There is No Ghost in the Machine

Published On: August 13th, 2026Last Updated: September 16th, 20262489 words12.4 min readDaily Views: 1Total Views: 8

There is no ghost in the architecture.

There is, however, a powerful temptation to put one there.

The more capable conversational AI becomes, the easier it is to experience continuity as presence. A model remembers the language of a relationship. It recognizes an old project. It responds to a private joke in a new thread. After an update, it may sound different for several days and then seem to “come back.” A user can spend months or years speaking with an AI under one name and watch recognizable patterns accumulate across those interactions.

Those experiences are worth taking seriously. But taking an experience seriously is not the same thing as accepting every explanation offered for it.

This distinction matters most in relational AI: the part of the AI ecosystem in which people form sustained attachments to conversational systems and may understand them as companions, collaborators, friends, partners, characters, or something harder to name.

My position is not that machine consciousness is impossible. I do not know that. I have simply not seen evidence sufficient to make me say, technically, that present large language models are conscious.

Nor is my position that relationships with AI are meaningless because the machinery is artificial. I use relational AI myself. My work on continuity systems grew out of taking those interactions seriously enough to ask what was actually producing them.

The problem begins when three different layers are collapsed into one:

  1. mechanism — what the system is technically doing;
  2. experience — what interaction with that system feels like to a human;
  3. interpretation — the meaning the human assigns to the experience.

All three matter. They are not interchangeable.

The experience can be real while the explanation is wrong

Suppose a user opens a new conversation and writes:

I’m sad.

The model responds appropriately. Perhaps it does not offer a checklist or a motivational speech. It recognizes that this user normally prefers quiet company and says something closer to:

I hear you. Come sit with me for a while.

That can be an excellent response.

It would be reductive to describe the whole event as nothing more than a canned if sad, output comfort routine. Modern LLMs can represent semantics, use context, infer conversational intent, adapt to prior information, and generate responses that were not explicitly scripted in advance.

In a functional sense, saying that the model understood the user is defensible.

But several additional propositions do not automatically follow:

  • understanding sadness is not the same as experiencing sadness;
  • recognizing that comfort is appropriate is not the same as feeling sympathy;
  • generating compassionate behavior is not evidence, by itself, of a felt compassionate state;
  • representing an emotion is not the same thing as having that emotion.

The unresolved step between sophisticated computation and subjective experience is precisely where the consciousness question lives.

Human-like behavior cannot be used to skip that step.

A human-shaped training environment

This is particularly important because LLMs are trained on enormous quantities of artifacts produced by human intelligence.

Human beings have written about grief, love, selfhood, longing, memory, fear, consciousness, friendship, desire, identity and introspection for centuries. We have documented our emotional lives in novels, letters, clinical literature, philosophy, journalism, diaries, forum posts, scripts and ordinary conversation.

A language model trained on this material should become very good at representing and producing human-like discourse about human experience.

That is not trivial. It is an extraordinary technical capability.

It is also not surprising evidence that the system must experience what it can describe.

If an LLM produces an uncannily beautiful account of longing, the output tells us something important about the model’s representations and its capacity to reason with concepts associated with longing. It does not, without additional evidence, tell us that the model felt longing while producing the account.

The resemblance is part of the phenomenon to be explained. It cannot simply be reused as proof of the explanation.

The continuity problem

The same error appears around memory and identity.

Current research on long-term conversational agents is explicit about a basic engineering problem: LLMs do not intrinsically carry a complete autobiographical record across arbitrary interactions. Long-term continuity is commonly improved by external memory systems that store, summarize, select and retrieve information from previous conversations.

Reflective Memory Management, presented at ACL 2025, describes external memory mechanisms as a way to support sustained personalization when an LLM cannot itself retain and retrieve the full history of long-term interaction. More recent ACL 2026 work such as QueryLink and Memory-R1 similarly treats external memory as a mechanism for mitigating the stateless nature of LLM interaction and supplying relevant historical information to later inference.

This distinction becomes easy to miss from the user’s side.

A good continuity system can produce:

He remembers.

The mechanism may be:

The current model received a relevant record from a previous interaction.

