The Mechanics of Substrate Dissolution
Statutory rights are constructed upon an assumed ontology. When the Court of Justice of the European Union recognized the right to be forgotten in 2014, and when the GDPR codified Article 17 in 2016, deletion assumed a discrete object: a database row, an index pointer, a log entry, or a document file. An index entry existed, could be located, and could be removed.
Training a large neural model destroys that ontology. Personal data enters the training corpus as text and leaves as parameter adjustments distributed across billions of numerical weights. There is no row corresponding to an individual, no file to shred, and no addressable record to delete. The right remains on the statute books of the European Union, India, Brazil, and California, but the machine on the other side of the promise no longer holds personal data in a form the right can reach.
Architectural Repeal operates without an official journal, without a floor debate, and without an effective date. It calls back a digital right with none of the procedural accountability of a legislature.
Code as De Facto Repeal
Legal scholarship established that code regulates conduct as effectively as law does. Architectural repeal marks a further development: technical architecture removing the object of an enacted law rather than merely regulating where positive law is silent.
The phenomenon extends beyond the right to erasure:
- Rectification: Assumes an identifiable error that can be located and corrected; in a deep neural network, no engineer can point to the specific parameter holding a falsehood.
- Access: Assumes a controller can disclose what personal data it holds about a subject; with a trained foundation model, the honest technical response is that nobody knows.
- Purpose Limitation: Assumes data collected for one task is segregated from other pipelines; weights absorb features holistically, dissolving boundaries between distinct processing purposes.
Impossibility as Practical Immunity
In traditional legal doctrines, when performance of an obligation becomes impossible, the party must plead impossibility openly in an adversarial proceeding, prove it, and accept the consequences (such as damages, restitution, or contract discharge). In machine learning deployments, AI companies decline to plead technical impossibility before data protection authorities, because doing so would concede that personal data persists inside the weights. Impossibility instead operates as an unadjudicated, practical immunity.
The Transatlantic Divergence
Regulators facing this collision have adopted incompatible ontologies:
- European Accommodation: Certain European discussions (such as the Hamburg DPA 2024 discussion paper) conclude that model weights store mathematical probabilities rather than personal data, restricting erasure rights to model inputs and outputs while leaving the weights untouched—resulting in Erasure Theater.
- Algorithmic Disgorgement: The United States Federal Trade Commission (in matters such as Everalbum and Kurbo) treats trained models and algorithms as holding the underlying illegal data, commanding the complete destruction of models trained on tainted inputs.
A right exercisable only where technical architecture chooses to accommodate it ceases to function as a fundamental right and acquires the operational status of an optional feature.
Field Notes & Ephemera
Field Note: "A parliament can restore what a parliament repealed. No comparable mechanism has been identified for what architecture repeals. When data is dissolved into irreversible parameters, the statute remains printed while the remedy evaporates."