SafeGen: Embedding Ethical Safeguards in Text-to-Image Generation

SafeGen addresses real harms from text-to-image AI: The framework tackles three concrete problems—societal bias amplification, high-quality disinformation, and intellectual property violations—that currently affect widely-used systems like DALL-E, Stable Diffusion, and Midjourney.

Ethical safeguards don't require sacrificing creative quality: SafeGen demonstrates that safety filters and image quality can coexist; the system achieved strong technical performance (FID score of 22.08, SSIM of 0.79) while filtering harmful prompts, suggesting responsible AI design doesn't mean limiting creative capability.

Two-layer protection is more effective than single approaches: The framework combines prompt filtering (BGE-M3 classifier achieving 0.81 F1-Score) with an optimized generation model (Hyper-SD), with ablation studies confirming that both components are necessary—neither alone is sufficient for comprehensive safety.

Multilingual and fairness-aware training matters: SafeGen was built on a curated dataset spanning English and Vietnamese with fairness-conscious training processes, suggesting that responsible AI systems must be intentionally designed for diverse languages and populations rather than defaulting to English-only approaches.

Practical applications extend beyond blocking misuse: Case studies show SafeGen's real-world value includes generating inclusive educational materials and maintaining academic integrity, indicating that ethical safeguards can actively support positive uses rather than only preventing harms.