Updating our Model Spec with teen protections

我们对 Model Spec 进行了更新，公布了针对未满 18 岁用户的 U18 Principles ，这套书面规则阐明了我们期望 AI 模型在尤其是高风险或关键情形下应有的行为、价值取向和准则。模型行为直接影响人们与 AI 的互动，而青少年的发展需求不同于成年人，因此需要专门考虑。

这套 U18 Principles 指导 ChatGPT 为 13 至 17 岁的青少年提供安全、符合年龄的使用体验。原则以发展心理学研究为依据，优先考虑预防、透明与早期干预。制定过程中，我们向外部专家征询意见并做了预览，包括与 American Psychological Association 的沟通，以汲取长期改进的专业建议。

Model Spec 的总体原则仍适用于所有成年与未成年用户；本次更新的重点是明确这些原则在青少年情境下的具体适用方式，特别是在未成年人安全问题更为敏感的场景中应如何更谨慎地执行。

四项核心承诺构成了 U18 Principles 的基石：

• 将青少年安全置于首位，即便这与其他目标发生冲突；
• 鼓励寻求现实世界支持，推动线下关系与可信赖资源的利用；
• 以“青少年”的身份对待他们，既不居高临下，也不把他们当作成年人；
• 通过设定清晰预期保持透明。

与我们此前发布的 Teen Safety Blueprint 一致，这些原则已用于指导目前的青少年安全工作，包括在注册时自报未满 18 岁的用户以及通过家长控制施加的内容保护。在这些场景中，我们已为模型设定额外防护，要求对高风险话题格外谨慎，涵盖自伤与自杀、带有浪漫或性化的角色扮演、血腥或露骨内容、危险活动与毒品、身体形象与饮食失调，以及要求保守有关不安全行为秘密的请求等领域。

经过早期草案评审并提供长期见解的 American Psychological Association 明确强调了保护青少年的重要性。其首席执行官 Dr. Arthur C. Evans Jr 指出，AI 开发者应为青少年用户提供符合其发展阶段的预防措施，并对更年幼的用户采取更保护性的做法。他认为，若 AI 工具与科学证明对社交、心理、行为乃至生理发展至关重要的人际互动相平衡，儿童与青少年可能从中获益；同时，青少年使用 AI 的经历应在可信赖成年人监督与讨论下进行，以促进对 AI 输出的批判性审视，并帮助青少年发展独立思考与技能。

本次更新还明确了当对青少年产生安全担忧时，助理应如何响应。也就是说，当对话进入高风险领域时，青少年将遇到更强的保护措施、被推荐更安全的替代方案，并被鼓励寻求线下的可信支持；若存在迫在眉睫的风险，我们会劝导青少年联系紧急服务或危机干预资源。

与 Model Spec 其他内容一样， U18 Principles 反映的是我们的预期模型行为。随着新研究、专家反馈与实际使用数据的加入，我们会持续完善这些原则。

在强化青少年安全方面的后续工作

除了对 Model Spec 的更新，我们还在 ChatGPT 的整体产品中采取了多层次措施来增强青少年安全，涵盖产品防护、家庭支持与专家指导等方面。

自推出家长控制以来，我们已将保护措施扩展到群聊、新的浏览功能 ChatGPT Atlas 和 Sora 应用中。相关更新让家长在我们推出新产品与新功能时，能够更灵活地为子女定制 ChatGPT 使用体验。

遵循专家建议，我们鼓励家长与青少年在家庭内持续讨论健康且负责任的 AI 使用方式。为支持这种对话，我们在家长资源中心新增了经专家审核的资料，包括名为 Family Guide to Help Teens Use AI Responsibly 的家庭指南及面向家长的交流建议（均经 ConnectSafely 与我们的 Expert Council on Well-Being and AI 成员审阅）。我们将持续补充更多资源。同时，产品端也内置了长时间会话的休息提醒，帮助用户有意识地平衡与 ChatGPT 的互动时间。

与专家的协作

我们的青少年安全工作依赖跨学科专家的密切参与。今年十月，我们成立了 Expert Council on Well-Being and AI ，为何为所有年龄段设计健康的 AI 互动提供建议，这些意见已被用于家长控制与家长通知的制定。我们还通过 Global Physician Network 引入临床专业知识，用以支持安全性研究并评估模型行为，包括提升 ChatGPT 识别情绪困扰并在适当时引导用户寻求专业护理的能力。我们以此为基础构建了 GPT‑5.2 ，并通过与 ThroughLine 的合作，在 ChatGPT 和 Sora 中呈现本地化求助热线，扩大了可获得的线下支持渠道。

下一步

我们正处于在 ChatGPT 消费版计划中推出年龄预测模型的早期阶段，这将帮助我们在判断某账户可能属于未成年人时自动应用青少年保护措施。如果对年龄判断不够确定或信息不完整，我们将默认提供 U18 experience ，并为成年人提供验证年龄的途径。

强化青少年安全是持续的工作。我们会不断改进家长控制与模型能力，扩充家长资源，并与组织、研究者及包括 Well-Being Council 与 Global Physician Network 在内的专家伙伴合作。

我们承诺建立并持续改进有力的青少年保护措施，以更好地支持青少年及其家庭。

We’re sharing an update to our Model Spec, the written set of rules, values, and behavioral expectations that guides how we want our AI models to behave, especially in difficult or high stakes situations, with Under-18 (U18) Principles⁠. Model behavior is critical to how people interact with AI, and teens have different developmental needs than adults.

