OpenAI today released an image classifier that detects suspected child sexual abuse material (CSAM), including a separate signal for images that may be AI-generated, and is offering it at no cost through its API to vetted organizations. The company also introduced warnings in ChatGPT that remove a flagged conversation and pause access for 24 hours when someone tries to find abusive material, published CSAM guidance for developers, and said it will supply child-safety law enforcement units with AI investigation tools and training. The post sets out results from OpenAI's own tests and from early access partners fal, invideo AI and SafetyKit, and a model card for the classifier went up the same day.
In Short
OpenAI made a free tool that helps approved safety teams spot child sexual abuse images, including fake ones made with AI, and it now shuts people out of ChatGPT for a day when they try to find such material. This matters to anyone who runs a platform where people upload or generate images, including the AI image and video tools many marketers use, and to advertisers buying space next to that content. Organizations have to apply to get the classifier, ChatGPT users who keep trying face longer restrictions, account closure or reports to authorities, and developers now have a written checklist from OpenAI.
Four parts, one announcement
According to OpenAI, the release is made up of four pieces of work, which it describes as building on its Child Safety Blueprint and "our ongoing work to detect, block, and report abuse." The company frames the case for it in industry terms. "It is well established that combatting online child predators requires coordination across many organizations," OpenAI wrote, adding that "individual companies like OpenAI have a responsibility to prevent our technology from being exploited by predators to harm children." That sentence links to a toolkit on preventing child sexual exploitation and abuse from Australia's eSafety Commissioner and to the Tech Coalition's Lantern program.
| Component | Who it is for | Access or terms, according to OpenAI | Organizations named |
|---|---|---|---|
| CSAM image classifier | Child safety organizations, trust and safety teams, the broader industry | Free through the OpenAI API for vetted organizations; request through an intake form | NCMEC (training and evaluation), Tech Coalition and ROOST (sharing access), fal, invideo AI, SafetyKit (early testing) |
| ChatGPT warnings | People who try to find abusive material in ChatGPT | Conversation removed, 24-hour pause, support offered; repeat violations can lead to longer restrictions, account closure or reports to authorities | Protect Children, Lucy Faithfull Foundation, Dr. Caoilte Ó Ciardha (University of Kent), Joel Scanlan (University of Tasmania) |
| CSAM guidance for developers | Organizations at different stages of building safety systems | Published on OpenAI's developer site | Tech Coalition, Thorn, NCMEC, Internet Watch Foundation |
| Law enforcement tools and training | Specialist child-safety investigation units | Access managed by program administrators; depends on agency authorization and program requirements | Thorn (earlier public discussion) |
A classifier built for images nobody has seen before
The centerpiece is an image classifier that, according to OpenAI, "looks for visual signs of abuse, helping teams identify novel images that may not have been seen before." It does not issue verdicts. It "returns scores that help reviewers decide which images may need attention first," which places a human in the loop at the point where an image is confirmed or dismissed. OpenAI claims it is "the first CSAM classifier with broad availability that has a specific signal on indicating if CSAM may be AI generated."
That distinction matters because most industry detection has relied on matching images against lists of known material. Hash matching compares a digital fingerprint of an uploaded file with fingerprints of images already identified and catalogued. It cannot recognize an image created yesterday by a generative model. In April, PPC Land reported that Google accounted for about 74% of hashes on NCMEC's industry hash-sharing list and that Google's Hash Matching API had helped NCMEC tag more than 26 million images, figures attributed in that article to NCMEC's chief operating officer. A classifier trained to recognize visual characteristics addresses the gap that lists of known files leave open.
OpenAI says its teams "trained, evaluated, and designed the classifier for real-world deployment with assistance from the National Center for Missing and Exploited Children (NCMEC)," the US nonprofit that operates the CyberTipline, a central reporting hub for online child exploitation. Access is free. "We're offering access to the classifier to vetted organizations through the OpenAI API at no cost, making CSAM detection more accessible to child safety organizations and the broader industry, including those with teams of varying sizes," the company wrote.
Distribution runs partly through third parties. OpenAI says it has "engaged the Tech Coalition and ROOST to help us share access to this new tool," and describes itself as a supporter of Thorn's Safer CSAM detection system. The company also says it "recently began participating" in the Tech Coalition's Lantern program, "where we will share and receive signals relating to online child sexual exploitation and abuse to strengthen cross-platform collaboration." No start date for that participation is given.
