OpenAI publishes plan to protect children from AI-enabled sexual exploitation

OpenAI has published a new document called the Child Safety Blueprint, in which the company sets out its approach to preventing generative AI from being used to create and distribute child sexual exploitation material. The document arrives amid growing pressure from regulators, child-protection organizations and the public.

Generative AI has pushed the problem to a fundamentally new scale. Where producing CSAM (child sexual abuse material) once required contact with real victims, offenders can now create realistic synthetic images and video without physically coming into contact with any child. According to the National Center for Missing & Exploited Children (NCMEC), the number of reports linked to AI-generated content of this kind has grown many times over in the past two years.

OpenAI describes four key areas of work. First, prevention at the model level: its systems are trained to refuse requests that could lead to harmful content, and its models are tested for resistance to "jailbreaks," workaround techniques that allow users to bypass restrictions. Second, response and reporting: the company uses hash databases from NCMEC and the Technology Coalition to automatically block the transfer of known material, and it regularly files reports with law enforcement. Third, partnerships with organizations that work on child protection in practice. Fourth, research into new types of attacks and vulnerabilities as the technology develops.

Synthetic victims pose a particular difficulty: children who do not exist in reality, but whose realistic images are generated by AI. Legal regulation remains uneven: in some jurisdictions synthetic CSAM is treated as fully equivalent to material depicting real victims, while in others it sits in a legal gray area. OpenAI's position is unambiguous: all such content should be classified as CSAM regardless of how it was created.

The industry's reaction is mixed. Some experts see the document as an important step toward transparency and standardization. Other critics point out that the gap between stated commitments and the real effectiveness of protective systems may be significant: several research groups have already recorded cases where models were bypassed through multi-step conversations. How resilient OpenAI's defenses are to such attacks will be shown in practice.

In parallel, the industry is discussing common standards. The Technology Coalition, which brings together Amazon, Apple, Google, Meta*, Microsoft and a number of other companies, is developing a universal safety framework for generative AI developers. The pace of industry consolidation, however, still lags behind the pace at which new threats emerge.

The publication of the child safety plan signals that protecting minors is becoming a mandatory part of AI companies' corporate agenda, on par with data privacy and countering disinformation. Companies that have not developed their own standards, whether voluntarily or under regulatory pressure, risk facing legislative restrictions and serious reputational damage.

*Meta has been designated an extremist organization and is banned in Russia.