Hailuo AI

Scenario

During its global expansion, Hailuo AI faced highly complex and sensitive compliance challenges unique to AI-generated content (AIGC):
1. Subtle Pornography Detection: AI-generated images may contain phallic suggestions (e.g., sex toys, suspicious food shapes), genital contours, or suggestive facial expressions that traditional models often fail to capture.
2. Strict Minor Protection: The platform must identify and intercept P0-level regulatory violations involving minors, such as suggestive poses, nudity, pregnancy implications, or abusive themes.
3. High-Precision Performance Requirements: While supporting the rapid growth of large-scale AIGC services, the platform must maintain extremely low false positive and leakage rates for "Reject" categories to balance creative freedom with regulatory baselines.

Solutions

DeepCleer delivered a full-lifecycle compliance solution for Hailuo AI, covering "Rapid Integration, Multi-dimensional Detection, and Long-term Monitoring":
1. Rapid Integration & Standard Alignment: Leveraging collaborative experience from mature projects like Talkie, Shumei aligned overseas moderation standards within a minimal timeframe to support the quick launch of image/video generation scenarios.
2. Hyper-Granular Visual Recognition Labels:- General Pornography: Established specialized rules for granular dimensions including phallic objects, sexual acts, breast deformities (e.g., gynaecomastia), and genital contours.- Minor Protection: Implemented deep recognition for high-risk scenarios such as minor sexiness, suggestive themes (e.g., suspicious liquids), nudity, and pregnancy, ensuring full blockage except for specific exempt cases (e.g., infants).
3. Bi-directional Inspection & SLA Mechanism:- Hailuo Side: Conducts regular weekly assessments of 1,500 images.- Deepcleer Side: Performs daily targeted inspections of 2,000 images, covering random Reject/Pass checks and targeted Review risk points.
4. Dynamic Policy Optimization: Addresses AIGC-specific issues (e.g., false positives caused by abnormal anatomical structures) through continuous manual feedback loops to refine AI policies.