# Privacy-Aware Infrastructure in the AI Era: Meta's Asset Classification Approach How Meta combines LLMs and deterministic rules for scalable data governance. - **Source:** [Facebook Engineering](https://engineering.fb.com) - **Published:** 2026-06-25 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/a595746d-1490-41af-afa2-35c26dc46fba ## The Core Challenge of Privacy Controls Data privacy systems must understand data nature to enforce protection policies. A field like 'age' could be personal data or system metadata. This ambiguity grows more complex with AI products introducing new data modalities and rapid iterations. ## AI-Native Complexity AI systems bring multimodal data, rapid iterations, and derived features. Traditional manual reviews can't keep pace. Meta's solution is a hybrid model using LLMs strategically to handle ambiguity while maintaining human oversight for critical decisions. ## Hybrid Classification Model Meta's approach rests on three principles: Context beats prompts, human oversight remains at critical points, and LLM learnings are distilled into versioned rules. This gradually reduces LLM dependency in production while maintaining adaptability. ## Why Asset Classification Matters Accurate data asset classification forms the foundation for all privacy controls - retention periods, access policies, and sharing rules. Misclassification can cascade errors throughout the entire enforcement pipeline. ## Four Key Challenges The system addresses: Noisy signals, distributed context, evolving requirements, and actionable outcomes. For example, an 'age' field without context analysis might trigger incorrect pipeline-wide restrictions. ## Context is King Classification requires assembling scattered context - code lineage, ownership, semantic annotations, and usage patterns. Meta's model structures this into evidence briefs that help models make informed decisions. ## Human-Machine Collaboration Humans remain in the loop for two critical functions: Approving reference labels and reviewing rule promotions. LLMs interpret ambiguous cases but production rules require human validation. ## Rule Distillation Learnings from LLMs are gradually converted into versioned, human-reviewed deterministic rules. This makes the system more predictable over time while maintaining auditability. ## The Enforcement Pyramid Privacy-aware infrastructure has four layers: Understand (classification), discover, enforce, and demonstrate compliance. Classification is the load-bearing base - errors here affect everything above it. ## Data Lifecycle in AI Systems In AI systems, data constantly transforms - from raw signals to features, model inputs, or derived datasets. Classification must follow data meaning rather than just its structural form. ## Determinism for Auditing Production rules must be low-latency, replayable, and auditable. Meta's model preserves this determinism by using LLMs selectively and distilling their outputs into versioned rules. ## The Future of Privacy Infrastructure Meta's approach provides a scalable framework for AI-era data governance. This hybrid model successfully balances ambiguity handling with enforceable requirements, setting a pioneering example for the industry. ## Sources covering this story - https://engineering.fb.com/2026/06/25/security/privacy-aware-infrastructure-in-the-ai-native-era-an-asset-classification-case-study/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._