The Sora feed philosophy

原则

我们为 Sora 信息流设定的目标很简单：帮助人们了解可能性，并激发他们去创造。为实现这一愿景，我们制定了以下核心出发原则：

• 优先促进创造力。我们的推荐排序旨在偏向创造性和积极参与，而不是被动刷屏。我们认为这才是让 Sora 使用起来愉悦的关键。
• 让用户掌控体验。信息流提供可引导的排序，你可以明确告诉算法自己想看什么。家长也可以通过 ChatGPT 的家长控制，为未成年子女关闭信息流个性化和连续滚动功能。
• 强化连接性。我们希望 Sora 帮助人们加强并建立新的联系，尤其是通过有趣、充满魔力的 Cameo 互动流程。与他人关联的内容会优先于全球性、无关联的内容。
• 在安全与自由间找到平衡。信息流旨在广泛可访问且安全。强有力的护栏可在生成阶段就阻止不安全或有害内容，同时我们会主动屏蔽可能违反使用政策的内容。与此并行，我们也希望保留表达、创意与社区空间。我们知道推荐系统是不断演进的：随着真实使用情况的反馈，我们会在这些原则指导下调整细节。

运作方式

我们的推荐算法旨在为你提供个性化的推荐，激发你和他人的创作灵感。每个人的兴趣和口味不同，所以我们构建了个性化系统来更好地满足这一目标。

为了个性化你的 Sora 信息流，我们可能会考虑以下信号：

• 你在 Sora 上的活动：可能包括你的发帖、关注的账号、点赞和评论的帖子以及二次创作（remix）的内容。也可能包括你的设备访问 Sora 时的一般位置（例如城市），基于诸如 IP 地址等信息。
• 你的 ChatGPT 数据：我们可能会考虑你的 ChatGPT 历史，但你可以随时在 Sora 的数据控制（设置中）关闭此项。
• 内容互动信号：可能包括浏览量、点赞、评论和二次创作数。
• 作者信号：可能包括粉丝数量、其它发帖记录以及历史帖子的互动情况。
• 安全信号：该帖是否被判断为违规或合适。

我们可能会使用这些信号来预测某条内容是否是你可能想看到并以此进行创作的素材。

家长也可以通过 ChatGPT 的家长控制，为他们的青少年关闭信息流个性化并管理连续滚动功能。

我们如何平衡安全与表达

让 Sora 信息流对每个人都既安全又有趣需要谨慎把握：既要保护用户免受有害内容，又要为创造力留下足够空间。

第一道防线在创作环节。由于每条帖子都在 Sora 内生成，我们可以在源头构建强有力的护栏，在内容产生前就阻止不安全或有害内容。这包括对性内容、涉及真实人物的血腥暴力、极端主义宣传、仇恨内容以及鼓励自残或饮食失调的内容的限制。

在生成之外，信息流被设计为适合所有 Sora 用户（包括青少年），因此我们会过滤可能有害、不安全或不适龄的内容。我们优先处理可能造成最大伤害的内容，包括血腥自残、性或暴力的露骨内容；不健康的节食或运动行为；基于外貌的批评或比较与欺凌内容；未成年人可能模仿的危险挑战；极端冒犯性语言或美化仇恨/颓废或鼓励暴力的内容；以及促销年龄限制商品或活动（包括非法药物或有害物质）的内容。

我们使用自动化工具扫描所有信息流内容，以检测是否符合我们的全球使用政策（Global Usage Policies）和信息流资格要求。这些系统会随着我们对新风险的了解不断更新。如果你看到你认为不符合我们使用政策的内容，可以举报。

我们以人工审查作为补充。团队会监控用户举报并主动检查信息流活动，以捕捉自动化可能遗漏的问题。如果你看到你认为不符合我们使用政策的内容，也可以提交举报。

但安全不仅仅是严格过滤。过多限制会扼杀创造力，而过度自由又会破坏信任。我们的目标是取得平衡：在风险最高的领域采用前瞻性护栏，同时配合“举报 + 下架”机制，在给予用户探索与创作空间的同时，确保我们能在问题出现时迅速行动。在 ChatGPT 的 4o 图像生成模型中，这一方法已被证明有效，我们将在此基础上继续构建。

