Our approach to age prediction

我们正在为 ChatGPT 消费者计划推出年龄预测功能，以判断一个账号是否可能属于 18 岁以下的未成年人，从而为青少年提供更合适的使用体验和必要的保护措施。正如我们在 Teen Safety Blueprint 和 Under-18 Principles for Model Behavior 中所阐明的，年轻人既应享有技术带来的机会，也应得到对其身心健康的保护。

年龄预测是在现有保护措施基础上的补充。注册时明确表示未满 18 周岁的青少年，会自动获得额外的保护设置，减少接触敏感或可能有害内容的几率；同时，这也让我们能在确保安全的前提下，将成年人按成年人对待并满足他们的使用需求。

我们此前已公布了年龄预测的早期计划，现在在逐步部署的同时，提供更多细节说明。

年龄预测如何实现

ChatGPT 使用一套年龄预测模型来估计某个账号是否可能属于 18 岁以下的人。该模型综合考量行为和账号层面的信号，包括账号存在时长、活跃的典型时段、随时间变化的使用模式以及用户声明的年龄。部署后我们会观察哪些信号有助于提升准确性，并以此不断优化模型。

被错误归入未成年体验的用户，始终可以通过简便的方式迅速验证年龄并恢复全部访问权限——即通过 Persona （一项安全的身份验证服务）提交自拍完成验证。用户可随时前往 Settings > Account 查看账号是否已启用额外保护并发起验证流程。

当年龄预测模型判断某账号可能属于未成年人时， ChatGPT 会自动启用一系列附加保护，旨在减少接触敏感内容，例如：

• 露骨的暴力或血腥内容
• 可能鼓励未成年人从事危险或有害行为的病毒挑战
• 性、恋爱或暴力角色扮演
• 自伤描绘
• 倡导极端审美、不健康节食或羞辱体形的内容

这一做法有专家意见为指导，并基于儿童发展科学的学术研究，考虑到青少年在风险感知、冲动控制、同伴影响与情绪调节等方面的已知差异。尽管这些内容限制能降低青少年接触敏感材料的可能性，我们仍致力于持续完善保护措施，特别是应对规避防护的尝试。在对用户年龄没有把握或信息不完整时，我们采取更保守的默认体验。

除了上述保护外，家长还可以通过家长控制进一步自定义青少年的使用体验，包括设置禁止使用的静默时段、控制诸如记忆或模型训练等功能，以及在检测到急性困扰迹象时接收通知。

下一步

我们正从初期部署中学习，并会随着时间推进持续提升年龄预测的准确性。我们将密切监测上线情况，并据此改进。

在 EU 区域，年龄预测功能将于未来数周内上线，以符合当地要求。详见帮助页面。

这虽然是重要一步，但我们在保障青少年安全方面的工作仍在继续。我们会在与专家组织（包括 American Psychological Association 、 ConnectSafely 和 Global Physicians Network ）对话的基础上，持续公布进展与所获经验。

We're rolling out age prediction on ChatGPT consumer plans to help determine whether an account likely belongs to someone under 18, so the right experience and safeguards can be applied to teens. As we’ve outlined in our Teen Safety Blueprint⁠ and Under-18 Principles for Model Behavior⁠, young people deserve technology that both expands opportunity and protects their well-being.

Age prediction builds on protections already in place. Teens who tell us they are under 18 when they sign up automatically receive additional safeguards to reduce exposure to sensitive or potentially harmful content. This also enables us to treat adults like adults and use our tools in the way that they want, within the bounds of safety.

We previously shared our early plans⁠ for age prediction, and today we’re sharing more detail as the rollout is underway.

How age prediction works

ChatGPT uses an age prediction model to help estimate whether an account likely belongs to someone under 18. The model looks at a combination of behavioral and account-level signals, including how long an account has existed, typical times of day when someone is active, usage patterns over time, and a user’s stated age. Deploying age prediction helps us learn which signals improve accuracy, and we use those learnings to continuously refine the model over time.

Users who are incorrectly placed in the under-18 experience will always have a fast, simple way to confirm their age and restore their full access with a selfie through Persona, a secure identity-verification service. Users can check if safeguards have been added to their account and start this process at any time by going to Settings > Account.

When the age prediction model estimates that an account may belong to someone under 18, ChatGPT automatically applies additional protections⁠ designed to reduce exposure to sensitive content, such as:

• Graphic violence or gory content
• Viral challenges that could encourage risky or harmful behavior in minors
• Sexual, romantic, or violent role play
• Depictions of self-harm
• Content that promotes extreme beauty standards, unhealthy dieting, or body shaming

This approach is guided by expert input and rooted in academic literature about the science of child development and acknowledges known teen differences in risk perception, impulse control, peer influence, and emotional regulation. While these content restrictions help reduce teens’ exposure to sensitive material, we are focused on continually improving these protections, especially to address attempts to bypass our safeguards. When we are not confident about someone’s age or have incomplete information, we default to a safer experience.

In addition to these safeguards, parents can choose to customize their teen’s experience further through parental controls⁠ including setting quiet hours when ChatGPT can not be used, controlling features such as memory or model training, and receiving notifications if signs of acute distress are detected.

What’s next

We’re learning from the initial rollout and continuing to improve the accuracy of age prediction over time. We will closely track rollout and use those signals to guide ongoing improvements.

In the EU, age prediction will roll out in the coming weeks to account for regional requirements. For more detail, visit our help page⁠.

While this is an important milestone, our work to support teen safety is ongoing. We’ll continue to share updates on our progress and what we’re learning, in dialogue with experts including the American Psychological Association, ConnectSafely, and Global Physicians Network⁠.

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

Source