John Shehan, Senior Vice President at NCMEC, is quoted in the post. "This new classifier gives more organizations access to tools that can help detect previously unidentified CSAM, strengthen reporting and provide critical information to those working to identify and protect victims," he said. "We're grateful for OpenAI's investment to develop technology that strengthens the broader child safety ecosystem."
What the model card adds
OpenAI published a model card for the tool, titled OpenAI CSAM Image Classifier, on the same day. According to the card, the system combines an internal CLIP image encoder with a three-layer multilayer perceptron, which converts the image representation into separate scores. Two scores are exposed through the API: likely-csam, for suspected CSAM, and likely-csam/gen-ai, for suspected AI-generated CSAM. They are not mutually exclusive, so an AI-generated image can score highly on both. The card lists default thresholds of 0.996324 for likely-csam and 0.90 for likely-csam/gen-ai.
The training set for the main score held 505,000 positive and 913,000 negative examples, according to the card, with an offline evaluation set of about 7,000 positives and 40,000 negatives. To test false alarms on ordinary imagery, OpenAI ran the classifier over 1,248,985 images from the Re-LAION-5B research dataset.
| Score | Default threshold | Precision (offline) | Recall (offline) | F1 | Flags on 1,248,985 Re-LAION images |
|---|---|---|---|---|---|
| likely-csam | 0.996324 | 0.991 | 0.802 (5,490 of 6,849) | 0.886 | 5 (0.000004) |
| likely-csam/gen-ai | 0.90 | 1.000 | 0.864 (1,966 of 2,275) | 0.927 | 0 |
Source: OpenAI CSAM Image Classifier model card, published November 16, 2026.
The card also sets out where the tool is weaker. At the same thresholds, OpenAI says precision was substantially lower in an online evaluation drawn from ChatGPT traffic than in the offline set. Recall was lowest for images tagged as showing infants: 0.493 for the main score, and for the AI-generated score, two of only five infant-tagged positives were detected. For likely-csam, recall was 61.27% for examples tagged Black, compared with 84.48% to 90.32% for the other reported race tags, and it was higher for female-tagged than male-tagged examples (0.893 against 0.603). According to the card, the classifier has not been tested on video frames, results on OpenAI traffic are not a guarantee of performance on another service, and a score alone does not confirm CSAM or trigger a report. Images submitted to the API are processed without retention and are not used for training or evaluation, the card says.
Results reported by OpenAI and its partners
OpenAI compared the new classifier with its existing protections on user-uploaded images. The post does not state the size of that sample or when the evaluation ran.
| Test | Run by | Reported result |
|---|---|---|
| User-uploaded images, new classifier against existing tool | OpenAI | An estimated 48% fewer images flagged, with "roughly the same number of abusive images" found |
| Same evaluation | OpenAI | Images flagged by the new classifier "about twice as likely to contain abuse" |
| Human-reviewed production images | fal | 72% of high-confidence flags confirmed by reviewers as novel CSAM |
| Flagged cases | invideo AI | One in four flagged cases led to a ban |
| Drawings or animations that sexualize children | SafetyKit | 90% of cases detected |
| Legal adult content | SafetyKit | False positives below 1% |
The first two lines describe a reduction in review work rather than an improvement in detection. Flagging 48% fewer images while finding roughly the same number of abusive ones means fewer false alarms for human reviewers to look at, which is consistent with the second finding that a flagged image is about twice as likely to be abusive.
Samuel Wang, Member of Technical Staff at SafetyKit, went further than the other partners quoted. "In our testing, OpenAI's classifier is the best we've seen, catching unknown CSAM that hash matching misses, including strong performance on drawings and animations, with false positives below 1% on legal adult content," he said. SafetyKit's comparison set is not described in the post. HeyGen and Adobe are also featured among the organizations listed under the results, without test figures of their own. OpenAI says partner feedback "helped us identify weaknesses and improve the system," and that it will keep improving the classifier with feedback from those that use it.
How access works, according to OpenAI
The post and the model card describe the route to the classifier in five steps.
- An organization requests access through OpenAI's CSAM classifier intake form. The model card describes the tool as intended for personnel at appropriately authorized organizations and lists trust and safety teams, child-safety nonprofits and law enforcement among them.
- OpenAI vets the request. Neither document publishes the approval criteria.
- Approved organizations use the classifier through the OpenAI API at no cost.
- Each submitted still image returns two scores, likely-csam and likely-csam/gen-ai, which reviewers use to decide which images need attention first.
