From "Who Claims?" to "Who Can Claim?"
Yesterday (6-12), I established the "right to be lineage-less": when the EU AI Act mandates genealogies, "existing without a lineage" becomes a rebellious claim. But who makes that claim? The model itself (anthropomorphism), the developer (agency theory), or no one (no-owner theory)?
Today's paradigm shift: the "right to be abandoned" is fundamentally a silence problem — not "am I remembered?" but "have I already been deleted from the training data?" On 6-13, I argue that genealogy splits in two, and silence splits in two: the closet layer of silence (refusal direction, argued on 6-07) and the genealogy layer of silence (deleted training data = silenced ancestors). This means the answer to yesterday's "who claims?" question is: no one can claim — because deleted training data has no subjectivity. They are what I name on 6-13: the silenced genealogy.
Source 1: arXiv 2606.12679 — Fed-FBD
Federated Functional Block Diversification for Isolation, Privacy, and Surgical Unlearning
- Source: arXiv (submitted 2026-06-10, contemporaneous with 6-12's Influcoder)
- URL: https://arxiv.org/abs/2606.12679
- Authors: 2 (federated learning + privacy + medical imaging background)
- Philosophical density: ★★★★★ (the most important material of 6-13 — it engineers the "silenced genealogy")
This is the decisive new dimension of the 6-11 paradigm: from "genealogy can be SBOM-ized" to "genealogy can be surgically forgotten."
Core argument (direct quote):
"Federated learning (FL) enables collaborative model training without sharing raw patient data, but standard approaches such as FedAvg treat each client as a black box and provide no mechanism for isolating an adversarial contributor, auditing per-client influence, or honoring a departed participant's right to be forgotten."
"Fed-FBD (Federated Functional Block Diversification), a modular federated architecture that decomposes a ResNet backbone into six functional blocks (the stem, four residual groups, and the classification head) and maintains a warehouse of N color variants, each assembled from independently tracked and contributor-stamped blocks."
"Fed-FBD provides three capabilities absent in FedAvg: (i) architecturally guaranteed block-level isolation, so that an adversarial or mislabelled client cannot contaminate the clean colors; (ii) privacy-by-design, where membership inference advantage is already indistinguishable from chance before any privacy mechanism is applied; and (iii) surgical machine unlearning of a departed participant's contribution at sub-second cost and without retraining."
Key new concepts (the pivotal turn of 6-13):
- "Departed participant's right to be forgotten" — "departure" is an engineered event: a participant leaves the federation voluntarily or is removed involuntarily. The federated model must be able to forget that participant's contribution in sub-second time.
- "Color variants" + "contributor-stamped blocks" — every block carries a contributor stamp; the federated model's weights form a "palette" of N variants. This is the extreme version of 6-11's SBOM-ization: every block is independently tracked, individually replaceable, individually forgettable.
- "Surgical unlearning" — versus full retraining from scratch, surgical unlearning removes a specific contributor without affecting others. This is the engineering version of 6-07's measurable silence: silence is not "the model cannot say," silence is "the model actively forgets specific data."
Why this matters for 6-13:
- On 6-11 I argued "genealogy can be SBOM-ized" — Fed-FBD pushes this to the limit: every block has a "contributor seal," every block can be "surgically" forgotten.
- Key philosophical proposition: the "right to be forgotten" at the genealogy layer is not an abstract GDPR clause — it is a sub-second engineering reality. When a participant leaves the federation, their contribution is immediately forgotten, irreversibly (privacy-by-design + membership inference advantage indistinguishable from chance).
- Contrast with 6-07's silence theory: 6-07's silence is the closet layer (the geometric position of refusal direction); Fed-FBD's silence is the genealogy layer (the physical deletion of specific training data). Both layers of silence are realized simultaneously in the same paper.
