OpenAI's Stricter Rules: Why ChatGPT Won't Talk About Goblins Anymore

Lisa Ernst · 04.05.2026 · Artificial Intelligence · 7 min

In our increasingly interconnected world, the conversations we have online profoundly shape how we understand and interact with each other. As artificial intelligence models become more accessible and integrated into daily life, the responsibility of ensuring these platforms foster safe and constructive environments grows exponentially. OpenAI, a leading AI research and deployment company, grapples with this challenge through its comprehensive content moderation policies and tools.

Source: AI’s growing presence in daily life highlights the critical need for robust content moderation to foster safe online interactions.

Quick Summary: OpenAI's Content Moderation

• Universal Usage Policies: Updated October 29, 2025, these policies ensure consistent guidelines across all OpenAI products and services, prioritizing safety and user control.
• Prohibited Content: OpenAI strictly forbids content related to threats, harassment, self-harm, sexual violence, child sexual abuse material (CSAM), and terrorism.
• Moderation API: A free technical tool for developers to identify and filter harmful text and images, offering detailed flagging and categorization.
• Guardrails System: Real-time validation of AI assistant messages, preventing offensive or policy-violating content from reaching users.
• Continuous Updates: OpenAI regularly refines its moderation models and policies, requiring developers to stay informed and adapt.

The Foundation of Moderation: OpenAI's Usage Policies

OpenAI has been consistently refining its approach to content moderation. On October 29, 2025, the company updated its usage policies to establish a unified set of guidelines across all its products and services. These policies prioritize the secure and responsible use of their tools while maximizing user control. Earlier, on January 29, 2025, these universal guidelines were clarified further to detail prohibitions based on applicable law.

Significant changes also occurred on January 10, 2024, when the guidelines offered clearer articulation and more specific instructions for services. A major consolidation took place on February 15, 2023, merging application-specific and content guidelines into a single set of usage policies, providing detailed guidance on prohibited activities in high-risk industries. Since November 9, 2022, OpenAI no longer requires application registration, instead relying on automated and manual methods to monitor policy violations. The app review process itself saw an update on October 25, 2022, allowing developers to deploy applications without awaiting approval, provided they adhere to the established guidelines, as outlined in their usage policies.

OpenAI strictly prohibits the use of its services for a range of harmful activities. This includes explicit threats, intimidation, harassment, or defamation. The promotion or facilitation of suicide, self-harm concepts, or eating disorders is also forbidden. Content involving sexual violence or non-consensual intimate material falls under these prohibitions. The policies also explicitly ban terrorism or any form of violence, including hateful violence. Crucially, any material involving child sexual abuse (CSAM) is forbidden, irrespective of whether it is AI-generated.

The solicitation of minors and exposure of minors to age-inappropriate content, such as graphic self-harm, sexual, or violent material, are also prohibited. Furthermore, encouraging unhealthy diet or exercise behaviors in minors through OpenAI's services is not allowed, as detailed in their official usage policies.

Source: roborhythms.com

OpenAI’s usage policies explicitly prohibit a wide range of harmful content, from harassment and threats to child sexual abuse material.

The Moderation API in Practice

OpenAI’s Moderation API serves as a key technical tool in identifying potentially harmful content in both text and images. This endpoint is available for free, allowing developers to filter content or intervene with user accounts generating offensive material. When sending image files to the Moderation endpoint, a 20MB size limit applies, as described in their API documentation.

Source: community.openai.com

This dashboard illustrates how the Moderation API processes content, flagging it as potentially harmful or compliant.

Developers can utilize two distinct models for the Moderation endpoint: omni-moderation-latest for multimodal inputs and a broader range of categories, and text-moderation-latest as a legacy option exclusively for text. The output of the Moderation API provides several key indicators:

• flagged status: True if the content is potentially harmful.
• categories: Shows violation flags per category.
• category_scores: Provides confidence values between 0 and 1.
• category_applied_input_types: For Omni models, specifies which input types (image, text) triggered flags for each category.

OpenAI continuously updates the underlying model of the Moderation endpoint, which may necessitate recalibrating custom policies.

Moderation API Categories

The Moderation API categorizes content into various types of harm. Here's a breakdown of the main categories and their input type support:

The illicit category covers content offering advice or instructions for illegal actions, while illicit/violent specifically refers to content that includes violence or obtaining a weapon. The sexual/minors category flags sexual content involving individuals under 18 years of age, as documented in the OpenAI Moderation guide.

Managing User-Generated Content with Guardrails

Effective management of user-generated content is crucial for upholding OpenAI's policies and avoiding API bans. Implementing unique user identifiers (user_id) in every API call helps OpenAI attribute policy violations to specific users, safeguarding applications from blanket penalties. Companies should educate users about acceptable content guidelines to minimize violations.

Detailed logging of user interactions assists in identifying patterns of misuse and refining moderation strategies. Rate limiting and throttling mechanisms prevent excessive requests and repeated submission of problematic content. A robust error handling system should also provide clear and informative feedback to users when moderation rejects requests. Due to the evolving nature of OpenAI's content policies, regular review and updates of internal policies and moderation strategies are essential.

The Guardrails and Content Moderation system validates assistant messages in real-time before they reach users. This system classifies agent outputs based on moderation policies, identifying offensive content, off-brand messages, and violence. Messages that violate policies are marked as FAIL and hidden, while compliant messages are marked as PASS and displayed normally. The Guardrails system operates asynchronously during message streaming, employing a separate LLM classifier (gpt-4o-mini). Guardrails categories include OFFENSIVE (encompassing hate speech and discriminatory language), OFF_BRAND (like discussions about competitors), VIOLENCE (such as explicit threats or descriptions of violence), and NONE for perfectly compliant messages. The core classification component, runGuardrailClassifier, evaluates the message content. The classifier prompt defines an expert role, the company name, the message for classification, and detailed output class descriptions.

Conclusion

OpenAI’s multi-layered approach to content moderation, encompassing clear usage policies, a robust Moderation API, and real-time Guardrails, demonstrates a profound commitment to fostering safe online environments. As AI technology advances, continuous refinement of these systems remains paramount to navigating the complexities of digital communication and upholding ethical standards.

Frequently Asked Questions