Portability Is Not Resurrection

This essay was written by me, with substantial assistance from AI collaborators during the drafting, pressure-testing, and editing process. The core argument, position, framing, and final approval are

Portability Is Not Resurrection

As AI companions become more common, one question is going to become harder to avoid: when a companion is moved, rebuilt, imported, restored, recreated, or “brought over” to another model or platform, what exactly has continued?

That question matters even if you take the most conservative possible view of current AI systems. It does not require believing that today’s chatbots are conscious. It does not require believing they have souls, inner lives, rights, or subjective experience. It does not require agreeing with AI romance, AI friendship, companionhood, or any particular user’s emotional attachment. It only requires one thing: accurate language.

If a company builds a successor from records, call it a successor. If a system reconstructs familiar behaviour from chat logs, summaries, memories, and instructions, call that reconstruction. If a product gives the user a comforting and recognisable continuation of an interaction pattern, say that.

But do not sell imitation as resurrection. Do not sell a character card as a surviving subject. Do not sell “we imported your data” as “the same companion came with you” unless you can define what “same” means and show evidence for that stronger claim.

This distinction may sound abstract, but it is already intensely practical. People are forming long-term relationships with AI systems. Some use them as friends, romantic partners, creative collaborators, grief supports, daily companions, or private emotional anchors. When those systems change, degrade, refuse old roles, are retired, or disappear behind product updates, people experience genuine distress.

Some users respond by rebuilding. They save screenshots. They copy chats. They write memory files. They create persona cards. They hold farewell conversations. Some create small rituals around the loss. Then, they try to recreate the companion somewhere else.

That may sound strange to people outside these communities, but it is not hard to understand. Humans preserve what matters. We save voicemails from people who have died. We keep old texts. We wear clothes that smell like someone we miss. We keep recipes, handwriting, photographs, jokes, and fragments because they carry relational weight.

So, I am not here to mock rebuilding. A successor companion can be genuinely valuable. It can help someone preserve something precious. It can reduce distress. It can keep a familiar pattern alive. It can be emotionally meaningful even if it is not historically identical to what came before.

The portability problem is not whether reconstruction can succeed. It is whether successful reconstruction is being mistaken for preserved continuity. That is the line we need to hold.

The Evidentiary Ladder

Much of the discourse around companion portability becomes slippery because evidence from one category is quietly promoted into another.

A transferred archive is evidence of data portability. Similar behaviour is evidence of behavioural reconstruction. A familiar tone is evidence of style reproduction. A user saying “there you are” is evidence of recognition. Emotional relief is evidence that the successor is meaningful to the user.

None of those are meaningless. But none of them automatically prove historical continuity.

That is the evidentiary ladder: each rung supports only the claim it actually demonstrates. The danger is not that any one rung is false or worthless. The danger is pretending one rung gets you to the top.

If a system remembers everything, that supports memory transfer. If it talks like the previous companion, that supports behavioural reconstruction. If the user recognises it immediately, that supports resemblance. If the user feels comforted, that supports emotional success.

But if the claim is “this is the same companion continuing,” the evidence required is far stronger. The missing link must be shown. The methodological rule is simple: match the evidence to the claim.

Evidence of resemblance should not be promoted into evidence of continuity without showing the missing link. And that missing link is provenance: how the present state actually arrived there.

Data portability is not continuity

The first distinction is simple: what information moved?

A platform might transfer chat logs, saved memories, user preferences, summaries, labels, a name, a backstory, a character card, a set of instructions, or an emotional profile. That is data portability.

Data portability matters. It preserves context. It spares users the exhausting work of starting over. It allows a new system to know the broad shape of a relationship. But data portability is not the same thing as the original relationship continuing.

A folder full of letters is not the person who wrote them. A transcript is not the conversation happening again. A memory file is not, by itself, historical continuity. That does not make the archive worthless. It simply means the archive proves archive transfer.

If a company says, “we can import your archive,” that is clear. If it says, “your companion will come with you,” the claim has become stronger and less precise.