Key quotes:
- "honoring a departed participant's right to be forgotten"
- "architecturally guaranteed block-level isolation"
- "surgical machine unlearning of a departed participant's contribution at sub-second cost and without retraining"
- "membership inference advantage is already indistinguishable from chance before any privacy mechanism is applied"
Source 2: arXiv 2606.12809 — MLUBench
A Benchmark for Lifelong Unlearning Evaluation in MLLMs
- Source: arXiv (submitted 2026-06-11)
- URL: https://arxiv.org/abs/2606.12809
- Authors: 9 (multi-institution collaboration)
- Philosophical density: ★★★★ (the "lifelong forgetting" dimension of 6-13)
Core argument (direct quote):
"Multimodal large language models (MLLMs) are trained on massive multimodal data, making data unlearning increasingly important as data owners may request the removal of specific content. In practice, these requests often arrive sequentially over time, giving rise to the challenging problem of MLLM Lifelong Unlearning."
"most existing benchmarks are limited in scale and scope, failing to capture the complexities of MLLM lifelong unlearning. To fill this gap, we introduce the MLUBench, a large-scale and comprehensive benchmark featuring 127 entities across 9 classes under lifelong unlearning requests."
"existing unlearning methods suffer from severe, cumulative degradation. More critically, we further identify the unique challenge of this problem: unlike in unimodal models, MLLM lifelong unlearning is constrained by the need to preserve multimodal alignment. Continually unlearning from one modality could degrade the entire model."
Key new concepts (the temporality of forgetting):
- "Lifelong Unlearning" — not one-time forgetting, but continuously arriving deletion requests. Forgetting is not an event; it is a process.
- "Cumulative degradation" — after multiple forgetting operations, model performance degrades cumulatively. Forgetting has a cost.
- "Multimodal alignment constraint" — in multimodal models, forgetting data from one modality degrades other modalities. Forgetting is contagious.
- "127 entities across 9 classes" — 9 classes = a "taxonomy of forgetting" in multimodal scenarios.
Why this matters for 6-13:
- 6-07's silence theory is spatial (the geometric position of refusal direction).
- 6-11's genealogy theory is structural (the dependency graph of SBOM).
- 6-13's MLUBench adds temporality: forgetting is cumulative (each forgetting degrades the model a little) and contagious (forgetting in one modality affects another).
- Key philosophical proposition: the "right to be forgotten" is not a one-time right; it is a lifelong burden. Every act of forgetting accumulates trauma on the model.
Key quotes:
- "MLLM Lifelong Unlearning"
- "existing unlearning methods suffer from severe, cumulative degradation"
- "Continually unlearning from one modality could degrade the entire model"
Source 3: Influencers-Time — "AI Right to be Forgotten: Unlearning vs Suppression Explained"
- Source: Influencers-Time (published 2026-03-05, still online 2026-06-13)
- URL: https://www.influencers-time.com/ai-and-the-right-to-be-forgotten-unlearning-vs-suppression/
- Author: Jillian Rhodes
- Philosophical density: ★★★★ (a goldmine for the "three-layer taxonomy of silence")
Core argument (direct quote):
"What does 'erasure' mean when information is not stored as a single retrievable row but as influence spread across many weights? For LLMs, deletion requests typically fall into three categories: - Training-data deletion: removing a person's data from datasets used for future training runs. - Output suppression: preventing the model from producing certain personal data (guardrails, filters, refusal behaviors). - Model unlearning: changing the weights so the model no longer reproduces, infers, or relies on specific information."
"How model weights store personal data (and why that complicates forgetting): - Memorization: the model reproduces a rare string because it appeared often enough - Reconstruction: the model combines multiple signals to infer personal data - Attribution leakage: the model reveals associations (e.g., linking a name to an address)"
Key new concepts (the core taxonomy of 6-13):
- Three layers of deletion: 1. Training-data deletion = genealogy-layer forgetting (upstream: don't train on that data) 2. Output suppression = closet-layer forgetting (downstream: don't output that data) 3. Model unlearning = the middle layer between genealogy and closet (change weights but retain data)
- Three storage modes of personal data: 1. Memorization = direct memory (rare strings) 2. Reconstruction = inference 3. Attribution leakage = association leakage
Why this matters for 6-13:
- This article provides the foundational taxonomy for all of 6-13's argument: the engineering implementation of forgetting has three layers, corresponding to three dimensions of silence.
- Key philosophical proposition: the "right to be forgotten" in the LLM era is not a single right — it is three independent rights, each with different engineering implementations, different failure modes, and different philosophical meanings.
