Get the weekly research letter
Companies change their terms quietly. We read every version and catch what actually changed. One email a week on the changes that matter and what they mean. No account.
OpenAI states that all GPT-5 model variants incorporate a safety training approach called safe-completions, which the document describes as designed to prevent disallowed content from being produced.
This analysis describes what OpenAI's agreement states, permits, or reserves. It does not constitute a legal determination about enforceability. Regulatory applicability and practical outcomes may vary by jurisdiction, enforcement context, and individual circumstances. Read our methodology
This provision establishes a uniform safety training layer applied across all GPT-5 model variants, including those accessed via the API, which may affect the range of outputs available to developers and enterprise users building applications on the platform.
Interpretive note: The document does not define the scope of disallowed content covered by safe-completions, the testing methodology used, or how the mechanism interacts with developer-level API customization, making independent assessment of its operational scope uncertain.
This provision establishes that safe-completions governs output generation across all GPT-5 models, meaning queries that would produce disallowed content as defined by OpenAI's policies will be restricted regardless of which model variant or access method is used.
Cross-platform context
See how other platforms handle Safe-Completions Safety Training Across All Models and similar clauses.
Compare across platforms →Monitoring
OpenAI has changed this document before.
Receive same-day alerts, structured change summaries, and monitoring for up to 25 platforms.
"All of the GPT‑5 models additionally feature safe-completions, our latest approach to safety training to prevent disallowed content.Excerpt from OpenAI's GPT-5 System Card
1) REGULATORY LANDSCAPE: The safe-completions mechanism engages content moderation obligations that may be relevant under the EU Digital Services Act for large online platforms, the EU AI Act's requirements for general-purpose AI model providers regarding harmful content prevention, and FTC standards for AI system safety representations. The document does not define the scope of disallowed content in this provision, and the adequacy of the mechanism relative to regulatory content safety standards may require independent evaluation. 2) GOVERNANCE EXPOSURE: Low to Medium. The existence of a stated safety training layer is consistent with commonly disclosed AI platform practices. However, the document does not detail what content categories are covered by safe-completions, what testing was conducted, or what failure modes were identified, which limits independent verification of the mechanism's adequacy. 3) JURISDICTION FLAGS: EU/EEA deployers face heightened obligations under the EU AI Act and Digital Services Act regarding content safety documentation and auditability. Deployers in sectors such as healthcare, legal services, or education should assess whether safe-completions adequately addresses sector-specific content restrictions applicable in their jurisdiction. 4) CONTRACT AND VENDOR IMPLICATIONS: API developers building applications on GPT-5 should assess whether safe-completions restrictions are documented in OpenAI's API terms and whether the restriction scope aligns with their own product content policies and user-facing commitments. Applications that rely on specific output types should verify compatibility with safe-completions enforcement before deployment. 5) COMPLIANCE CONSIDERATIONS: Compliance teams should request documentation of the safe-completions testing methodology and disallowed content scope from OpenAI to support internal AI governance records. Organizations subject to AI transparency or auditability obligations should assess whether this disclosure is sufficient for their regulatory reporting requirements.
This provision establishes a uniform safety training layer applied across all GPT-5 model variants, including those accessed via the API, which may affect the range of outputs available to developers and enterprise users building applications on the platform.
This provision establishes that safe-completions governs output generation across all GPT-5 models, meaning queries that would produce disallowed content as defined by OpenAI's policies will be restricted regardless of which model variant or access method is used.
No. ConductAtlas is an independent monitoring service. We are not affiliated with, endorsed by, or sponsored by OpenAI.