Building more helpful ChatGPT experiences for everyone
OpenAI News我们致力于让 ChatGPT 尽可能地提供帮助，这项工作是持续不断的。我们看到人们在最艰难的时刻依赖它[https://openai.com/index/helping-people-when-they-need-it-most/]。因此，我们不断改进模型识别和应对心理及情绪困扰的能力，并以专家意见为指导。
这项工作已经在进行中，但我们希望主动预览未来120天的计划，让大家无需等待新功能发布就能了解我们的方向。工作将远超这段时间，但我们正集中精力争取今年推出尽可能多的改进。
上周，我们分享了在关键时刻帮助人们的四个重点领域[https://openai.com/index/helping-people-when-they-need-it-most/]：
- 将干预措施扩展到更多处于危机中的人群
- 让联系紧急服务和专家帮助变得更容易
- 促进与可信联系人建立联系
- 加强对青少年的保护
部分工作进展迅速，另一些则需要更多时间。
今天，我们将分享更多关于如何与专家合作指导工作、利用推理模型应对敏感时刻，以及重点领域之一——加强对青少年的保护的细节。
——
与专家合作
人工智能是新兴且不断发展的领域，我们希望确保进展由深厚的心理健康和福祉专业知识指导。我们的“福祉与人工智能专家委员会”和“全球医生网络”共同提供了专业医疗知识的深度和视角的广度，帮助指导我们的做法。我们将在120天计划期间分享更多相关工作。
福祉与人工智能专家委员会
今年早些时候，我们开始召集一支由青少年发展、心理健康和人机交互领域专家组成的委员会。该委员会的职责是制定一个清晰、基于证据的愿景，指导人工智能如何支持人们的福祉并助其成长。
他们的意见将帮助我们定义和衡量福祉、设定优先事项、设计未来的保护措施（如家长控制的后续版本），并结合最新研究。委员会将为我们的产品、研究和政策决策提供建议，但最终决策权仍由 OpenAI 承担。
全球医生网络
该委员会将与我们的全球医生网络协同工作。该网络涵盖了来自60个国家的250多名医生，过去一年我们与他们合作开展了健康基准评估[https://openai.com/index/healthbench/]等项目，旨在更好地衡量 AI 系统在健康领域的能力。
在这批医生中，已有来自30个国家的90多名精神科医生、儿科医生和全科医生参与了我们关于模型在心理健康场景中应如何表现的研究。他们的意见直接影响我们的安全研究、模型训练和其他干预措施，帮助我们在需要时迅速联系合适的专家。
我们还在不断扩充网络，新增在饮食失调、物质使用和青少年健康等领域拥有深厚专业知识的临床医生和研究人员。
——
利用推理模型应对敏感时刻
我们的推理模型（如 GPT-5-thinking 和 o3）设计为花更多时间思考和推理上下文后再回答。通过一种称为“深思熟虑对齐”的训练方法[https://openai.com/index/deliberative-alignment/]，测试显示(opens in a new window) [https://cdn.openai.com/gpt-5-system-card.pdf]，推理模型更能持续遵守安全指南，并且对对抗性提示更具抵抗力[https://openai.com/index/trading-inference-time-compute-for-adversarial-robustness/]。
我们最近推出了一个实时路由器[https://openai.com/index/introducing-gpt-5/]，可根据对话上下文在高效聊天模型和推理模型之间切换。我们将很快开始将一些敏感对话（如系统检测到急性困扰迹象时）路由到推理模型（如 GPT-5-thinking），以提供更有帮助和益处的回应，无论用户最初选择了哪个模型。我们会谨慎迭代这一方法。
——
加强对青少年的保护
许多年轻人已经在使用 AI。他们是首批“AI 原住民”，从小就将这些工具视为日常生活的一部分，就像早期几代人对互联网或智能手机的使用一样。这带来了支持、学习和创造的真实机会，但也意味着家庭和青少年可能需要帮助，制定适合青少年独特发展阶段的健康使用指南。
家长控制
今年早些时候，我们开始构建更多家庭共用 ChatGPT 的方式，让他们决定最适合自家情况的设置。未来一个月内，家长将能够：
- 通过简单的电子邮件邀请，将自己的账户与青少年的账户（最低年龄13岁）关联
- 通过默认开启的年龄适宜模型行为规则，控制 ChatGPT 对青少年的回应方式
- 管理禁用的功能，包括记忆和聊天记录
- 在系统检测到青少年处于急性困扰时收到通知。该功能将由专家意见指导，以支持家长与青少年之间的信任
这些控制措施是我们为所有用户推出的功能的补充，包括长时间使用时的应用内休息提醒。
这些只是开始。我们将继续学习并加强方法，以专家为指导，目标是让 ChatGPT 尽可能地提供帮助。期待在未来120天内分享我们的进展。
Our work to make ChatGPT as helpful as possible is constant and ongoing. We’ve seen people turn to it in the most difficult of moments. That’s why we continue to improve how our models recognize and respond to signs of mental and emotional distress, guided by expert input.
