L6. AI Data Governance: Classification, Lineage & Residency
Course outlineLesson 6 of 21
The data that flows into an AI system (training data, retrieval context, user inputs) carries the same classification and compliance obligations as data in any other system. But AI creates new data governance challenges that traditional DLP and classification tools were not built for.
Why AI Creates New Data Governance Challenges
Traditional data governance assumes data sits in structured stores (databases, file systems) where classification labels, access controls, and DLP rules can be applied. AI systems disrupt this model in three ways:
- Training data aggregation: a model trained on data from multiple sensitivity levels can memorize and reproduce sensitive content in unexpected contexts
- Inference-time mixing: user queries and retrieved context from different classification levels are mixed in a single prompt, making per-field classification insufficient
- Unstructured content: AI systems routinely ingest unstructured text (emails, documents, web pages) that traditional DLP tools classify poorly
Data Classification for AI Pipelines
Classification must be applied at every stage of the AI data lifecycle:
| Stage | Data Types | Classification Consideration |
|---|---|---|
| Training data collection | Web scrapes, licensed datasets, internal documents | PII detection, sensitive category identification, copyright status |
| Fine-tuning datasets | Curated examples, human feedback | Same as above, plus annotation data provenance |
| RAG retrieval corpus | Internal documents, knowledge bases | Apply existing classification labels to retrieval index |
| Inference-time inputs | User queries, uploaded files | Classify and log; apply to output generation policy |
| Model outputs | Generated text, summaries | Inherit classification from inputs; apply DLP to outputs |
Data Lineage
Data lineage for AI means being able to answer: what training data influenced this model's behavior? This is harder than traditional lineage because:
- The training process blends data points: no single record maps cleanly to a specific model output
- Fine-tuning on private data changes behavior globally: you cannot "remove" a training point without retraining
- Retrieval-augmented generation (RAG) provides more tractable lineage: each retrieved chunk can be logged alongside the response
Data Residency
AI APIs introduce residency risks that are easy to overlook. When you send a prompt to a cloud AI provider, the prompt content (which may include PII or confidential data) is processed in the provider's infrastructure, potentially in a different jurisdiction.
Controls for residency compliance:
- Use private endpoints or VNet injection to keep traffic within your network perimeter
- Choose AI service regions aligned with your data residency requirements (EU data to EU regions)
- Review provider data processing agreements for training opt-out clauses
- Consider on-premises or private cloud AI deployment for the highest-sensitivity use cases
- ✓AI creates three new governance challenges: training data aggregation (memorization), inference-time classification mixing, and unstructured content ingestion
- ✓Classification must be applied at every stage: training collection, fine-tuning, RAG corpus, inference inputs, and model outputs
- ✓Training data lineage: maintain a manifest recording source, version, classification, and consent basis for each dataset
- ✓RAG provides more tractable lineage than trained models: each retrieved chunk can be logged alongside the response
- ✓Data residency risk: prompts sent to cloud AI providers may be processed in a different jurisdiction; use private endpoints and region-specific deployments
1. Why is training data lineage harder to maintain for fine-tuned models than for RAG systems?
2. Which control best addresses data residency compliance when using a cloud AI API?
Recommended: Pluralsight
Pluralsight's AI security courses cover threat modeling, governance, and practical red teaming for AI systems to complement what you learn here.