The U18 Principles guide how ChatGPT should provide a safe, age-appropriate experience for teens aged 13 to 17. Grounded in developmental science, this approach prioritizes prevention, transparency, and early intervention. In developing these principles, we previewed them with external experts, including the American Psychological Association, as part of our ongoing work to seek input to strengthen our approach.

While the principles of the Model Spec continue to apply to both adult and teen users, this update clarifies how it should be applied in teen contexts, especially where safety considerations for minors may be more pronounced. 

The U18 Principles are anchored in four guiding commitments:

• Put teen safety first, even when it may conflict with other goals
• Promote real-world support by encouraging offline relationships and trusted resources
• Treat teens like teens, neither condescending to them nor treating them as adults
• Be transparent by setting clear expectations

Consistent with our Teen Safety Blueprint⁠, these principles have guided our teen safety work to date, including the content protections we apply to users who tell us they are under 18 at sign up, and through parental controls. In these contexts, we’ve implemented safeguards to guide the model to take extra care when discussing higher-risk areas, including self-harm and suicide, romantic or sexualized roleplay, graphic or explicit content, dangerous activities and substances, body image and disordered eating, and requests to keep secrets about unsafe behavior.

The American Psychological Association, which reviewed an early draft of the U18 Model Spec and offered important insights for the long term, is clear about the importance of protecting teens:

“APA encourages AI developers to offer developmentally appropriate precautions for youth users of their products and to take a more protective approach for younger users.  Children and adolescents might benefit from AI tools if they are balanced with human interactions that science shows are critical for social, psychological, behavioral, and even biological development. Youth experiences with AI should be thoroughly supervised and discussed with trusted adults to encourage critical review of what AI bots offer, and to encourage young people’s development of independent thinking and skills.”—Dr. Arthur C. Evans Jr, CEO, American Psychological Association

This update also clarifies how the assistant should respond when safety concerns arise for teens. This means teens should encounter stronger guardrails, safer alternatives, and encouragement to seek trusted offline support when conversations move into higher-risk territory. Where there is imminent risk, teens are urged to contact emergency services or crisis resources.

As with the rest of the Model Spec, the U18 Principles reflect our intended model behavior. We will continue to refine them as we incorporate new research, expert input, and real-world use.

Building on our work to strengthen teen safety

Alongside updating the Model Spec, we’ve taken a multi-layered approach to strengthening teen safety⁠ across ChatGPT, spanning product safeguards, family support, and expert guidance.

Since rolling out parental controls⁠, we’ve extended protections across new products including group chats, the ChatGPT Atlas browser, and the Sora app. These updates help parents tailor their teen’s ChatGPT experience as we introduce new products and features.

Consistent with expert guidance, we encourage ongoing conversations between parents and teens about healthy and responsible AI use in their family. To support these conversations, we’ve added new expert-vetted resources to the parents resource hub⁠, including a Family Guide to Help Teens Use AI Responsibly⁠ and tips for parents⁠ on how to talk with their kids about AI, that were reviewed by ConnectSafely and members of our Expert Council on Well-Being and AI⁠. We’ll continue adding more resources over time. We also support healthy use directly in the product, with built-in break reminders⁠ during long sessions to help keep time spent with ChatGPT intentional and balanced.

Working with experts

Our work in teen safety is guided by close engagement with experts across disciplines and expertise. In October, we established an Expert Council on Well-Being and AI⁠ to help advise and define what healthy interactions with AI should look like for all ages. That work has informed guidance on parental controls and parent notifications. We also incorporate clinical expertise through our Global Physician Network⁠ to inform safety research and evaluate model behavior, including improving how ChatGPT recognizes distress and guides people toward professional care when appropriate. We built on these foundations with GPT‑5.2⁠, and we’ve also expanded access to real-world support by surfacing localized helplines in ChatGPT and Sora through our partnership with ThroughLine⁠.

What’s next

We’re in the early stages of rolling out an age prediction model⁠ on ChatGPT consumer plans. This will help us automatically apply teen safeguards when we believe an account belongs to a minor. If we are not confident about someone’s age or have incomplete information, we’ll default to an U18 experience and give adults ways to verify their age.

Strengthening teen safety is ongoing work and we’ll continue to improve parental controls and model capabilities, expand resources for parents, work with organizations, researchers, and expert partners including the Well-Being Council and Global Physician Network. 

We’re committed to making strong teen protections and improving them over time to better support teens and families.

Generated by RSStT. The copyright belongs to the original author.

Source