我们也意识到一开始不可能做到十全十美。推荐系统和安全模型是不断演进的系统，你的反馈对我们完善它们至关重要。我们期待与大家一起学习、改进并随着时间不断提升。

Principles

Our aim with the Sora feed is simple: help people learn what’s possible, and inspire them to create. Here are some of core starting principles to bring this vision to life:

• Optimize for creativity. We’re designing ranking to favor creativity and active participation, not passive scrolling. We think this is what makes Sora joyful to use.
• Put users in control. The feed ships with steerable ranking, so you can tell the algorithm exactly what you’re in the mood for. Parents can also turn off feed personalization and continuous scroll for their teens through ChatGPT parental controls.
• Prioritize connection. We want Sora to help people strengthen and form new connections, especially through fun, magical Cameo flows. Connected content will be favored over global, unconnected content.
• Balance safety and freedom. The feed is designed to be widely accessible and safe. Robust guardrails prevent unsafe or harmful generations from the start and we proactively block content that may violate our Usage Policies. At the same time, we also want to leave room for expression, creativity, and community. We know recommendation systems are living, breathing things. As we learn from real use, we’ll adjust the details—in service of these principles.

How it works

Our recommendation algorithms are designed to give you personalized recommendations that inspire you and others to be creative. Each individual has unique interests and tastes so we’ve built a personalized system to best serve this mission.

To personalize your Sora feed, we may consider signals like:

• Your activity on Sora: This may include your posts, followed accounts, liked and commented posts, and remixed content. It may also include the general location (such as the city) from which your device accesses Sora, based on information like your IP address.
• Your ChatGPT data: We may consider your ChatGPT history, but you can always turn this off in Sora’s Data Controls, within Settings.
• Content engagement signals: This may include views, likes, comments, and remixes.
• Author signals: This may include follower count, other posts, and past post engagement.
• Safety signals: Whether or not the post is considered violative or appropriate.

We may use these signals to predict if this content is something you may like to see and riff off of.

Parents are also able to turn off feed personalization and manage continuous scroll for their teens using parental controls in ChatGPT.

How we balance safety & expression

Keeping the Sora feed safe and fun for everyone means walking a careful line: protect users from harmful content, while leaving enough freedom for creativity to thrive.

Our first layer of defense is at the point of creation. Because every post is generated within Sora, we can build in strong guardrails that prevent unsafe or harmful content before it’s made. This includes restrictions on sexual content, graphic violence involving real people, extremist propaganda, hate content, and content that promotes self-harm or disordered eating.

Beyond generation, the feed is designed to be suitable for all Sora users, including teens—so we filter out content that may be harmful, unsafe or age-inappropriate. We prioritize content that could cause the most harm, including graphic self-harm, sexual, or violent content; unhealthy dieting or exercise behaviors; appearance-based critiques or comparisons and bullying content; dangerous challenges likely to be imitated by minors; extremely offensive language or content glorifying hatred or depression or promoting violence; or promotion of age-restricted goods or activities including illegal drugs or harmful substances.

We use automated tools to scan all feed content for compliance with our Global Usage Policies and feed eligibility. These systems are continuously updated as we learn more about new risks. If you see something you think does not follow our Usage policies, you can report it.

We complement this with human review. Our team monitors user reports and proactively checks feed activity to catch what automation may miss. If you see something you think does not follow our Usage Policies, you can report it.

But safety isn’t only about strict filters. Too many restrictions can stifle creativity, while too much freedom can undermine trust. We aim for a balance: proactive guardrails where the risks are highest, combined with a reactive “report + takedown” system that gives users room to explore and create while ensuring we can act quickly when problems arise. This approach has served us well in ChatGPT’s 4o image generation model, and we’re building on that philosophy here.

We also know we won’t get this balance perfect from day one. Recommendation systems and safety models are living, evolving systems, and your feedback will be essential in helping us refine them. We look forward to learning together and improving over time.

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

Source