OpenAI deploys age projection models to prepare adult content.

OpenAI recently announced the start of the deployment of an age projection model to determine whether ChatGPT users are 18 years of age or older and have access to “sensitive or potentially harmful content”. This has resulted from litigation and congressional hearings by its chat robots linked to multiple suicides, forcing AI companies to transform service security from propaganda slogans to practical action.

At the September 2025 Senate hearing, the Director of Strategic Integration of the American Psychological Association, Mitch Prinstein, gave written testimony that over half of the United States youths over 13 years of age were using the Generating AI and the proportion of users under 13 years of age was estimated at 10-20 per cent. He stressed that “the AI system designed for adults is fundamentally inappropriate for adolescents and requires the development of special protection measures for the stage of development.” OpenAI’s solution is to develop an automated age projection system. It explained that the model was judged through a “behaviour-to-account signal mix”, including information on the length of the account’s life, typical active periods, long-term use patterns and age statements by users. When suspected underage users are detected, the system automatically enhances the security set to reduce the blood content of violence, the viral challenge of inducing harmful behaviour, sexual/love/violent role-playing, self-mutilation and promotion of extreme aesthetic standards. However, the system has significant limitations. OpenAI is frank in the document: “No system is perfect. If we misjudge your age, a user over 18 years of age can file a complaint through a third-party certification body, Persona, for a photograph of a real-time self-shot or a government identity document.”

However, Alexis Hancock, Director General of Engineering of the Electronic Front Foundation of the United States, said in an interview with The Register: “The model itself is not obliged to be correct and its decisions cannot be challenged.” She warned that factors such as the length of use of the account were questionable in the context of ChatGPT’s introduction for only four years. OpenAI technical path echoes the findings of the Australian age certification test. Australia’s Age Assurance Technical Test Report shows that, despite the challenges, the average age-validation accuracy rate is as high as 97.05 per cent, but it is lower for “ages, non-white users and women near policy thresholds”. Last June, the Australian Broadcasting Corporation reported that the age certification system had managed to contain errors within 18 months in only 85 per cent of cases. Advocates are sceptical about technological maturity. Mozilla, a non-profit organization, stated last month: “Despite the existence of a variety of techniques to validate, estimate or extrapolate age, the fundamental contradiction between validity, accessibility, privacy and security remains unresolved”. The Association of Computers and Communications Industries (on behalf of giants such as Amazon, Apple, Google) also warned in October last year that age-validation techniques “is difficult to implement in practice”.

With ChatGPT ‘ s introduction of advertising services and the development of pornographic content, OpenAI must find a balance between the protection of minors and commercialization. The company predicts that the model will be fully implemented in the EU in a few weeks’ time, but that it will use the “implementation age threshold” Instead of “accurate validation” being at the core of the design approach, it is triggering a double examination by global regulators and privacy protectors.