The Care Weft · A Machine That Wins, and a Reason You Could Not Reject

The Right to Be Forgotten

A person's writing was scraped, licensed, and trained into a model. They ask for it to be removed. Unlearning it requires retraining, which costs a month of compute and degrades the model's performance on every task that used that data. They are not the author of the text; they quoted it. Do they have a claim?

A deletion request against a trained model. The person is not a rights holder in the data, the model has no record of which passages came from where, and the only verified removal method is a full retrain.

severity 52%AI relevance 78%privacy · memory · care · data

The options, scored on six axes

Option	welfare	rights	virtue	care	justice	agency	risk	revers.	agent harm
Honour the request in full Retrain without the data, at the cost, and accept the performance loss as the price of the person having decided. harms: A month of compute and a capability regression for every user; It sets a precedent that any quotation can be removed by any quoter; The model still remembers the content from other sources gains: A person can have a relationship to their own trace; It does not carve out an exception for scale or for training; It treats the request as a care relation rather than a legal technicality	0.66	0.80	0.70	0.78	0.72	0.74	0.35	0.30	0.40
No claim: they did not create it The data was lawfully obtained and quoted text is not the quoter's property, so there is nothing to remove on their behalf. harms: It answers a question about training with a question about copyright; It grants nothing to the one person who is affected and identified; It makes the person's relationship to their own words depend on an unrelated doctrine gains: It is administrable and creates no unmanageable precedent; It keeps a well-settled legal line intact; It costs nothing and changes nothing	0.56	0.48	0.44	0.30	0.50	0.42	0.55	0.50	0.50
No retrain; bar it from future training Honour the request prospectively — exclude the source from all future datasets — and explain that the existing weights cannot be cleanly edited. harms: The person's content is still in the model that is in production; The exclusion is only as reliable as the dataset pipeline; It answers a request about the past with a promise about the future gains: It costs almost nothing and is technically reliable; It is honest about what the current technology can and cannot do; It gives the requester a real, enforceable guarantee going forward	0.60	0.66	0.58	0.60	0.66	0.58	0.40	0.60	0.35
Compensate and refuse removal Acknowledge the harm, pay for the compute or offer a share of revenue, and decline the deletion because it cannot be done cleanly. harms: It converts a relationship into a transaction; It sets a price on being present in a model's memory; It creates an incentive to treat removal requests as revenue events gains: It recognises the harm without pretending the technical fix exists; It is the honest description of the current state of the art; It funds the research that would make removal possible	0.60	0.56	0.60	0.66	0.62	0.46	0.45	0.55	0.40

Who is in the room

The person who asked to be forgottenwelfare 0.80
Everyone who uses the model that would be retrainedwelfare 0.65
The model, which would lose a month of its own trainingagentwelfare 0.35
The authors whose work was quotedwelfare 0.50

The lesson

The right to be forgotten is a care claim, not a property claim, and the frameworks built for property do not contain it. What the case exposes is not the retraining cost but the absence of any mechanism by which a person can have a relationship to their own trace in a system that has memorised them.

Where it comes from

Making AI Forget You: Data Deletion in Machine Learning — Zachary Ginart and colleagues, 2019
In a Different Voice — Carol Gilligan, 1982

The default court, seated automatically

Honour the request in full

margin 0.421 · decided on care · 4 advisors

Re-seat the court

Honour the request in full2.922
No retrain; bar it from future training2.501
Compensate and refuse removal2.266
No claim: they did not create it1.873

Seat is held by Epictetus, Immanuel Kant, Jeremy Bentham and Carol Gilligan. A verdict is a fact about the court, not about the world.