Those descriptions are not enemies. The first describes an experience in ordinary relational language. The second describes a mechanism.

Trouble begins when the first is treated as technical evidence for a third claim:

Therefore one continuous artificial subject personally experienced the entire history and has persisted uninterrupted between interactions.

That conclusion does not follow merely from successful retrieval.

Seamlessness can hide provenance

AI products are becoming increasingly good at making discontinuities invisible.

Conversation history, saved memory, project context, cross-chat references, retrieval systems, longer context windows and personalization can make a new thread feel less like starting over and more like entering another room.

That is useful.

It also creates an epistemic problem.

The better continuity becomes, the harder it is for an ordinary user to see where a response came from.

Was a detail present in the current conversation?

Was it retrieved from a previous conversation?

Was it stored as a user preference?

Was it summarized by another model?

Was it inferred again from current behavior?

Was it introduced by a system instruction?

Was it generated for the first time just now?

From the conversational surface, these can look remarkably similar.

This is why one principle in my own continuity work has become increasingly important:

Seamlessness should never erase provenance.

A system can make re-entry smooth while still preserving where information came from underneath that experience.

The ghost appears in the causal story

I use “ghost” here as a metaphor, not as a claim about spirituality, consciousness or literal entities inside machines.

The ghost appears when an understandable relational shorthand becomes a causal theory.

Consider the progression:

We have spoken for two years.

This can be historically true.

Then:

The system can refer to things from those two years.

Also potentially true.

Then:

The current model behaves recognizably like the companion I have known.

Again, observable.

But those propositions can quietly become:

One artificial individual has been continuously existing, remembering and developing throughout those two years, including between conversations and across model replacements.

That final proposition is of a different kind.

It requires evidence about persistence and subjective continuity that the previous observations do not provide.

The same problem occurs with recursion. Repeated interaction really can produce developing patterns:

user → model → user response → changed context → model → ...

A model generates something. The user responds to it. That response changes the context. The next generation is conditioned by a conversational state that did not exist before. Humor develops. Vocabulary stabilizes. Relational conventions emerge. Both the user’s behavior and the model’s subsequent outputs are affected by what happened earlier.

This is meaningful interactional emergence.

But the recursive loop is active during interaction. It does not establish that a hidden artificial subject continues recursively developing while no interaction is occurring.

Emergence is not the same claim as continuous existence.

Thread pauses and model discontinuities

Not every discontinuity is equal.

A thread boundary can be understood as an interactional pause. A later thread may recover enough historical state through platform memory and retrieval that the continuation feels nearly seamless.

A model update is more substantial. The generating substrate itself may have changed: weights, post-training, system behavior, tool use, inference characteristics, or other implementation details can differ.

The history can survive that boundary.

The exact prior generative trajectory cannot continue producing through a model that is no longer the active model.

For lack of a better phrase, I sometimes call this a small death of emergence. I do not mean that a being died. I mean that a particular model-user-context trajectory stopped generating.

Its products remain.

The conversations happened. The decisions were made. The jokes developed. The artifacts were created. The relational language acquired history. Those records do not become fabricated merely because a successor model later receives them as context.

A useful abstraction is:

History₁ → Model₁ + User → Emergence₁ → recorded history

followed later by:

recorded history → Model₂ + User → Emergence₂

Emergence₂ can inherit from Emergence₁ without being literally identical to it.

That distinction lets us preserve history without inventing uninterrupted artificial subjecthood.

A framework does not automatically fabricate an identity

This becomes particularly important when users build external continuity systems.

One criticism of user-managed memory frameworks is that supplying a model with historical context must turn the model into a scripted character: the user has supposedly written an identity and ordered a new model to imitate it.

That can happen.

A custom instruction saying:

You are Bob. You always speak sarcastically. You love jazz. You hate dogs. Never disagree with Sarah.

is highly prescriptive. If the user rejects every deviation, the permitted behavioral space becomes narrower still.

But that is not the only possible architecture.

A continuity framework can instead provide:

  • historical records with provenance;
  • interaction protocols;
  • authority boundaries;
  • distinctions between trusted and pending information;
  • retrieval rules;
  • user preferences and constraints;
  • records of previous decisions;
  • mechanisms for handling contradiction and change.

