Right to Erasure

The right is not automatic; it must be invoked under one of six specific legal bases defined in GDPR Article 17(1).

• Consent Withdrawal: The data subject withdraws consent and no other legal basis for processing exists.
• Objection to Processing: The data subject objects under Article 21(1) and there are no overriding legitimate grounds.
• Unlawful Processing: The personal data was processed in violation of GDPR.
• Legal Obligation: Erasure is required to comply with a Union or Member State legal obligation.
• Purpose Expiration: The data is no longer necessary for the original collection purpose.
• Child Data: Data collected in relation to information society services offered to a child.

Unlike databases where records can be deleted, neural networks memorize training data within their weights. Erasure requires addressing this fundamental technical challenge.

• Weight-Level Memorization: Large language models can verbatim reproduce training sequences, including PII, when prompted adversarially.
• No Direct Deletion: There is no SQL DELETE equivalent for a trained model; the influence of a data point is distributed across millions of parameters.
• Exact vs. Approximate Unlearning: Exact unlearning requires proving the model is indistinguishable from one trained without the data. Approximate unlearning aims to remove the data's influence to a statistically negligible level.

Machine unlearning is an active research area with several emerging techniques to comply with erasure requests without full retraining.

• SISA (Sharded, Isolated, Sliced, Aggregated): Training data is partitioned into isolated shards. Erasure requires retraining only the affected shard, dramatically reducing compute cost.
• Gradient Ascent: Reversing the training process by applying the negative of the original gradient updates for the target data, effectively 'unlearning' it.
• Differential Privacy Unlearning: Training with DP guarantees ensures that no single data point significantly influences the model, making certified removal feasible.
• Scrubbing: Post-hoc techniques that identify and suppress specific knowledge within model weights without retraining.

Article 17(3) carves out critical exceptions where the right to erasure does not apply, balancing privacy with other fundamental rights.

• Freedom of Expression: Erasure cannot override the right to freedom of expression and information, particularly relevant for journalistic or artistic content.
• Legal Claims: Processing necessary for the establishment, exercise, or defense of legal claims is exempt.
• Public Interest: Archiving purposes in the public interest, scientific or historical research, or statistical purposes where erasure would render impossible or seriously impair the achievement of those objectives.
• Public Health: Processing necessary for reasons of public interest in the area of public health.

Data controllers face strict procedural requirements when a valid erasure request is received.

• Response Deadline: The controller must act without undue delay and at the latest within one month of receiving the request.
• Extension: This period may be extended by two further months where necessary, taking into account the complexity and number of requests.
• Notification Duty: The controller must communicate the erasure to each recipient to whom the data was disclosed, unless this proves impossible or involves disproportionate effort.
• Verification: Controllers must take reasonable steps to verify the identity of the requester before proceeding with erasure.

A critical distinction exists between deleting data from storage and removing its influence from a model. Compliance requires both.

• Storage Deletion: Removing the raw personal data from databases, data lakes, and backup systems is the straightforward first step.
• Model Retraining: The complex second step involves retraining the model from scratch without the deleted data or applying machine unlearning techniques.
• Cascade Effect: If a model trained on deleted data was used to generate synthetic data or fine-tune downstream models, those derivative artifacts may also require remediation.
• Audit Trail: Controllers must maintain a Record of Processing Activities (RoPA) documenting the erasure process to demonstrate compliance to supervisory authorities.

The technical process of removing the influence of specific data points from trained model weights post-deployment. Unlike simple database deletion, unlearning requires gradient-based scrubbing or retraining from a cleansed checkpoint to eliminate memorized patterns. Key approaches include:

• Exact unlearning: Retraining the model without the target data (computationally prohibitive)
• Approximate unlearning: Using influence functions to estimate and reverse parameter updates
• Sharding-based unlearning: Partitioning training data across sub-models so only one shard requires retraining

This remains an open research challenge, as models can retain latent representations even after scrubbing.

A core privacy principle under Article 5(1)(c) of GDPR mandating that data collection be limited to what is strictly necessary for a specific purpose. This directly challenges the large-scale, indiscriminate scraping practices common in foundation model training. Key implications:

• Training datasets must demonstrate proportionality between data volume and stated purpose
• Data hoarding for speculative future AI use violates this principle
• Organizations must implement automated culling to remove stale or unnecessary records from training corpora

Data minimization serves as a proactive defense—less data ingested means less data to erase upon request.

A GDPR principle requiring that personal data be kept in an identifiable form for no longer than necessary for the original processing purpose. When applied to AI training pipelines:

• Training corpora must have automated expiration policies tied to purpose completion
• Periodic model retraining from fresh, compliant datasets becomes mandatory
• Stale data lingering in vector databases or embedding stores creates ongoing compliance liability

This principle mandates the systematic deletion of opted-out data from all AI-accessible storage layers, including backup archives and intermediate processing caches.

The automated tracking of data's origin, movement, and transformation over time, providing a forensic audit trail to verify that training data has not been contaminated by unauthorized or opted-out sources. Critical capabilities include:

• Provenance graphs mapping every data point to its consent record
• Transformation logging to trace how raw data becomes training examples
• Downstream impact analysis to identify all models affected by a single erasure request

Without robust lineage, organizations cannot confidently certify that erased data has been removed from all derivative models and embeddings.

A legal provision under Article 21 of GDPR granting individuals the absolute right to object to processing of their personal data for direct marketing or legitimate interest purposes, including AI profiling and training. Unlike the Right to Erasure, this right:

• Requires no demonstration of harm when objecting to direct marketing
• Forces controllers to demonstrate compelling legitimate grounds that override individual rights
• Applies immediately upon request, mandating cessation of processing while the objection is evaluated

This right is increasingly invoked to halt the use of personal data in foundation model training pipelines.

A specific data subject right under Article 22 of GDPR allowing individuals to refuse to be subject to solely automated processing—including AI profiling—that produces legal or similarly significant effects. Key provisions:

• Covers decisions made without meaningful human intervention
• Requires human review mechanisms for contested automated outcomes
• Extends to model training if the resulting system makes consequential decisions about the data subject

This right intersects with erasure when individuals demand deletion of the inference data and profiling outputs derived from their personal information.