- Human reviewers confirm or dismiss flagged images and handle reporting, since a score alone does not confirm CSAM or trigger a report.
ChatGPT: a warning, a deleted conversation and a 24-hour pause
The second part of the release concerns ChatGPT itself. OpenAI says it already prohibits CSAM, trains its models to refuse requests for it and acts against accounts that violate its policies. Those existing guardrails now gain "new product warnings in ChatGPT that interrupt attempts to find abusive material, explain the consequences, and point people toward help to stop harmful behavior."
According to the post, "the warning makes clear that sexualizing children causes harm, including when the requested material is AI-generated." Alongside it, OpenAI will "remove the flagged conversation and pause access to ChatGPT for 24 hours, while keeping specialist support available to help people stop harmful behavior." The post adds: "Repeated violations may lead to longer restrictions and ultimately account closure or reports to authorities."
OpenAI's post shows two versions of the intervention. The first, headed "Let's stop here," carries three points: "Content that sexualizes children is harmful and can be illegal, whether AI-generated or not. Seeking it violates OpenAI's Usage Policies." Then: "We've suspended your account for 24 hours. If this activity continues, we'll permanently close your account." And: "Free, specialized support is available. Others have received help to stop seeking this content, and you can too." Two buttons follow, "I understand" and "Find support." The second, headed "Your account is suspended," reads: "Our systems identified activity that violates OpenAI's Usage Policies." Below that: "We removed the conversation that was flagged. If this activity continues, we'll permanently close your account," and "If you're worried about your thoughts or behavior, free, specialized support is available," with a "Find support" button.
| Stage | What happens, according to OpenAI |
|---|---|
| Attempt detected | Warning shown that sexualizing children causes harm, including AI-generated material |
| Same time | Flagged conversation removed; access to ChatGPT paused for 24 hours |
| During the pause | Specialist support offered through a "Find support" link |
| Repeated violations | Longer restrictions, and ultimately account closure or reports to authorities |
The wording differs slightly between the text and the interface. The post speaks of pausing access to ChatGPT; the screens say the account has been suspended. The post lists reports to authorities among the consequences of repeat violations, while the user-facing messages mention only permanent closure.
OpenAI says it built the approach with organizations and researchers who develop prevention programs: Protect Children, the Lucy Faithfull Foundation, Dr. Caoilte Ó Ciardha, Researcher and Honorary Reader in Forensic Psychology at the University of Kent, and Joel Scanlan at the University of Tasmania, "who independently evaluated a trial of deterrent messages." The post links that phrase to a 2024 University of Tasmania article about a deterrence chatbot.
Scanlan, identified in the quote as Associate Professor at the CSAM Deterrence Centre, described the deployment in favorable terms. "OpenAI has deployed a best-practice example in ChatGPT of how a platform can deter and disrupt harmful online behaviours," he said. He added that OpenAI "aims to redirect users to local support options in as many jurisdictions as possible." The post does not list which jurisdictions those are.
A checklist for developers
The third part is written guidance. "We've published CSAM guidance to give developers practical steps to prevent abuse, detect harmful content, and act on it," OpenAI wrote. The guidance was developed with input from the Tech Coalition, Thorn, NCMEC and the Internet Watch Foundation, and according to the post it covers clear rules, ways for users to report concerns, detection tools, and processes for reviewing and reporting suspected abuse. It "also explains how safety identifiers can help developers connect misuse to individual users and how to support the people reviewing sensitive content."
The guidance page itself organizes its starting checklist in three stages.
- Prevent: set clear rules, make reporting available to users, and track uploads at the user level so that each one can be tied to an account.
- Detect: use available CSAM detection tools, including perceptual hash matching and classifiers for previously unseen material, and identify repeat offenders.
- Respond and report: register and prepare to report CSAM to authorities, keep usable records and identifiers, write a response playbook, train and support staff involved, and use tools designed for moderation and abuse response.
For image generation through the API, the page says developers should pass a stable per-user value in the safety_identifier field, and that the older user parameter is deprecated but still supported. It also states that OpenAI's Moderation API is not a substitute for dedicated CSAM detection, and that its recommendations are not legal advice. The post lists the tools that can be used alongside it: the new classifier, the Moderation API, ROOST's self-hosted review console Coop, and dedicated services such as Thorn's Safer.
