Data Lineage Tracking for AI: Complete Guide | QuizBy Eyal Doron / December 6, 2025 / 1 minute of reading Data Lineage Tracking for AI: Complete Guide | Quiz 1 / 8 1. Why does the article say lineage adds minimal overhead despite concerns? 1. Async metadata capture and proper tooling minimize impact while missing lineage costs far exceed implementation 2. Overhead concerns only apply to real-time systems 3. Lineage requires no resources at all 4. Only large enterprises need to worry about overhead Correct! Why: With async metadata capture and proper tooling lineage adds minimal performance impact – the cost of missing lineage during an incident or audit far exceeds implementation overhead. Context: This addresses the misconception that lineage adds too much overhead. Remember: Async capture plus proper tooling equals minimal impact. 2 / 8 2. What does the EU AI Act require regarding training data according to the article? 1. Training data documentation for high-risk systems and demonstrable traceability requirements 2. Documentation is optional for all risk levels 3. No documentation is required for any AI systems 4. Only the model output needs to be documented Correct! Why: The EU AI Act requires training data documentation for high-risk AI systems demonstrating what data trained the model and its characteristics plus traceability requirements. Context: Lineage is the technical foundation for meeting these regulatory requirements. Remember: Document training data plus demonstrate traceability. 3 / 8 3. How does lineage support GDPR right to erasure according to the article? 1. Lineage shows which models were trained on a person's data enabling accurate deletion compliance 2. Erasure only requires deleting the original source data 3. GDPR does not apply to AI training data 4. Lineage automatically deletes data when requested Correct! Why: If someone requests deletion you need to know which models were trained on their data – lineage answers this question and without it you cannot comply accurately. Context: Right to erasure creates complex challenges for AI that only lineage can address. Remember: Deletion requests require knowing which models used the data. 4 / 8 4. What tools does the article recommend for different lineage roles? 1. MLflow for experiment tracking – DVC for data versioning – Apache Atlas for enterprise lineage 2. Any database system works equally well 3. Only spreadsheets and manual documentation 4. Custom tools must be built from scratch Correct! Why: The article recommends MLflow for experiment tracking and model-to-data linkage and DVC for dataset version control and Apache Atlas for enterprise lineage and regulatory audits. Context: Tool integration is one approach for practical lineage implementation. Remember: MLflow for experiments – DVC for data – Atlas for enterprise. 5 / 8 5. What is the critical link for backward lineage according to the article? 1. API authentication tokens 2. Network connection between servers 3. Model-to-data linkage connecting each trained model to its training dataset versions 4. Database foreign keys Correct! Why: Model-to-data linkage explicitly connects each trained model to its training dataset versions – without it you cannot trace a prediction back to its training data. Context: Dataset version identification assigns unique identifiers to training data snapshots. Remember: No model-to-data link equals no backward traceability. 6 / 8 6. What metadata should be captured during the data collection stage? 1. Just the database connection string 2. Only metadata required by the AI model 3. Source system identification – collection timestamps – consent and permission metadata 4. Only the file size and format Correct! Why: The article specifies capturing source system identification (which database or API) and collection timestamps (when data was extracted) and consent and permission metadata (legal basis for use). Context: This metadata becomes critical for GDPR compliance. Remember: Source – Timestamp – Consent. 7 / 8 7. Why does feature engineering obscure data origins according to the article? 1. Features are stored in different databases than source data 2. Feature engineering deletes the original data 3. Derived features like ratios and aggregations create indirect connections to dozens of underlying data points 4. Engineering transforms data into unreadable formats Correct! Why: When you derive new features like ratios and aggregations and embeddings the connection to original data becomes indirect – a customer_risk_score might derive from dozens of underlying data points. Context: This is one of several factors that make AI lineage harder than traditional data lineage. Remember: Derived features hide their sources. 8 / 8 8. What is the difference between forward and backward lineage? 1. Forward traces source to output while backward traces output to source 2. Forward is for training while backward is for inference only 3. Forward is for new data while backward is for historical data 4. Forward is automatic while backward requires manual effort Correct! Why: Forward lineage traces data from source to output answering what happened to this data while backward lineage traces from output to source answering where did this prediction come from. Context: Both directions matter – forward supports compliance and auditing while backward enables debugging and explanation. Remember: Forward equals source to output – Backward equals output to source. Your score isThe average score is 0% Restart quiz Download PDF Please leave this field empty๐ The AI Security Manager's Newsletter Weekly insights on AI risk management, EU AI Act compliance, and practical security strategies. We donโt spam! Read our privacy policy for more info. Thank you! Please check your inbox to confirm your subscription.
