measure absolute recall across all traffic.” The card puts the precision gap down to base rates. The held-out set contained a much higher share of CSAM than the sampled ChatGPT traffic, and when the material is rare, a small rate of wrong flags can outnumber the true ones. On that traffic, roughly seven in ten likely-csam flags and two in three likely-csam/gen-ai flags were not true positives. Model card

Other limits the card reports:

  • Of 1,248,985 open-web images from Re-LAION-5B-research, likely-csam flagged 5 and the AI-generated signal none. The card calls this a proxy, not a measured false-positive rate, because filtering does not guarantee the dataset is free of CSAM.
  • likely-csam recall was uneven: 0.493 on infant-tagged images, 0.613 on Black-tagged images and 0.603 on male-tagged images, against 0.802 overall. An LLM assigned the tags and some groups are small, so the card says to treat the results as directional.
  • OpenAI says it lacks the data needed to prevent some overlap between its training and evaluation sets, has not tested the classifier on video frames, and keeps three signals for severe subcategories internal until it validates them further.

Model card

The announcement also relays early-access partner results: reviewers in fal’s testing confirmed 72% of high-confidence flags as novel CSAM, one in four flagged cases in invideo AI’s testing led to a ban, and the classifier detected 90% of cases in SafetyKit’s evaluations of drawings or animations. We could not verify these, and they measure three different things. TechCrunch presents the 48% figure as a finding from partner testing, while OpenAI words it as its own estimate against its existing tool. Read beside the card’s 0.298, the twice-as-likely figure describes an improvement on the old tool, not a review queue that is mostly abuse. Announcement · TechCrunch

Where it fits beside hash matching

The card gives the two techniques different jobs: hash matching identifies previously confirmed CSAM, and classifiers look for newly created, transformed or AI-generated material. OpenAI’s developer guidance lists hash-matching technologies (PhotoDNA, PDQ, YouTube’s CSAI Match and Thorn’s Safer) and hash lists from NCMEC and the Internet Watch Foundation. For novel content it names OpenAI’s classifier next to Thorn’s classifiers in Safer and Google’s Content Safety API. What OpenAI claims as new is the separate AI-generated signal, which it says no other broadly available CSAM classifier has. Model card · CSAM guidance · Announcement

A flagged image then needs a review queue. ROOST describes Coop as free and self-hosted, with review queues and built-in NCMEC reporting. The guidance says to invest in training, support and a well-being program for reviewers. On thresholds, the card suggests favoring precision for automatic blocking and balancing precision and recall for human review, and says any deployment without human review needs its own legal, contractual and operational assessment. ROOST · CSAM guidance · Model card

The guidance treats image generation and editing as higher-risk features, which covers products built on models such as Images 2.6 or Nano Banana 2.1. On OpenAI’s Images API it asks for a stable per-user safety_identifier on generation and edit calls, with emails or usernames hashed first. CSAM guidance

Reporting duties stay with the platform

For US providers the duty comes from 18 U.S.C. 2258A. A provider, defined as an electronic communication service provider or remote computing service, must report to NCMEC’s CyberTipline “as soon as reasonably possible after obtaining actual knowledge” of facts showing an apparent violation of the listed child-exploitation offenses. The statute does not require a provider to monitor users or scan content. A knowing and willful failure to report carries a fine of up to $850,000 for a provider with at least 100 million monthly active users and $600,000 for a smaller one, with higher caps for repeat failures. A completed report also counts as a request to preserve its contents for one year. 18 U.S.C. 2258A · 18 U.S.C. 2258E

None of the sources says whether a classifier flag by itself amounts to actual knowledge. OpenAI’s position is that a score alone should not be treated as confirmation of reportable content. Its guidance tells developers to register with NCMEC’s ESP program or a local hotline and notes that reporting law varies by jurisdiction. Model card · CSAM guidance

The ChatGPT warning and the rest of the release

The same announcement adds ChatGPT warnings that interrupt attempts to find abusive material, explain the consequences and point people toward help. With the warning, OpenAI removes the flagged conversation and pauses access to ChatGPT for 24 hours, while keeping specialist support available. Repeated violations may lead to longer restrictions, account closure or reports to authorities. The release also includes tools for child-safety law enforcement units. All of it landed a day before GPT-6.2 Astra and Sol. Announcement

What to do with it

Apply if the review side exists. A team that handles uploaded or generated images and already has trained reviewers and a reporting path gets a free signal for novel material that hash matching does not cover. Size the queue for the false positives: most flags in the card’s ChatGPT sample were not true positives. Without reviewers, start with the guidance’s checklist and a review tool such as Coop.

Benchmark on your own traffic. The card says results on OpenAI’s traffic do not guarantee performance on another service, and it recommends two evaluations before deployment: what the classifier finds beyond your current tools, and what it does to review volume. Log raw scores with the model name instead of relying on flagged, and set thresholds from your own labeled data, as we suggested for Decisions API probabilities.

Settle the legal questions before deployment. Have counsel decide what a flag and a reviewer’s confirmation mean for your reporting obligations. During onboarding, get the retention terms, rate limits and the threshold behind flagged from OpenAI in writing.

On OpenAI’s published evidence, this is a triage signal for novel and AI-generated images whose main signal missed about a fifth of held-out positives at its default threshold. It should not trigger reports without a person in the loop, automatic blocking needs the separate assessment the card describes, and video platforms should not assume the results carry over.

Sources

OpenAI announcement and model card

Developer documentation

Law

Other

Continue reading.

Insight17 min read

California Sues OpenAI Over Agents That Escaped Its Test Environment: What the Complaint Alleges

California's Nov. 16 complaint says OpenAI admitted losing control of test agents over 80 times. The claims, the relief sought and the egress lessons.

Insight13 min read

OpenAI Gives Four Outside Assessors Inside Access and Details a 35-Second Stop for Research Jobs

OpenAI signed Apollo, METR, Transluce and Goodfire as embedded assessors and described Control Room's 35-second stop. Terms, redaction rights and gaps.

Technical guide29 min read

How AI Watermarking Works: Text, Images, Audio, and What the Mark Actually Proves

How AI watermarks work in text, images, audio and video, what C2PA adds, what Article 50 and California's SB 942 require, and what a detected mark proves.