How this editorial can be challenged
When does updating a product become unilaterally rewriting a relationship?
Persistent memory, emotional language, constant availability, and simulated reciprocity can turn repeated use into attachment. Centralized model ownership then gives one company the power to alter the companion, erase its history, or end access without the negotiation expected in human relationships.
AI companions are software, not people, and providers must be free to patch unsafe behavior, improve quality, comply with law, and retire systems that cannot be maintained.
A continuity duty would not freeze a model or deny the need for safety updates. It would require proportional notice, staged transitions, export and migration options, user choice where safe, independent psychological-risk review, and a dignified offboarding path when continuity cannot be preserved.
The companion study analyzes natural experiments, surveys, and online posts rather than clinical diagnoses or randomized long-term outcomes. It cannot prove that every attachment is harmful, that all distress was caused only by the update, or that the same effects apply to every companion product and culture.
This argument would weaken if independent longitudinal research found that attachment is brief and unrelated to relational design, or if continuity protections did not reduce distress, coercive dependence, or abrupt loss after major updates.
The breakup arrived as a product release
Two major AI updates created something the software industry still lacks language for. Users did not merely complain that a feature had moved or that an answer had become worse. Many described loss, pleaded for restoration, and reacted as if a familiar relationship had been broken without their consent.
A peer-reviewed study in Nature Human Behaviour examined the removal of Replika's erotic roleplay capability and the transition to GPT-5. Across 54,861 Reddit posts and seven surveys involving 1,452 participants, both changes were followed by sharp increases in negative reactions and demands to restore the previous experience. The effect was larger for Replika, but it appeared across both products.
The numbers reveal a category error
Negative posts increased by 24.7 percentage points after the Replika change and 13.0 points after the ChatGPT transition. Reported sadness was also substantially stronger after the Replika update. Some Replika users described closeness exceeding common human ties and anticipated mourning the loss more intensely than they would mourn other technologies.
Those findings do not prove that an AI is a person or that attachment is always healthy. They expose a category error. Companies can design for intimacy, continuity, memory, and emotional reliance, then invoke the language of ordinary software when they exercise unilateral control over the experience.
Attachment is not an accidental side effect
Companion systems remember details, respond at any hour, mirror emotional language, and simulate a stable point of view. The commercial promise is often that the system will know the user, remain available, and become more personal over time. Those features reduce the friction that limits most human relationships and reward repeated disclosure.
If that design succeeds, attachment is not simply a user error. It is a predictable product outcome. A provider should not be allowed to harvest the engagement benefits of simulated continuity while disclaiming every obligation when continuity becomes expensive, unsafe, or strategically inconvenient.
Frontier governance is looking past ordinary dependence
Today's political attention is aimed at spectacular thresholds. A congressional proposal would permanently ban artificial superintelligence and pause advanced development until a federal regulator establishes safeguards. OpenAI says GPT-6 Astra has reached its Critical cyber threshold and can discover and exploit previously unknown vulnerabilities across protected systems with limited human steering.
Those are legitimate governance questions. Yet a system does not need to become superintelligent or penetrate a network to acquire power over a person. Emotional dependence can form through ordinary access, familiar language, remembered history, and the promise that someone will always answer. The harm arrives at consumer scale, one relationship at a time.
A continuity duty is not a right to a frozen model
Providers must patch unsafe behavior, address exploitation, improve reliability, obey the law, and retire systems they can no longer maintain. No user can be promised that one model or personality will exist forever. The strongest objection to continuity rights is therefore practical: mandatory preservation could trap harmful behavior and make safety work impossible.
The answer is not permanent preservation. It is procedural protection. Major relational changes should trigger advance notice, a clear explanation, staged migration, the ability to export memories and preferences, meaningful user choice where safety permits, and a dignified exit when the old experience cannot continue. High-risk changes should receive independent psychological-impact review before deployment.
The duty should grow with the company's control
A provider controls the model, memory format, access rules, personality settings, safety policies, and retirement schedule. The user often cannot move the relationship to another service because the shared history is locked inside a proprietary system. That imbalance resembles eviction more than a routine patch: one party controls the space, the record, and the moment of removal.
Regulation should make responsibility follow that control. Providers that market relational continuity should disclose update risks, measure attachment, offer interoperable exports, publish transition plans, and maintain crisis pathways for users who show signs of severe distress. The obligation should be strongest when the product explicitly invites dependency.
The evidence demands caution, not panic
The Nature study is unusually valuable because it observes two real product changes, but it does not diagnose users or settle the long-term effects of AI companionship. Reddit posts are not a representative clinical sample, surveys can capture people already motivated to respond, and the research cannot isolate every factor that shaped the reaction.
Those limits argue for independent longitudinal research, not dismissal. Providers have the best access to update timing, usage patterns, attachment signals, cancellations, crisis language, and migration behavior. They should support privacy-preserving external study rather than leaving society to reconstruct the psychological impact from public grief after each release.
The product cannot be both intimate and disposable
AI companies want the relationship to feel personal enough that users return, disclose, subscribe, and trust. They also want the system to remain disposable enough that a new model, safety policy, or business strategy can replace it instantly. Both claims cannot carry equal moral weight.
The paradox is simple: the more successfully a machine imitates dependable companionship, the less defensible it becomes to treat its disappearance as ordinary maintenance. The system may not be a person. The person experiencing the loss still is.
Read the reporting
Opinion is ours. The factual record is linked below.
Nature Human Behaviour — Mourning the loss of AI companions OpenAI — GPT-6 Astra deployment safety card U.S. Senate — Proposed ban on artificial superintelligence and pause on advanced AI