Data Subject Rights

The right to erasure ('right to be forgotten') compels controllers to delete personal data without undue delay when it is no longer necessary for the original purpose, consent is withdrawn, or processing was unlawful. In machine learning contexts, this presents a unique challenge: simply deleting a record from a database is insufficient if the data has already been used to train a model. True erasure may require model unlearning—the computationally intensive process of removing a data point's influence from trained weights without full retraining. Techniques like SISA training (Sharded, Isolated, Sliced, Aggregated) partition data to limit the blast radius of deletion requests.

The right of access entitles individuals to obtain confirmation of whether their personal data is being processed and, if so, to receive a copy along with contextual metadata. In AI governance, this extends beyond structured database records to include:

• Inferred data: Derived attributes and profile scores generated by models
• Training provenance: Whether the individual's data was included in a training corpus
• Logic of automated decisions: Meaningful information about the reasoning behind algorithmic outcomes Fulfilling access requests in Retrieval-Augmented Generation (RAG) systems requires tracing which source documents contributed to a generated output.

The right to rectification mandates that inaccurate or incomplete personal data be corrected without undue delay. In enterprise AI systems, this right cascades through multiple layers: the source record must be corrected, any feature stores derived from that record must be updated, and downstream models that consumed the erroneous data may need retraining or fine-tuning. This creates a data lineage dependency—organizations must maintain end-to-end traceability from raw ingestion through feature engineering to model output. Rectification requests often trigger data quality workflows that validate corrections against business rules before propagation.

The right to data portability allows individuals to receive their personal data in a structured, commonly used, and machine-readable format and to transmit it to another controller without hindrance. For AI systems, this right applies to data provided by the data subject, not inferred or derived data. Key technical considerations include:

• Format standardization: JSON, CSV, or Parquet exports with clear schemas
• Scope limitation: Excludes proprietary model weights or aggregated insights
• Interoperability: Enabling seamless transfer between competing platforms This right is a cornerstone of data sovereignty, preventing vendor lock-in by ensuring individuals maintain control over their digital footprint.

Article 22 of the GDPR establishes the right not to be subject to solely automated decisions that produce legal or similarly significant effects. This includes profiling that evaluates personal aspects such as work performance, economic situation, health, or behavior. In practice, this requires organizations to implement human-in-the-loop (HITL) or human-on-the-loop (HOTL) override mechanisms. Exceptions exist where the decision is necessary for a contract, authorized by law, or based on explicit consent. Even when automated decisions are lawful, controllers must provide meaningful information about the logic involved and the envisaged consequences.

The right to restriction acts as a temporary freeze on data processing while a dispute is resolved. Unlike erasure, the data is preserved but marked as restricted, with processing limited to storage only unless consent, legal claims, or public interest grounds apply. This right is invoked when:

• Accuracy is contested: Processing halts while the controller verifies correctness
• Processing is unlawful: The data subject opposes erasure and requests restriction instead
• Purpose is exhausted: The controller no longer needs the data, but the subject requires it for legal claims In data lakehouse architectures, restriction requires tag-based access controls that dynamically block compute engines from accessing flagged records.