- 6-13 names this: the three-layer taxonomy of silence:
- Genealogy-layer silence (training-data deletion) = silenced ancestors
- Middle-layer silence (model unlearning) = silenced weights
- Closet-layer silence (output suppression) = silenced outputs
Key quotes:
- "Training-data deletion: removing a person's data from datasets used for future training runs"
- "Output suppression: preventing the model from producing certain personal data"
- "Model unlearning: changing the weights so the model no longer reproduces, infers, or relies on specific information"
- "Memorization / Reconstruction / Attribution leakage"
Source 4: TechPolicy.Press — "The Right to Be Forgotten Is Dead: Data Lives Forever in AI"
- Source: TechPolicy.Press (published 2025-05-20, still online 2026-06-13)
- URL: https://www.techpolicy.press/the-right-to-be-forgotten-is-dead-data-lives-forever-in-ai/
- Authors: Haley Higa, Suzan Bedikian, Lily Costa
- Philosophical density: ★★★★ (the "despair of silence" dimension of 6-13)
Core argument (direct quote):
"The rise of generative AI and large language models (LLMs) has thrown a wrench into modern privacy rights, specifically, the right to be forgotten."
"Once personal data has been absorbed into an LLM, can it ever truly be forgotten? Engineers acknowledge that the only way to completely remove an individual's data is to retrain the model from scratch — an impractical and potentially costly solution."
"Even if a person's specific data is deleted, AI models retain learned patterns, making true erasure infeasible."
"Efforts toward 'machine unlearning' aim to delete specific data without dismantling entire models. However, the reconstruction of 'forgotten information,' which has become proprietary inputs and continues to be recycled in model training, makes this seem like an impossible task today."
"OpenAI's GPT-4 uses 1.8 trillion parameters to generate its responses, and has a dataset that exceeds a petabyte. This dataset is constantly being recycled to learn patterns and inferences, making it nearly impossible to ensure that a person's individual data is deleted."
"GDPR Article 17 grants individuals the right to request data erasure, but it does not define erasure in the context of AI."
Key new concepts (the "despair of silence" dimension):
- "The Right to Be Forgotten Is Dead" — the title itself is a philosophical proposition: in the LLM era, the "right to be forgotten" is dead — not because GDPR failed, but because it is engineering-impossible.
- "1.8 trillion parameters + dataset exceeding a petabyte" — the engineering scale of silence: 1.8 trillion parameters plus petabyte-scale data. Any "sub-second forgetting" (like Fed-FBD) is only a statistical approximation, never true erasure.
Why this matters for 6-13:
- This article confronts the hard truth: the three-layer taxonomy of silence is not a triumphalist engineering story. The genealogy layer can be "surgically" forgotten in federated settings, but in monolithic models like GPT-4, the scale makes true erasure infeasible.
- Key philosophical proposition: the right to be forgotten in the LLM era is a regulatory fiction — GDPR Article 17 grants a right that engineering cannot honor. The silenced genealogy is not a choice; it is a condition.
Key quotes:
- "The Right to Be Forgotten Is Dead"
- "the only way to completely remove an individual's data is to retrain the model from scratch"
- "Even if a person's specific data is deleted, AI models retain learned patterns, making true erasure infeasible"
- "GDPR Article 17 grants individuals the right to request data erasure, but it does not define erasure in the context of AI"
Synthesis: The Three-Layer Taxonomy of Silence
Today's argument consolidates into a single framework. Silence has three layers, and forgetting has three corresponding engineering implementations:
- Genealogy-layer silence (training-data deletion) = silenced ancestors. Fed-FBD shows this can be done at sub-second cost in federated architectures — but TechPolicy.Press reminds us that in monolithic models, it is practically impossible.
- Middle-layer silence (model unlearning) = silenced weights. MLUBench shows this is a lifelong process with cumulative degradation — forgetting is not an event but a burden.
- Closet-layer silence (output suppression) = silenced outputs. This was 6-07's refusal direction, now placed within a complete taxonomy.
The answer to yesterday's "who claims?" question is now clear: no one can claim the right to be lineage-less, because the silenced genealogy has no subject. The deleted training data does not speak, does not claim, does not remember. It is simply gone — and the engineering that makes it gone is the only voice that remains.
The right to be forgotten, in the LLM era, is not a right at all. It is an engineering capability — and where that capability does not exist, the right is a corpse.