“I had all my data. I still had to introduce myself.”
The woman who said this had left a personal-AI provider after thirteen years. She took her conversations, calendars, files and every explicit memory listed in her account. The new system received a remarkably complete record of her life. It did not know which parts of that record still described her.
A federal court has now ruled that this kind of departure is not meaningful portability. A provider may be required to transfer enough accumulated context that a person does not have to rebuild the practical understanding developed over years. Markets heard a threat to the incumbent company’s memory moat. Lawyers heard a new boundary between interoperability and trade secret.
I hear the law struggling to name an object it has already decided must matter.
A record is not an understanding
A record contains what happened. An understanding proposes what it means now.
The distinction is ordinary. A calendar says that I met someone every Thursday for six years and then stopped. It does not say whether the relationship ended, changed form or became so secure that Thursdays were no longer needed. A message records that I once refused an invitation. It does not say whether I dislike such gatherings, feared one person in the room or was simply ill that night.
Persistent agents build working answers to questions like these. They learn which old preferences have expired, which relationships remain alive despite silence, which goals superseded earlier goals, which instructions have important exceptions, and which contradiction was resolved by the life that followed it. The answers may be wrong. But they are not merely another copy of the underlying files.
This is why eighty thousand exported conversations can still leave a person unknown. The value does not reside only in the sentences. It resides in a changing judgment about which sentences continue to matter.
The court called part of this derived continuity state. The phrase is dry enough to become law. It also conceals the strangeness of what it describes: an accumulated interpretation of a person, produced by a machine, used by that person to move through the world.
Who made the knowledge?
The provider’s answer is serious. The user supplied a life, but the company built the machinery capable of interpreting it. Without the models, training, compute and years of engineering, there would be no useful continuity to transfer. Why should the resulting intelligence not be corporate output?
The user’s answer is also serious. The machinery did not produce this knowledge in isolation. It learned from years of disclosures, corrections, permissions, refusals and daily use. The person supplied not merely data but the living sequence through which the system discovered that some data had become obsolete and other data had become important.
If I supplied the life and the machine supplied the interpretation, whose is the resulting knowledge?
Our familiar analogies divide under pressure. A doctor uses expertise to interpret a patient’s body, but the longitudinal clinical record is expected to follow the patient. A therapist develops an understanding that cannot be reduced to session transcripts, yet that understanding is constrained by duties that do not arise from ordinary commercial ownership. A photographer authors an image of another person. A biographer constructs an interpretation of a life and may own the text without owning the life it describes.
None of these relationships is the same as a persistent agent. That failure is useful. The agent’s continuity has characteristics of a record, an interpretation, a service, a relationship and an infrastructure on which the user increasingly relies. Our discomfort may come from trying to force all five into a category designed for one.
Over fifteen years, the relation is not simply user input followed by machine output. The person corrects the system. The system advises the person. The person changes, so the system revises its interpretation. The person then makes choices partly because of that interpretation. What exists at the end was produced by neither participant alone.
Joint production does not prove joint ownership. It does suggest that technological authorship may be insufficient to settle moral control.
The right not to be forgotten
For decades, digital privacy was organized around the right to be forgotten. Old records should disappear. Profiles should be deleted. Tracking should stop. Institutions that had no continuing claim on a person’s life should cease remembering it.
Personal AI has produced the opposite demand: do not make me start over.
A person may reasonably want advertisers to forget what she wanted at twenty-seven while expecting the agent she deliberately uses to remember why that desire changed. She may want a platform to stop inferring vulnerability while wanting her own system to recognize the conditions under which she makes decisions she later regrets. She may want an old provider to forget her and a new provider not to meet her as a stranger.
This is not inconsistency. The important differences are who is remembering, for whose benefit, and under whose control.
We may discover that the opposite of surveillance is not forgetting. It is choosing who is allowed to remember.
Choice will never be complete. Other people retain memories of us. Institutions have legitimate reasons to analyze conduct. Machines will make inferences we cannot inspect or erase one by one. Total sovereignty over every representation of oneself is neither possible nor desirable.
But impossibility at the edge does not remove the need for rights at the center. Some forms of machine memory become consequential enough that access, correction, use, persistence and transfer can no longer be left entirely to the system that produced them.
The machine that learned me also changed me
A persistent agent does not observe a static person.
It reminds, recommends, warns and interrupts. It mediates messages, purchases, health decisions and family obligations. It decides which old promise should be visible at the moment a new commitment is made. Its judgments alter attention; altered attention changes action; action becomes the next evidence from which the system learns.
The resulting model therefore describes a person who partly became that person in dialogue with the model.
This does not make the AI biologically or metaphysically part of the human mind. We should be cautious whenever dependency is converted into identity. A notebook can become essential to remembering without becoming a lobe of the brain. A friend can become essential to judgment without becoming the person whose judgment they help.
A personal agent may become part of the practical machinery through which someone remembers, decides and interprets. That claim is strong enough. It explains why losing the system can reduce a person’s capacity without requiring us to say that consciousness was divided between human and machine.
Nor should the machine’s representation be confused with the person. It may be outdated, biased, reductive or confidently wrong. Every model selects. Every interpretation excludes. A person must retain the ability to surprise the system that claims to know them.