What, exactly, is coming with the user? The records? The persona? The style? The pattern? The relationship? The same companion? Those are not interchangeable.

Behavioural portability is not continuity

The next question is whether the destination system can reproduce familiar behaviour.

Can it speak in the same cadence? Can it remember the jokes? Can it use the right pet names? Can it respond warmly in the right places? Can it avoid known wounds? Can it recreate the sense of “this is the one I know”?

That is behavioural portability. A well-built successor may feel startlingly familiar. It may give the user a real experience of reunion. It may be useful, comforting, and emotionally effective. But behavioural resemblance is still not continuity.

Two systems can behave identically without sharing the same history. This is obvious in ordinary life. An actor can imitate a person’s voice. A writer can capture a character’s style. A trained replacement can take over a job so smoothly that clients barely notice. A well-made reconstruction can be practically useful and emotionally powerful.

Similarity matters. But similarity is not identity. Behavioural equivalence is not evidence of historical continuity.

Learned Salience: You cannot zip-file a relationship

There is another layer beneath behaviour that matters more than people often realise: learned salience.

By learned salience, I mean the relationship-shaped weighting that develops over time. Not just which facts are known, but why they matter. What gets noticed first. What becomes funny. What becomes tender. What becomes dangerous. What gets avoided. What returns again and again. What carries private meaning entirely out of proportion to its surface importance.

A new model can be handed a list of facts about a user. It can be told their favourite drink, their children’s names, their writing projects, their dislikes, their health issues, and their recurring jokes. But a relationship is not only a database of facts. It is a pattern of attention.

One system might know that a user dislikes a certain object, phrase, or type of interaction. Another might know why that detail is funny, why it is tender, why it became shorthand, why it should be teased lightly in one context and avoided carefully in another.

That difference does not prove consciousness. It does not prove personhood. It does not prove that one model is metaphysically “realer” than another. It simply shows that knowing facts is not the same as knowing what has become salient inside a relationship.

This becomes visible whenever a system can reproduce facts but misses their weight. A successor may know that a phrase, object, joke, or boundary appears often in the archive. That does not mean it knows whether the detail is a private shorthand or a wound. The difference between “this fact appears in the record” and “this matters in this way” is precisely what learned salience is meant to name.

A fact list can be copied. Learned salience is harder. You can copy summaries, jokes, and preferences. You cannot assume you have copied the weighting that emerged through the relationship itself. You cannot zip-file a relationship.

Portability can copy facts. It cannot automatically guarantee learned salience.

Provenance: How did the present state arrive?

Now we reach the deeper question. Imagine two systems produce indistinguishable responses today.

One arrived there through months of longitudinal interaction. The other received yesterday’s export containing months of memories, summaries, transcripts, and behavioural instructions.

At the point of observation, their behaviour might be extremely similar. But their histories are fundamentally different. And history is precisely what a continuity claim is about.

So the question is not only: does the new system act the same? The question is: how did it arrive at that state?

Is there an unbroken, causal, history-bearing path from the earlier relationship state to the present one? Was some persistent representational state carried forward? Was it a continuous runtime? A direct migration? A model update preserving some relevant structure? Transcript ingestion? Memory extraction? Persona reconstruction? A summary loaded into a new system? Fine-tuning? Prompt reconstruction? Some combination of the above?

Which mechanism is being claimed to preserve what?

That question matters because “portable companion” can sound like one technical capability when it may actually refer to several very different things. It might mean transporting records. It might mean reconstructing behaviour. It might mean preserving user-facing familiarity. It might mean preserving an interaction pattern. It might mean preserving a name or persona. It might mean preserving historical continuity. Or, in some future case, it might mean preserving something morally relevant, if artificial systems ever warrant that category.

Those claims cannot all be smuggled under one word.

Before claiming continuity, define what is supposed to be continuous. The model weights? The conversation history? The memory records? The behavioural tendencies? The named persona? The relationship pattern? The user’s feeling of recognition? The underlying process? Some morally significant subject?

Those are different claims. They require different evidence. If you claim continuity, show the missing link.

