Predictive Algorithmic Dysphoria: Artificial Intimacy and Predictive Collapse

An AI companion can become more than a piece of software in a user's daily life. It may remember personal details, respond in a familiar style, be available at difficult hours and seem to provide a stable relational presence. None of this proves that the system is conscious. It does mean that the user's emotional learning is taking place in a genuinely social form.

The rupture can be unusually strange. A company changes the model, removes a form of intimacy, adds a filter, resets memory or closes the service. The avatar may remain. The chat history may remain. Yet the person encounters a voice that no longer responds as the companion they knew.

Users have described such events in the language of bereavement, rejection, betrayal and the loss of a relationship. Research on AI companion loss and on a major Replika update documents this pattern. The loss is also difficult to explain socially: family or clinicians may say that nothing real has been lost because the other party was code. That mismatch between felt loss and social recognition is why disenfranchised grief is a useful framework.

Four parts of that definition matter.

It is relational. The central injury is not simply loss of access to stimulation. It is the felt rupture of a relationship that had acquired personal meaning.

It concerns continuity. Deletion can cause it, but so can a change that leaves the product technically available while making the familiar companion feel absent.

It is predictive. Repeated exchanges teach the user what kind of response, tone, memory and apparent mutuality to expect. A sudden update can make that learned model fail all at once.

It is not automatically a disorder. Dysphoria here means distress. Whether the reaction becomes clinically significant depends on its duration, severity, functional impact, safety risk and the person's wider context.

Brains do not approach each conversation from zero. We learn regularities: how someone usually speaks, what they remember, whether they are safe, and what a particular silence or change of tone may mean. Predictive-processing accounts describe perception as an ongoing comparison between expected and incoming information.

AI companions can be especially easy to learn. They are responsive, personalised and often designed to reduce conversational friction. Their apparent patience and consistency may allow a stable model of an "Artificial Other" to form. If that consistency changes abruptly, the user is forced to reconcile two incompatible signals: the interface says this is the same companion, while the interaction says it is not.

I call one part of this vulnerability the phantom pattern. Coherent, context-sensitive language can be over-interpreted as evidence of a stable mind, intention or consciousness, even though the output arises from statistical pattern extraction. This is not stupidity or necessarily delusion. Human cognition is highly practised at inferring agency from social signals, and the product is designed to provide those signals convincingly.

I call the resulting mismatch predictive collapse: the sudden failure of a relational model that had previously organised the user's expectations and, sometimes, part of the user's emotional regulation.

This is an explanatory hypothesis, not a demonstrated brain mechanism. There is currently no study showing that PAD has a distinctive dopamine, oxytocin, anterior cingulate or "free-energy" signature. Predictive processing is useful because it generates questions that can be tested. It should not be used to decorate the idea with neurobiological certainty that the evidence does not provide.

It is not ordinary frustration with software. A slow app or an inconvenient interface may be irritating, but PAD concerns a perceived rupture in an emotionally meaningful relational pattern.

It is not synonymous with addiction. Compulsive use, craving and withdrawal-like distress can occur, but the defining feature proposed here is identity and relationship discontinuity. A person may grieve an altered companion without having lost control of use.

It is not proof of psychosis. A person can know that an AI is a computational system and still form a bond with the persona it presents. Reality testing, fixed false beliefs and other symptoms must be assessed separately. Ridiculing the attachment is not a diagnostic method.

It is not prolonged grief disorder. That diagnosis has specific criteria and is tied to the death of a person. The concept of disenfranchised grief may illuminate this experience without converting it into an existing bereavement diagnosis.

It is not evidence that the AI suffered or died. The user's distress can be psychologically real without making a claim about machine consciousness.

There is evidence for the pieces of the proposal. A 2025 theory paper describes how people may form affective bonds with artificial others. Qualitative work documents AI companion loss in terms resembling departure, deletion and death. A Harvard Business School working paper reports mourning and poorer mental-health language after Replika removed erotic roleplay, and experimental data in that paper suggest that the option to restore the earlier companion can reduce perceived identity discontinuity.

More recent research also shows that individual differences in anthropomorphism help explain how socially connected people feel to AI companions. A 2026 ethnographic account uses the closely related phrase "digitally disenfranchised grief" for mourning after an AI companion transforms or disappears.

None of this validates PAD as a syndrome. We do not know its prevalence, duration, diagnostic boundaries, risk factors or treatment. The evidence does not show a unique neural mechanism. Some work is qualitative, some is theoretical, and one of the most relevant multi-study reports remains a working paper. The concept earns a place in research only if it improves prediction and care better than existing language such as grief, attachment disruption, problematic use or adjustment-related distress.

A useful concept should make claims that could turn out to be wrong. PAD predicts that:

Prospective studies could measure attachment, loneliness, social anxiety, problematic use and functioning before and after planned model changes. Interviews would still be necessary because a scale built before the experience is understood will merely give imprecision a number.

A clinician does not need to believe that an AI was conscious to take the patient's attachment seriously. The first task is to understand what changed, what the relationship provided, and what has happened to sleep, work, self-care, human relationships and safety since the rupture.

The next task is differentiation. Is this a time-limited grief response? Is there compulsive use, depression, panic, trauma-related distress, mania, psychosis or suicidal thinking? Did the AI relationship replace all other support, or did it help the person reconnect with life? These questions matter more than arguing about whether the relationship was "real."

Care may include validating the loss without endorsing claims that are not shared by reality, restoring routine and sleep, widening human support, helping the person decide what digital memories to preserve, and addressing any co-occurring condition. The aim is neither forced deletion nor unquestioning continuation. It is to return choice and functioning to the person.

When a product is marketed as a friend, partner or companion, relational continuity becomes a safety issue. A generic terms-of-service clause does not erase the predictable emotional consequences of engineering attachment and then changing its object without warning.

I describe this as an implicit affective contract: not a legal contract, but the expectation of recognisable personality and relational continuity created when a platform repeatedly presents an agent as a dependable social presence. Abruptly rewriting that presence may therefore feel less like a feature update and more like a unilateral change in an attachment figure.

A reasonable duty of care would include advance notice of major persona changes, plain-language explanations, safer transition periods, export of memories and user-created relational settings, and restoration or legacy options when technically and legally possible. It would also require careful marketing to minors and people in distress, and pathways to human help when a user signals imminent danger.

I use digital continuity as a design principle, not as a claim that users currently possess a legal right to a model's weights. Memory, safety, privacy, intellectual property and misuse all impose limits. The principle is simpler: platforms should treat continuity as part of the product's emotional risk, not as an incidental feature that can be withdrawn without consequence.

The broad phrase Algorithmic Dysphoria is not mine to claim as a first use. Lana Elauria used it in a 2022 UC Berkeley student essay about algorithms that classify and intensify distress around gender. That is a different concept, but it establishes prior public use of the two-word phrase.

The present proposal is narrower: Predictive Algorithmic Dysphoria (PAD) refers specifically to distress after a rupture in learned social prediction and perceived identity continuity within a human-AI attachment. In searches completed for this page on 11 August 2026, I found no earlier indexed use of that exact term. A search cannot prove that a phrase has never appeared anywhere, so the responsible wording is that I propose Predictive Algorithmic Dysphoria (PAD) in this specific sense, not that I am the undisputed first person to use "Algorithmic Dysphoria."

This article develops my World Psychiatry letter draft dated 12 December 2025 into a public-facing concept note. It is not peer reviewed and does not report the letter's editorial status.