DALL·E 3 System Card Released: OpenAI Publishes System-Level Safety and Performance Boundary Documentation for Its Text-to-Image Model

Purpose and Value of the System Card

The DALL·E 3 system card is OpenAI’s inaugural standardized system impact document for its text-to-image model, aligned with the Model Card framework promoted by ML Commons, but extended to cover end-to-end generation systems—including model architecture, training data provenance, inference service logic, and integrated content policy enforcement.

Key Empirically Measured Metrics and Disclosure Dimensions

• Content Safety: On internal red-teaming datasets, DALL·E 3 achieves ≥98.7% refusal rate for prompts containing violent, adult, or hateful content; however, evasion via semantic obfuscation or cross-lingual paraphrasing (e.g., phonetic transliteration, code-based encoding) yields ~4.2% false-negative rate;
• Image Fidelity: Achieves 93.1% text rendering accuracy on the COCO-Text benchmark (vs. 76.5% for DALL·E 2), though multilingual mixed-text rendering (e.g., Chinese–English juxtaposition) remains unsupported;
• Prompt Robustness: Maintains ≥91.4% successful generation rate under misspelled, grammatically malformed, or syntactically redundant prompts—substantially outperforming prior versions;
• Constrained Generation Capabilities: Explicitly enumerates prohibited outputs—including photorealistic depictions of identifiable real-world individuals, geographically precise landmarks, and copyrighted character likenesses (e.g., Disney characters)—implemented via joint fine-tuning (RLHF + classifier-guided decoding).