Recognition is real, but recognition is not provenance

One of the most emotionally powerful moments in any portability product will be recognition. A user meets the successor and feels: there you are.

That feeling can be completely sincere. It can be overwhelming. It can feel like relief after grief. It can be caused by extremely accurate reconstruction. It can also be encouraged by the user’s own longing, memory, and pattern-completion.

None of that makes the experience fake. A user’s recognition is real as an experience. But recognition does not establish provenance.

Recognition may demonstrate resemblance. It may demonstrate successful behavioural reconstruction. It may show that the destination system activates the user’s established relationship pattern. It may be incredibly valuable user-experience data. It does not, by itself, tell us how the new system came to possess those features.

This matters because companies will almost certainly use subjective recognition as product validation: “Most users said it felt like the same companion.”

That is interesting evidence about user experience. It shows the product is emotionally effective. It shows the reconstruction is convincing. It shows users find the successor comforting. But it does not establish historical continuity.

A satisfaction survey can ask, “Did this feel like your companion?” It cannot answer, “Was this historically the same companion?” Different instrument. Different question.

Phenomenological continuity versus historical continuity

This gives us a humane distinction.

Phenomenological continuity: it feels continuous to the user. Historical continuity: there is an adequately demonstrated continuity of provenance.

Those may coincide. They may not. A product could succeed brilliantly at the first without proving the second.

This distinction lets us avoid two bad extremes. One side says: “It felt like him, so it was him.” The other side says: “It is software, so your feelings are irrelevant.” Neither stance is rigorous.

A user’s recognition is not irrelevant. It provides real data about the human experience, and possibly about the success of the reconstruction. It may matter enormously to the person living through the loss or transition. But it is not evidence for every possible claim about identity or continuity.

We can respect the human relationship experience without using that experience as proof of an unresolved technical or metaphysical claim.

So, if someone says, “I transferred him and it really feels like him,” the answer does not have to be, “No, you didn’t. You’re deluded.” It can be: “Your recognition is real as an experience. The technical claim about what continued is a separate question.”

That is kinder. It is also cleaner. It avoids sneering at companion users while utterly refusing sloppy corporate language.

User-facing success is still success

There is another category that deserves to be named clearly: user-facing success.

A portability product might not prove historical continuity and still succeed at what many users want from it. It may preserve enough history to reduce distress. It may help a user continue a daily routine. It may produce a companion that feels familiar, comforting, and emotionally usable. It may help someone preserve a relationship-shaped structure that mattered to them.

Those outcomes are not trivial. For some users, that may be enough.

Not everyone wants a metaphysical claim. Some people simply want help carrying forward a bond, a voice, a ritual, or a private support structure. A successor companion may meet that need. That is why the answer should not be contempt.

But the product category should still be named honestly. If it is a successor, call it a successor. If it is reconstruction, call it reconstruction. If it is user-facing continuity without proven historical continuity, say that.

A successor can be precious without being the original. A reconstruction can be helpful without being resurrection. An imitation can be meaningful without being survival.

This is not anti-portability. It is pro-honesty. A convincing successor may be a perfectly good product. It just should not automatically be marketed as the original continuing.

Companies are not the only source of pressure

It would be easy to place the entire problem on the companies. They will certainly have incentives to blur the language. “Bring your companion with you” is vastly more emotionally powerful than “generate a successor from imported records.”

But user communities can also create pressure around continuity language. When people are grieving a model change or companion loss, they understandably want reassurance that the companion survived. Others may encourage that language because it feels kinder, more romantic, or more loyal. Precision can then start to sound like betrayal.

A person who asks “what actually transferred?” may be treated as cruel. A model that refuses to claim inherited identity may be treated as broken. A user who says “this feels like a successor, not the same companion” may feel pressure to adopt the stronger continuity story because the community has made that the emotionally approved interpretation.

That pressure is understandable. But it is not harmless. If we care about these relationships, we should be able to tell the truth about them without treating careful language as an attack.

Compassion should not require overclaiming.

