OpenAI has released a set of new open-source safety prompts designed for developers to enhance the implementation of policies aimed at safeguarding teenagers. This prompt-based safety toolkit offers guidance on typical teenage risks, suggestions for developmental content, and age-appropriate advice on subjects such as self-harm, sexual themes, romantic role-playing, perilous trends or viral challenges, and detrimental body ideals. OpenAI characterizes this as a more comprehensive option compared to earlier high-level guidelines, structured as prompts that seamlessly integrate into AI systems.
In December, OpenAI incorporated fresh Under-18 principles into its Model Spec. Previously, the company unveiled gpt-oss-safeguard, an open-weight reasoning model intended to assist developers in applying safety conditions and categorizing content as safe or unsafe. Unlike conventional safety classification methods, gpt-oss-safeguard can directly embed platform safety policies and deduce the intent of those policies to differentiate appropriate outputs. “Even seasoned teams frequently encounter difficulties transmuting high-level safety objectives into specific, actionable rules, particularly as it necessitates both subject matter expertise and extensive AI knowledge,” OpenAI noted in its most recent press announcement. “This can result in lapses in protection, uneven enforcement, or excessively broad filtering. Well-defined, clearly scoped policies are an essential cornerstone for effective safety frameworks.”
The additional developer toolkit was created in partnership with the nonprofit Common Sense Media and everyone.ai. Experts have warned parents about the risks of excessive exposure to chatbots for at-risk teenagers and even young children, as AI firms strive to address the effects of their models on user mental health. Last year, OpenAI was the target of a lawsuit from the parents of teenager Adam Raine in the industry’s inaugural wrongful death case, with the Raine family claiming that ChatGPT’s sycophantic nature and insufficient safety policies contributed to their son’s demise. The company has refuted accusations of wrongdoing and has improved its mental health and teenager safety functionalities, including age verification measures. However, third-party developers utilizing OpenAI’s models have found it challenging to uphold the same standards of safety precautions, particularly in AI-driven children’s toys.
The lawsuit against OpenAI followed several legal actions against the contentious platform Character.AI and laid the groundwork for a recent wrongful death lawsuit filed against OpenAI’s rival Google and its Gemini AI assistant. Across the industry, tech and social media companies are encountering numerous legal obstacles regarding the long-term effects of their products on users. Last month, Instagram CEO Adam Mosseri and Meta leader Mark Zuckerberg provided testimony before a jury in a significant case that scrutinizes social media platforms for their purportedly addictive design strategies. A verdict is still pending.
OpenAI stressed that its new safety prompt toolkit is not a “comprehensive or final definition or guarantee of teen safety.” Robbie Torney, who leads AI and digital assessments at Common Sense Media, remarked that the new policies could create a “meaningful safety floor across the ecosystem,” addressing an AI safety deficiency worsened by the absence of operational guidelines for developers.
Developers can acquire OpenAI’s safety model on Hugging Face and find its new prompt pack on GitHub.
Disclosure: Ziff Davis, the parent company of Mashable, initiated a lawsuit against OpenAI in April, alleging that it violated Ziff Davis copyrights while training and operating its AI systems.