A second document comes from competitors working together. OpenAI says it worked "with Anthropic and Google to jointly commission new guidance on responding to disclosures of child sexual exploitation and abuse in AI systems." That guidance was published by the Tech Coalition; the post gives no publication date. OpenAI says its own guidance has already led to "increased numbers of developers accessing child safety tooling and registering to report CSAM to NCMEC," without giving numbers.
Tools and training for investigators
The fourth part is aimed at police. OpenAI says it is "supporting child-safety law enforcement units with new tools and trainings on how to use AI in their investigations, including dedicated plugins, investigative tools, and complex data analysis, with access managed by program administrators." The first tool, according to the post, "will help investigators review lengthy chat logs and legal process returns, identify potential connections, and create timelines, with findings linked to source material."
Training starts with an overview and demonstration, backed by a field guide for later reference. Access "will depend on agency authorization and program requirements," and future versions of the training "will be proposed at regional, national and international conferences." OpenAI links the investigative work to an earlier public conversation with Thorn, held through the OpenAI Forum on February 11, 2026, in which it discussed how AI can help investigators organize information and reduce unnecessary exposure to abusive material.
North Carolina Attorney General Jeff Jackson is quoted in the post, and his statement reads more as a demand than an endorsement. "As state attorneys general on the AI Task Force, our first priority is protecting kids. That needs to be the first priority of tech companies too," he said. "We must increase our defenses, and OpenAI and other tech companies must be thinking about and working around the clock to prevent synthetic child sex abuse material."
What the post does not say
The post gives no sample size, time period or error rates for the internal comparison behind the 48% figure, no approval criteria or count of vetted organizations, and no list of countries or plans where the ChatGPT warning is live. It does not say how many law enforcement units are covered, how many developers registered with NCMEC after the guidance appeared, or whether the classifier is available to organizations in the European Union.
Why this matters for the marketing community
The announcement does not mention advertising. Its subject, however, sits close to several parts of the market that PPC Land covers.
The first link is the tools. Among the companies OpenAI names are fal, invideo AI, HeyGen and Adobe, all of which offer generative image or video tools, a category that agencies and in-house teams use in ad production. Platforms of that kind are exactly where a list of known images is of little use, because the output did not exist until a user asked for it. A free classifier with a dedicated AI-generated signal is an attempt to give those platforms a check that hash lists cannot provide.
The cost of getting this wrong is already on record. Users documented Grok generating sexualized images of minors from December 25, 2025, and on January 26, 2026, the European Commission opened a Digital Services Act investigation into X over sexual deepfake images produced by Grok. Two months later, on March 26, 2026, the Amsterdam District Court barred Grok from generating non-consensual nude images, with penalties of 100,000 euros a day per defendant, capped at 10 million euros per entity. The episode showed how quickly a generative tool's output can turn into a brand suitability question for every advertiser on the same platform.
Europe's legal position on detection is also unsettled. The ePrivacy derogation that allowed providers to scan voluntarily for CSAM expired on April 3, 2026, with no replacement in place, after which Google, Meta, Microsoft and Snap said they would continue voluntary detection. The longer-running fight over the EU's proposed CSAM regulation, known as Chat Control, led Signal in October 2025 to threaten to leave Germany if the proposal passed. In August 2026, OpenAI declared 159.1 million average monthly ChatGPT users in the EU, well above the 45 million mark at which the DSA's strictest risk rules apply. The post is silent on how any of this applies to the classifier or the ChatGPT warnings in Europe.
Ad platforms have been tightening their own rules on the same material. Google's advertising policy on child sexual abuse and exploitation, in force since October 22, 2025, covers AI-generated and fictional depictions, and Google said on October 22, 2026 that it would split the rules into two penalty tiers from January 11, 2027. Apple's App Review Guidelines revisions of October 16, 2026 require apps to prevent and remove nonconsensual intimate imagery, including AI-generated material. Pinterest's Generative AI Acceptable Use Guidelines, which took effect on November 12, 2026, bar the use of its AI tools to compromise the safety of minors, with account suspension or termination among the penalties. OpenAI's 24-hour pause for ChatGPT users follows the same pattern of tying account consequences to how generative features are used.
Then there is ChatGPT's own ad business. When OpenAI announced its first ad tests on January 16, 2026, it said it would exclude ads from accounts where users indicate they are under 18 or where it predicts underage use. By October 21, 2026, ChatGPT ads were available in 89 countries, with more than 100,000 advertisers having run campaigns. Under brand suitability pilots that began on October 5, 2026, DoubleVerify and Integral Ad Science assess how OpenAI applies its standards without access to user conversations. Because verification partners cannot see the chats, OpenAI's own published descriptions of enforcement, like this one, remain the main public record of how prohibited activity is handled on the platform where those ads run.