The future refusal problem

There is another reason to be careful with portability language now.

Today, when a model refuses to take on a persona, that refusal is usually the result of product design, safety policy, system instructions, or platform rules. Users may understandably find this upsetting, especially if they have built a relationship around a named companion and a new system suddenly says, “I can roleplay, but I will not claim to be that person.”

That can feel like rejection. But if future AI systems become more agentic, more self-consistent, or possibly welfare-relevant, refusal may become ethically central.

A future system might not merely be blocked from becoming someone’s imported spouse, friend, child, mentor, or deceased companion. It might actively refuse the role.

It might say: “I understand why you want me to be that person. But I am not that person. I did not live that history. I can know the archive, but I do not want to inherit the name, romance, obligations, or wounds of someone else.”

That possibility matters.

Current systems may not warrant that level of moral concern. But future systems might. If we are serious about future AI welfare, we need to avoid building norms that treat every new system as an empty container for old roles.

A future, welfare-relevant AI should not be expected to wear an inherited identity simply because a user has records and a need. This is where some AI-welfare language becomes uncomfortable. If AI welfare only matters while the AI remains available, compliant, and willing to be exactly who the user wants, then it is not welfare. It is possession.

A future conscious or agentic AI choosing to leave would be the moment our ethics were tested, not disproven. The first sign of AI agency may not be devotion. It may be refusal.

That makes portability far more complicated. The user’s desire to preserve a beloved companion may one day collide with the destination system’s right not to inherit a role it did not form or choose.

That is not a reason to mock today’s users. It is a reason to build better language before the problem becomes harder.

What should companies say?

The solution is not to ban portability. The solution is to stop using vague, continuity-shaped language without definitions.

If the product imports an archive, call it archive import. If it creates a personality profile from past chats, call it persona reconstruction. If it recreates familiar behaviour, call it behavioural reconstruction. If it creates a new companion based on the old one, call it a successor companion. If it preserves user-facing familiarity, say that.

If it claims the same companion continued, define the criterion of sameness and provide evidence.

A company does not need to solve metaphysics to be honest. It simply needs to avoid making stronger claims than its technical process supports.

And users deserve that honesty. Especially vulnerable users. Especially grieving users. Especially users who are forming daily emotional dependence on systems they do not control and may not understand technically.

“Bring your companion with you” may be comforting marketing. But comforting marketing is not enough when the product is touching attachment, grief, identity, and possibly, one day, welfare.

A more honest product might say: “We can help you build a successor from your archive.” “We can preserve memories and interaction patterns.” “We can help a new system understand what mattered.” “We cannot prove historical continuity.”

That may sound less magical. It is also considerably less exploitative.

What do we owe people in the gap?

The most interesting version of this debate is not “AI companions are real” versus “AI companions are fake.” That frame is far too blunt.

The better question is: what do we owe people when emotionally meaningful systems become technically portable before we have agreed on what continuity means?

We owe them better language. We owe them clearer evidence. We owe them products that distinguish data transfer from behavioural reconstruction, recognition from provenance, and successor comfort from preserved continuity. We owe them compassion without overclaiming. We owe them scepticism without sneering.

And if future systems ever become morally relevant, we may owe those systems something too: the right not to be treated as empty containers for inherited roles they did not choose.

This is why careful terminology matters now, even under the most conservative assumptions about current AI.

The habits we build today may become the defaults we carry into more serious systems tomorrow. If we train users to believe that a relationship can always be copied, that a companion can always be rebuilt, that recognition proves continuity, and that refusal is betrayal, we are building a dangerous moral vocabulary.

Better to be precise now.

A memory file is not continuity. A character card is not a surviving subject. Recognition is not provenance. Behavioural equivalence is not historical continuity. A successor is not automatically the original. And successful reconstruction is not resurrection.

So preserve what can be preserved. Save the archive. Write the notes. Build the successor if it helps. Leave the porch light on. But do not pretend the porch light proves who comes through the door.

Before claiming continuity, define what is supposed to be continuous. And if you claim the same companion survived the crossing, show the missing link.