This work has already been underway, but we want to proactively preview our plans for the next 120 days, so you won’t need to wait for launches to see where we’re headed. The work will continue well beyond this period of time, but we’re making a focused effort to launch as many of these improvements as possible this year.
Last week, we shared four focus areas when it comes to helping people when they need it most:
- Expanding interventions to more people in crisis
- Making it even easier to reach emergency services and get help from experts
- Enabling connections to trusted contacts
- Strengthening protections for teens.
Some of this work will move very quickly, while other parts will take more time.
Today, we’re sharing more on how we’re partnering with experts to guide our work, leveraging our reasoning models for sensitive moments, as well as details on one of our focus areas: Strengthening protections for teens.
Partnering with experts
AI is new and evolving, and we want to make sure our progress is guided by deep expertise on well-being and mental health. Together, our Expert Council on Well-Being and AI and our Global Physician Network provide both the depth of specialized medical expertise and the breadth of perspective needed to inform our approach. We’ll share more about these efforts during our 120-day initiative.
Expert Council on Well-Being and AI
Earlier this year, we began convening a council of experts in youth development, mental health, and human-computer interaction. The council’s role is to shape a clear, evidence-based vision for how AI can support people’s well-being and help them thrive.
Their input will help us define and measure well-being, set priorities, and design future safeguards—such as future iterations of parental controls—with the latest research in mind. While the council will advise on our product, research, and policy decisions, OpenAI remains accountable for the choices we make.
Global Physician Network
This council will work in tandem with our Global Physician Network—a broader pool of more than 250 physicians who have practiced in 60 countries—that we have worked with over the past year on efforts like our health bench evaluations, which are designed to better measure capabilities of AI systems for health.
Of this broader pool, more than 90 physicians across 30 countries—including psychiatrists, pediatricians, and general practitioners—have already contributed to our research on how our models should behave in mental health contexts. Their input directly informs our safety research, model training, and other interventions, helping us to quickly engage the right specialists when needed.
We are adding even more clinicians and researchers to our network, including those with deep expertise in areas like eating disorders, substance use, and adolescent health.
Leveraging reasoning models for sensitive moments
Our reasoning models—like GPT‑5-thinking and o3—are built to spend more time thinking for longer and reasoning through context before answering. Trained with a method we call deliberative alignment, our testing shows that reasoning models more consistently follow and apply safety guidelines and are more resistant to adversarial prompts.
We recently introduced a real-time router that can choose between efficient chat models and reasoning models based on the conversation context. We’ll soon begin to route some sensitive conversations—like when our system detects signs of acute distress—to a reasoning model, like GPT‑5-thinking, so it can provide more helpful and beneficial responses, regardless of which model a person first selected. We’ll iterate on this approach thoughtfully.
Strengthening protections for teens
Many young people are already using AI. They are among the first “AI natives,” growing up with these tools as part of daily life, much like earlier generations did with the internet or smartphones. That creates real opportunities for support, learning, and creativity, but it also means families and teens may need support in setting healthy guidelines that fit a teen’s unique stage of development.
Parental Controls
Earlier this year, we began building more ways for families to use ChatGPT together and decide what works best in their home. Within the next month, parents will be able to:
- Link their account with their teen’s account (minimum age of 13) through a simple email invitation.
- Control how ChatGPT responds to their teen with age-appropriate model behavior rules, which are on by default.
- Manage which features to disable, including memory and chat history.
- Receive notifications when the system detects their teen is in a moment of acute distress. Expert input will guide this feature to support trust between parents and teens.
These controls add to features we have rolled out for all users including in-app reminders during long sessions to encourage breaks.
These steps are only the beginning. We will continue learning and strengthening our approach, guided by experts, with the goal of making ChatGPT as helpful as possible. We look forward to sharing our progress over the coming 120 days.
Generated by RSStT. The copyright belongs to the original author.