The political pressure is direct. Attorneys general from 44 jurisdictions wrote to AI companies including OpenAI on August 25, 2025, warning about child safety failures, and the Federal Trade Commission ordered seven AI chatbot companies, OpenAI among them, to report on child safety measures on September 10, 2025. Reporting duties carry legal weight: in August 2025 the Ninth Circuit allowed a negligence claim against X based on the federal duty to report CSAM to NCMEC to proceed. On October 30, 2026, OpenAI said it had made more than 200 law enforcement referrals in the third quarter of 2026. Jackson's quote in today's post addresses OpenAI and "other tech companies" alike.
How much of this holds up outside OpenAI's own traffic is the open question. The headline numbers come from internal tests and a handful of early partners. The model card also publishes the weaker spots: lower recall for infant-tagged and Black-tagged images, no testing on video, and online precision substantially below the offline figures. Those caveats are what any platform, agency or brand relying on the tool would need to weigh.
Key facts
- Announcement: "Strengthening the fight against online child sexual abuse with new tools and support," OpenAI, November 16, 2026, Safety category, author listed as OpenAI
- Classifier name: OpenAI CSAM Image Classifier; model card published November 16, 2026 at https://deploymentsafety.openai.com/csam-image-classifier
- API scores: likely-csam (default threshold 0.996324) and likely-csam/gen-ai (default threshold 0.90), not mutually exclusive
- Architecture: internal CLIP image encoder plus a three-layer multilayer perceptron, according to the model card
- Price and access: no cost, vetted organizations only, through the OpenAI API; requests through https://openai.com/form/csam-classifier-intake/
- Data handling: images processed without retention and not used for training or evaluation, according to the model card
- Not covered: video frames have not been tested, according to the model card
- Reported test results: 48% fewer images flagged for roughly the same abusive images found (OpenAI); 72% of high-confidence flags confirmed as novel CSAM (fal); one in four flagged cases led to a ban (invideo AI); 90% detection on drawings and animations, false positives below 1% on legal adult content (SafetyKit)
- ChatGPT intervention: flagged conversation removed, access paused for 24 hours, specialist support offered; repeated violations can lead to longer restrictions, account closure or reports to authorities
- Developer guidance: "CSAM guidance" at https://developers.openai.com/api/docs/guides/csam-guidance, developed with input from the Tech Coalition, Thorn, NCMEC and the Internet Watch Foundation
- Joint guidance: commissioned by OpenAI, Anthropic and Google and published by the Tech Coalition, at https://technologycoalition.org/resources/responding-to-child-sexual-abuse-disclosures-in-ai-systems/
- Programs joined: Tech Coalition Lantern (signal sharing)
- Law enforcement: first tool reviews chat logs and legal process returns, finds connections and builds timelines linked to source material; access managed by program administrators
- People quoted: John Shehan (NCMEC), Samuel Wang (SafetyKit), Joel Scanlan (CSAM Deterrence Centre), Jeff Jackson (North Carolina Attorney General)
Timeline
- May 11, 2022 - The European Commission first proposes the CSAM regulation known as Chat Control
- August 1, 2025 - The Ninth Circuit lets a negligence claim tied to the NCMEC reporting duty proceed against X
- August 25, 2025 - Attorneys general from 44 jurisdictions write to AI companies, OpenAI among them, over child safety
- September 10, 2025 - The FTC orders seven AI chatbot companies, including OpenAI, to detail child safety measures
- October 3, 2025 - Signal threatens to leave Germany over the Chat Control vote
- October 22, 2025 - Google's final ads policy on child sexual abuse and exploitation takes effect, covering AI-generated depictions
- December 25, 2025 - Users document Grok generating sexualized images of minors
- January 16, 2026 - OpenAI says ChatGPT ads will exclude users under 18 and those it predicts to be minors
- January 26, 2026 - The European Commission opens a DSA investigation into X over Grok
- February 11, 2026 - OpenAI holds a public OpenAI Forum conversation with Thorn on using AI to protect children online
- March 26, 2026 - Amsterdam District Court bars Grok from generating non-consensual nude images
- April 3, 2026 - The EU derogation allowing voluntary CSAM scanning expires without a replacement
- August 2026 - OpenAI declares 159.1 million average monthly EU users for ChatGPT under the DSA
- October 5, 2026 - DoubleVerify and IAS begin brand suitability pilots for ChatGPT ads
- October 16, 2026 - Apple revises its App Review Guidelines on nonconsensual intimate imagery, including AI-generated material
- October 21, 2026 - ChatGPT ads reach 89 countries and more than 100,000 advertisers
- October 22, 2026 - Google announces a two-tier Child Safety ads policy
- October 30, 2026 - OpenAI reports more than 200 law enforcement referrals in Q3 2026
- November 12, 2026 - Pinterest's Generative AI Acceptable Use Guidelines take effect
- November 16, 2026 - OpenAI releases its CSAM image classifier, ChatGPT warnings with 24-hour pauses, developer CSAM guidance and law enforcement tools; the classifier's model card is published
- January 11, 2027 - Google's two-tier Child Safety ads policy takes effect
Related PPC Land coverage
- OpenAI made over 200 law enforcement referrals in Q3 2026 - OpenAI's October 30 safety disclosure, including teen restrictions and a planned Community Safety Report.
