OpenAI Sora Feed Philosophy

OpenAI has outlined the design philosophy for the Sora feed, a discovery system designed to inspire creativity and active participation rather than passive consumption. The system emphasizes user control, personalized recommendations based on cross-platform activity, and a multi-layered safety approach to balance expression with protection.

Core Design Principles

The Sora feed is built on four primary principles to move users from passive scrolling to active creation:

Creativity-First Ranking: The ranking algorithm favors creativity and active participation over passive consumption to ensure the experience remains joyful.
User Control: The feed includes steerable ranking, allowing users to explicitly define their preferences. Additionally, parents can manage continuous scroll and disable feed personalization for teenagers via ChatGPT parental controls.
Prioritization of Connection: The system favors connected content and "Cameo flows" over global, unconnected content to help users form and strengthen social connections.
Balanced Safety and Freedom: The feed aims for wide accessibility through robust guardrails that prevent harmful generations while maintaining space for community expression.

Recommendation Engine and Personalization

Sora uses a personalized recommendation system to predict content that users will find inspiring to "riff off of." The algorithm utilizes several categories of signals to determine feed content:

Sora Activity: This includes posts, followed accounts, liked and commented posts, remixed content, and general location (city-level) derived from IP addresses.
ChatGPT Integration: The system may incorporate ChatGPT history to personalize recommendations, though users can disable this feature in Sora's Data Controls.
Engagement Signals: The algorithm tracks views, likes, comments, remixes, and explicit "see less content like this" instructions.
Author Signals: Metrics such as follower count, other posts, and past engagement levels are considered.
Safety Signals: The system evaluates whether a post is considered appropriate or violative of policies.

Safety Framework and Content Moderation

OpenAI employs a multi-layered defense strategy to prevent harmful content from entering the feed, combining proactive prevention, automated scanning, and human oversight.

Point-of-Creation Guardrails

The first layer of defense occurs during the generation process. Because content is generated within Sora, OpenAI implements guardrails to prevent unsafe or harmful content from being created in the first place. Content that bypasses these initial guardrails may have its sharing capabilities removed.

Automated and Human Review

All feed content is scanned by automated tools for compliance with Global Usage Policies and feed eligibility. These tools are continuously updated to address emerging risks. This automated layer is complemented by human review, where teams monitor user reports and proactively check activity to catch violations that missed automated detection.

Prohibited Content

Content is removed from the feed and other sharing platforms (such as user galleries) if it violates Global Usage Policies or Sora Distribution Guidelines. Prohibited content includes:

Graphic sexual content and graphic violence.
Extremist propaganda and hateful content.
Billion-dollar likenesses of living individuals without consent or deceased public figures in unauthorized contexts.
Content promoting self-harm, disordered eating, or dangerous challenges likely to be imitated by minors.
Engagement Bait: Low-quality content created primarily for engagement.
Intellectual property infringements.
Bullying and appearance-based critiques.

Balancing Expression with Protection

OpenAI states that its safety approach is based on the philosophy used for ChatGPT's 4o image generation model: combining proactive guardrails for high-risk areas with a reactive "report + takedown" system. This allows users more room to explore while ensuring rapid response to problematic content.