Addendum to GPT-4o System Card: 4o image generation

Source: OpenAI Date: 2025-03-25 URL: https://openai.com/index/gpt-4o-image-generation-system-card-addendum

Summary

System card addendum covering GPT-4o’s native image generation capability, released in March 2025. The addendum documents safety evaluations specific to image generation: CSAM resistance, photorealism limits, political content handling, and deepfake/impersonation guardrails. It supplements the original GPT-4o system card with image-specific red-teaming findings and mitigation commitments.

Implications

Safety/policy thread. Addendum format signals that image generation introduced risk categories not covered in the base system card — OpenAI is acknowledging that multimodal expansion requires separate safety accounting. The photorealism and impersonation sections directly respond to public pressure following the viral rollout of 4o images in early 2025.

Product thread. Native image generation in 4o (rather than a separate DALL-E endpoint) marks the consolidation of generation into the flagship chat model. The system card addendum is the governance artifact that enabled this consolidation to ship.