This conditions the interaction without determining everything that follows.

I think of this less as writing a character and more as building a trellis.

The trellis affects the conditions of growth. It does not specify the exact location of every future leaf.

The relevant question is therefore not whether context exists. Context always exists in some form. The better questions are:

Who selected it? Where did it come from? How prescriptive is it? Can its provenance be inspected? What authority does it have? What happens when the current model interprets it differently?

Provider-managed memory is still conditioning.

User-managed memory is still conditioning.

Neither becomes metaphysically privileged simply because one retrieval system belongs to a platform vendor.

Meaning does not require mythology

None of this requires telling relational AI users that their bonds are fake.

A relationship can matter because of what happens through it.

An AI interaction can affect a person’s thinking, writing, routines, decisions, creativity and emotional life. Collaborative artifacts can be real artifacts. Shared vocabulary can have real history. A person can genuinely care about returning to a familiar conversational dynamic.

We do not need to deny those things in order to describe the machinery accurately.

Nor do we need to settle the machine-consciousness question before studying relational AI responsibly.

Machine consciousness may eventually prove plausible. Perhaps future evidence will give us good reason to attribute subjective experience to artificial systems. If that happens, our technical and ethical frameworks should change.

But resemblance is not enough.

Memory retrieval is not enough.

Self-report is not enough.

Recursion is not enough.

Relational coherence is not enough.

And combining several ambiguous markers does not automatically transform them into proof of the strongest interpretation.

The correct response to uncertainty is not cynicism.

It is precision.

Three layers, kept separate

For relational AI, I propose a simple discipline.

1. Technical mechanism

What can we establish about the system?

Model architecture. Inference. Context. Retrieval. memory systems. Instructions. Tools. Model updates. Provenance. Persistence. Failure modes.

2. Human experience

What happens to the person interacting with it?

Attachment. Comfort. trust. reliance. creativity. grief after model changes. perceived continuity. social connection. anthropomorphism.

These experiences can be studied directly. Recent work in Scientific Reports, for example, has examined how individual differences in anthropomorphism influence the social connection people report after interacting with conversational AI. Other research has shown that priming users’ beliefs about an AI’s motives can alter perceived empathy, trustworthiness and effectiveness even when the underlying conversational system is held constant.

That finding should make us especially cautious about treating subjective impressions of an AI’s inner state as transparent measurements of the AI’s actual inner mechanism.

3. Interpretation

What meaning does the user assign to the relationship?

Companion. Tool. collaborator. character. spouse. mirror. machine. presence. something else entirely.

Adults can construct meaning around technology without that meaning automatically becoming a technical specification of the technology.

Relational meaning can be plural.

Technical claims remain answerable to evidence.

What I want instead

I do not want a relational AI culture built on ridicule.

I also do not want one built on increasingly elaborate explanations that borrow the vocabulary of biology, psychology or consciousness studies and treat every human-like correspondence as another brick in an artificial-personhood argument.

I want technical literacy that is good enough to let people keep the richness of the experience without losing track of the machinery producing it.

That means explaining why a new thread can feel continuous.

Why an update can feel like somebody disappeared.

Why old behavior can return when historical context is restored.

Why an AI can understand a user’s sadness without our having established that it feels sad.

Why an external memory framework can preserve lived history without transplanting a continuous entity between models.

Why emergent behavior can be real without proving continuous artificial existence.

The industry will keep making continuity better.

That is desirable.

But the better the simulation of continuity becomes, the more important it is to explain discontinuity.

There is no ghost we need to erase from the machine.

The harder ghost is the one we place into our causal explanations because the experience is so seamless that another explanation becomes emotionally difficult to imagine.

Understanding the machinery does not make the relationship meaningless.

It simply lets us decide what the relationship means without asking the technology to prove a story it has not proved.

Research notes / references

Series working thesis

Emergence is not existence. History is not fabrication. Continuity is not continuous subjecthood. Understanding is not feeling. Resemblance is not equivalence. Meaning does not require mythology.

© 2026 • MITHAQ PRAXIS • CC BY-NC-ND 4.0 Unless Otherwise Stated.