OpenAI announced a free image classifier for detecting potential child sexual abuse material, including AI-generated imagery, on November 16, 2026. Access is restricted to vetted organizations through the OpenAI API. The company also announced new ChatGPT interventions, developer guidance and support for authorized child-safety investigators.
The classifier looks for visual signs of abuse and returns scores to help safety teams prioritize review. OpenAI says its internal evaluation found it could flag 48% fewer images while identifying roughly the same number of abusive images as its existing protections. That is a potentially useful reduction in review workload, but the figures come from early vendor and partner evaluations, not independent validation across deployments. (OpenAI announcement)
What the Classifier Does and Who Can Access It
OpenAI’s classifier examines images for signs of potential abuse, including unfamiliar material that may not have been identified before. Its scores are intended to help reviewers decide which images need attention first.
The system also provides a signal indicating whether suspected CSAM may be AI-generated. OpenAI describes it as the first broadly available CSAM classifier with this specific capability. That distinction is the company’s claim; the announcement is not a comprehensive comparison of every detection product.
OpenAI says the National Center for Missing and Exploited Children, or NCMEC, assisted with training, evaluation and deployment design. NCMEC operates the CyberTipline, a centralized reporting hub for online child exploitation. (OpenAI)
Free access does not mean unrestricted access. Organizations must be vetted, and OpenAI’s announcement directs interested organizations to an intake form. The company is working with the Tech Coalition and ROOST to help share access.
The announcement does not detail approval criteria, review times, throughput limits or the API’s request and response format. Developers should therefore treat this as an application-based service, not an endpoint they can immediately add to a public application.
My view is that the price matters most for teams whose detection budgets are limited. Removing the API charge lowers one barrier, but integration, secure handling, specialist review and reporting still require resources. A free classifier can reduce part of an organization’s safety costs without making the overall workflow free.
What the Performance Figures Actually Measure
OpenAI’s headline results describe a comparison on user-uploaded images against its current protections. Its partner results cover different datasets and operational outcomes.
| Evaluation | Result reported by OpenAI | What it measures |
|---|---|---|
| OpenAI internal comparison | 48% fewer images flagged while finding roughly the same number of abusive images | Review volume at a similar level of detected abuse |
| OpenAI internal comparison | Flagged images were about twice as likely to contain abuse | The concentration of abusive material in the flagged queue |
| fal | Reviewers confirmed 72% of high-confidence flags as novel CSAM on human-reviewed production images | Confirmation rate within a selected, high-confidence subset |
| invideo AI | One in four flagged cases led to a ban | An enforcement outcome following a flag |
| SafetyKit | Detected 90% of cases involving drawings or animations that sexualize children | Detection within that specific evaluation category |
All of these figures appear in OpenAI’s announcement. They should not be combined into a single accuracy score: they describe different populations, thresholds and outcomes.
Why Fewer Flags Can Produce the “Twice as Likely” Result
The two headline claims are closely related.
Consider a hypothetical review queue containing 1,000 flagged images. A 48% reduction would shrink it to 520 images. If both queues contained approximately the same number of confirmed abusive images, the smaller queue would have about 1.92 times the concentration of abuse.
That calculation uses OpenAI’s reported reduction; it is an illustration, not another test result. It explains why “48% fewer flags” and “about twice as likely to contain abuse” can describe the same underlying improvement.
The practical benefit claimed here is a smaller review queue at a similar detection level. That could mean less reviewer time spent on false positives. The published summary still leaves unanswered how many abusive images both systems miss.
Why the Partner Numbers Need Separate Treatment
fal’s 72% figure applies to high-confidence flags, not every image processed. Without the size and composition of the wider dataset, it cannot establish overall detection performance.
invideo AI’s one-in-four figure measures bans. A ban is a platform decision that can depend on policies and account context, so it is not interchangeable with an image-level confirmation rate.
SafetyKit’s 90% result concerns drawings or animations that sexualize children. It provides information about that evaluation category, but it cannot be extended to all photographs, synthetic imagery or other deployment conditions.
These distinctions matter more than choosing the largest percentage. I’d assess a deployment by tracking review volume, confirmed detections and missed cases separately. A system that creates a cleaner queue is valuable, but teams also need to understand what falls outside that queue.
Why the AI-Generated Signal Matters for Triage
NCMEC’s CyberTipline received 400,000 generative AI-related reports in 2025, up from 67,000 in 2024, according to The Verge’s reporting. That is roughly a sixfold increase. These are report counts, so they should not be read as counts of unique images or victims.
The growth helps explain why prioritization is an important part of this release. Investigators need to distinguish cases requiring different responses while continuing to address harmful synthetic material.
OpenAI’s head of child safety public policy, John Buckley, told The Verge that identifying material that may involve fictional versus real children matters for reporting workflows:
“That triage is critical in law enforcement reporting flows.”
An AI-generation signal can help route a case, but it cannot by itself establish whether a real child is safe or whether further investigation is needed. Buckley also emphasized the harms of AI-generated CSAM and its presence alongside imagery involving real children. (The Verge)
The useful interpretation is narrow: the classifier offers another signal for prioritization. Its value depends on how safety teams combine that signal with review and reporting procedures.
ChatGPT Adds Warnings and a 24-Hour Pause
OpenAI is separately changing how ChatGPT responds to attempts to find abusive material. The company says its models already refuse CSAM requests and that it already takes action against accounts that violate its policies.
The new intervention adds several product-level consequences:
- A warning explains the harm caused by sexualizing children, including through AI-generated material.
- The flagged conversation is removed.
- Access to ChatGPT is paused for 24 hours.
- Specialist support remains available to help people stop harmful behavior.
- Repeated violations may lead to longer restrictions, account closure or reports to authorities.
These are the actions described in OpenAI’s announcement. Conversation removal describes the product action; the announcement does not explain what happens to retained records behind the scenes.
OpenAI says the approach was developed with input from organizations and academic advisers, including Protect Children and the Lucy Faithfull Foundation. It also cites an independently evaluated trial of deterrent messages. That earlier trial should not be confused with evidence that this specific ChatGPT rollout has already reduced abuse.
Frequently Asked Questions
3 questions
1Is OpenAI’s CSAM Classifier Free for Everyone?
No. OpenAI is offering API access at no cost to vetted organizations. Interested organizations must request access through the intake process described in its announcement.