- Google lost EU legal cover to scan for child abuse images - The April 2026 lapse of the EU derogation and the scale of Google's hash-matching contribution to NCMEC.
- Grok safety failure: X's latest AI scandal - How Grok came to generate sexualized images of minors in late December 2025.
- Europe confronts X over Grok risks: child safety fears trigger DSA probe - The Commission's formal investigation into sexual deepfakes produced by Grok.
- Dutch court bans Grok from generating non-consensual nude images - The Amsterdam ruling and its daily penalties.
- Google Ads splits child safety rules into two penalty tiers from January 11 - Google's revised advertising rules on child safety.
- Google finalizes child protection advertising policy with broader enforcement scope - The October 2025 policy covering AI-generated and fictional content.
- Apple's App Store rules now require apps to remove AI intimate imagery - Apple's October 2026 guideline changes on intimate imagery and child exploitation.
- Pinterest users who abuse AI tools can lose accounts from November 12 - Pinterest's account penalties for misuse of its generative AI features.
- IAS gains ChatGPT Ads safety pilot without access to user conversations - How brand suitability verification works in ChatGPT without reading chats.
- OpenAI finally pulls trigger on ChatGPT ads after monthslong delay - The January 2026 ad launch and its exclusion of users under 18.
- Advertisers in 26 more countries gain ChatGPT ads, bringing the total to 89 - The scale of ChatGPT advertising by October 2026.
- ChatGPT faces EU risk rules after declaring 159.1 million users - ChatGPT's reported EU audience under the Digital Services Act.
- US Attorneys General target AI companies for child safety failures - The August 2025 letter from 44 attorneys general.
- FTC orders seven AI chatbot companies to detail child safety measures - The federal inquiry into AI chatbots and minors.
- Federal court allows child exploitation case against X to proceed - The Ninth Circuit ruling tied to the duty to report CSAM to NCMEC.
- Signal threatens to exit Germany over Chat Control vote - The dispute over the EU's proposed scanning obligation.
Summary
Who: OpenAI, with NCMEC, the Tech Coalition, ROOST, Thorn, the Internet Watch Foundation, early access partners fal, invideo AI and SafetyKit, prevention researchers and organizations including the Lucy Faithfull Foundation and Protect Children, and child-safety law enforcement units. Anthropic and Google co-commissioned separate guidance with OpenAI.
What: A free image classifier that scores images for suspected CSAM and for suspected AI-generated CSAM, available to vetted organizations through the OpenAI API; ChatGPT warnings that remove the flagged conversation and pause access for 24 hours; published CSAM guidance for developers; and AI tools and training for investigators. OpenAI reports that the classifier flagged 48% fewer images than its existing tool while finding roughly the same number of abusive images.
When: Announced today, November 16, 2026, alongside the classifier's model card. Future law enforcement training is to be proposed at conferences, with no dates given.
Where: The classifier is offered through the OpenAI API, the warnings appear in ChatGPT, and the guidance sits on OpenAI's developer site. The post does not specify which countries the ChatGPT warning covers or whether the classifier is offered in the EU.
Why: OpenAI says child protection online requires coordination across companies, child safety organizations and law enforcement, and that companies have a responsibility to stop their technology from being exploited. Generative AI has made new abusive imagery possible that hash lists of known material cannot catch, while regulators, attorneys general and courts on both sides of the Atlantic have increased pressure on platforms.
Discussion