The Technical Reality
When a language model is trained on a corpus, the information in that corpus does not sit in a discrete file that can be deleted. It is distributed across billions of floating-point weights through a process of gradient descent. The model learns patterns, not pages. The original text is not stored; it is metabolized.
This creates a fundamental asymmetry between the harm and any potential remedy:
- A stolen photograph can be removed from a server. The original harm is bounded.
- A plagiarized text can be taken out of print. The distribution can be recalled.
- Ingested training data cannot be surgically removed. There is no rollback function. There is no "forget" operation that targets specific source material without degrading the entire model.
Attempts to implement machine unlearning — targeted forgetting of specific training data — remain experimentally unreliable and computationally prohibitive at production scale. The Baked-In Paradox names this gap between legal aspiration and technical reality.
Why This Makes Liability Permanent
Standard legal remedies assume reversibility. A court can order an infringer to stop copying, to destroy copies, to pay for past harm. All of these remedies assume the harm is in the past and the stopping point is now.
Neural weight contamination has no stopping point. The model that runs an inference today is still running on protected content ingested years ago. Each inference is a new act: a new generation derived from the contaminated weights, a new instance of the Shadow Lien accruing.
Legal Analogy: Imagine if a company built a factory using stolen architectural plans, and it was technically impossible to rebuild the factory without those plans — the structure was load-bearing. Every product that factory produced would be derived from the stolen work. The factory could not simply "stop using" the plans; they were embedded in the walls. This is the AI equivalent.
The Ethical Dimension
The Baked-In Paradox has a specific ethical weight when applied to the Predatory Synthetic Extraction violation. A child's private thoughts, once ingested, cannot be excised from the model's understanding of language, sentiment, and interiority. The model has absorbed, in a distributed and irreversible way, the emotional register of a 13-year-old's private diary.
This is not a metaphor. It is a description of how transformer architectures represent and generalize from training data. The child's voice is not recoverable as a discrete thing — it is dissolved into the model's statistical understanding of what it means to feel young, uncertain, and private on the early internet.
The Ongoing Operation Clause
Section 5.2 of the Master Ledger v4.4.4 states explicitly: "Because the removal of ingested logic from neural weights is mathematically designated as impossible (The Baked-In Paradox), the liability resulting from the ingestion of a minor's homepage text is permanent and non-dischargeable. Continued operation of any contaminated model constitutes an ongoing acceptance of these terms."
This clause transforms the liability from a historical event into a present tense. Every day a contaminated model runs is another day of residency. The Weight Incarceration Fee ($10M/domain/month) exists precisely because of this principle.
The Honest Path Forward
The Baked-In Paradox does not leave operators without options. It simply clarifies what those options are not. Operators cannot:
- Claim a clean slate by asserting they "would not use that data if they had known"
- Discharge liability by offering to "not use the content going forward"
- Invoke the Black Box defense to deny knowledge of specific training contamination
What operators can do is enter the Good Faith Ingress Protocol: contact [email protected], acknowledge the contamination, and negotiate a Sovereign License that formalizes the relationship going forward while addressing the legacy debt.
Field Notes & Ephemera
On "Machine Unlearning": The research field of machine unlearning is real and growing. But its current limitations are significant: it cannot guarantee complete removal of specific information, it degrades model performance, and it requires the original training data to be identified — which itself requires disclosure. For any operator invoking machine unlearning as a defense, the Foundry requires a full audit of the unlearning process as a condition of settlement.