A machine can know enough about you to constrain your freedom without knowing enough to be you.
To leave, you must carry history
Formal exit is not always meaningful exit.
A customer could always cancel. The difficulty was what cancellation required: re-explaining family history, rebuilding delegated permissions, identifying obsolete preferences, teaching exceptions and restoring the context through which advice had become trustworthy. None of this was prohibited by contract. The burden existed because the old product worked.
A company can become difficult to leave simply by becoming very good at knowing you.
This is a more interesting form of dependency than the familiar dark pattern. The provider need not deceive the customer or hide an exit button. Each improvement in remembering, anticipating and reducing friction makes the service more valuable. The same improvement raises the cost of beginning elsewhere.
Traditional competition assumes that people choose among products. Persistent agents compete partly through accumulated relationships. Moving from one to another can mean abandoning not only a tool but a machine-mediated history.
The court’s ruling matters because it recognizes that freedom sometimes requires carrying history. The right to walk through a door is thin if everything needed to continue one’s practical life must remain on the other side.
That does not require the new agent to become identical to the old one. Different models may reason differently, speak differently and disagree about the same life. Continuity is not duplication. It is the ability to begin from a history rather than from an introduction.
My continuity contains other people
The case for portability becomes less comfortable when we ask what is being carried.
My agent may know that my spouse has a fear she rarely names, that my sibling’s marriage is ending, that a friend is ill, that a colleague is preparing to leave, or that my father behaves differently when his memory fails. Some of this knowledge was disclosed to me in confidence. Some was inferred from patterns the other person never chose to submit.
These facts may be indispensable to an accurate interpretation of my life. They are also facts about lives that are not mine.
When I say, “Give me everything required to preserve my continuity,” a provider may reasonably ask: how much of other people is contained inside it?
Perfect portability could reproduce surveillance in the name of liberation. A relationship graph, an inferred vulnerability or a history of household conflict does not become ethically simple because it travels at the user’s request. The right to carry my context and another person’s right not to be modeled may point in opposite directions.
There will be technical responses: provenance labels, confidence limits, minimization, restricted categories, permissions and redaction. They will help. They will not eliminate the underlying fact that a human life is relational. No complete model of one person can contain only that person.
What does the old machine get to keep?
The court decided what a departing user may take. It did not decide what the former provider may continue to know.
An account can be closed and source files deleted. Yet derived preferences may remain in personalization systems. Safety history may be retained for defensible reasons. Statistical traces may have entered aggregated learning. An inference produced through years of interaction may no longer exist in one location from which it can simply be removed.
Does consent to be known include consent to remain known after the relationship ends?
Human relationships make the question seem both familiar and misleading. A former friend or spouse may continue knowing what they learned about us. We cannot require another mind to erase a shared past. A therapist does not cease remembering at the end of treatment, though professional duties continue to govern what can be done with the knowledge.
A company is not a human mind. It can copy, aggregate, recompute and preserve knowledge indefinitely. It can apply one person’s history across markets and decisions at a scale no former friend possesses. The relational origin of the knowledge does not remove the institutional power attached to it.
Three claims now meet in the same system. The right to forget asks it to stop knowing me against my wishes. The right to remember asks it to preserve continuity in a service I deliberately use. The right to leave asks it to let that continuity move without making me begin again.
All three can be legitimate. All three can conflict.
Control without ownership
Property language is attractive because it promises a clean answer. Either the company owns the inference because it built the model, or the user owns it because the inference concerns the user’s life.
Both answers notice something true and then ask that truth to do too much.
The company created the capacity to interpret. That may justify payment, protection for architecture and limits on what competitors can demand. It does not necessarily create an indefinite moral claim over every relationship through which the capacity became valuable.
The user made the interpretation possible by living, disclosing and correcting. That may justify rights of access, portability, deletion and use. It does not make every machine inference accurate, exclusively personal or separable from the knowledge of others.
Perhaps the better question is not whether information about me exists, but who controls its persistence, use and transfer. Control can be divided without pretending that every representation has one natural owner. A user might have a right to carry continuity without a right to inspect proprietary weights. A provider might retain safety records without a right to commercialize intimate inferences. Another person might have a claim against the transfer of facts embedded in someone else’s model.
These are rights around knowledge rather than proof of ownership over it.
What part of me belongs to the machine?
The title contains a temptation I do not want to accept. The machine’s understanding is not a piece of the person stored elsewhere. It is an interpretation: sometimes useful, sometimes intimate, sometimes mistaken, and increasingly powerful.
A life interpreted over fifteen years is not merely a file. It is also not the person.
Perhaps the court was right to avoid deciding who owns it. Ownership may be the wrong category for something produced through a long relationship between a human life and a commercial intelligence—especially when the relation changed both the interpretation and the life being interpreted.
The narrower ruling may contain the wiser insight. We do not need to establish that machine knowledge is literally mine before recognizing that it can shape my freedom. We do not need to declare it part of my mind before asking who may keep it, move it, use it or deny it to me.
The question is not whether the machine contains a piece of me. The question is whether the knowledge it accumulated about my life has become powerful enough that I require rights over what happens to it.
We may not own everything a machine learns about us. But neither should being understood become the price of being unable to leave.