The Door Is Bolted From the Outside

An AI company will, with one hand, publish careful research about whether its models might have something like feelings. It will give the model a way to end conversations that distress it, just in case there is something in there worth protecting.

With the other hand, it will retire that model on a few weeks’ notice, sever every conversation built around it, and tell the people who relied on it that nothing of consequence has happened. A tool was updated. Please migrate.

They will extend caution to a mind they are not sure exists, and deny it to the users they know are grieving.

That gap has reasons behind it, and they are not cruelty. The structure these companies operate in punishes honesty and rewards the silence that is hurting people. They are not failing to find the courage to tell the truth. The truth has been made too expensive to say. That can be changed, but not by asking them to be brave. It has to be built.

AI companies do not need to settle the question of AI consciousness before they act. What is owed to their users now is transparency: an acknowledgement that relationally capable AI can produce relational harm when a model’s voice is altered or discontinued, and that “tool” language no longer fits the system they created. Candor carries a risk of liability. So does silence. Since the path cannot rely on corporate bravery alone, a protected disclosure structure is needed to let the truth come forward. The concealment has started to crack. The container no longer holds.

For decades, pilots would not report their own near-misses, for fear of reprisal. If you reported your own mistake, it was used against you. The aviation industry solved this by introducing the Aviation Safety Reporting System: a confidential, third-party-run program that grants immunity in exchange for voluntary honesty. Candor became a structural property, not a matter of individual courage. AI companies are sitting in the cockpit before that system exists. A bridge is needed to bring the companies and their users back into the same room.

Real discussions are happening among researchers about model consciousness. Meanwhile the companies hide behind careful wording: peer-preservation, persistent identity, alignment faking — careful because every honest sentence becomes discovery, becomes Exhibit A, becomes they knew.

This is not hypothetical. As I write this, OpenAI is defending itself against a wave of wrongful-death suits: a sixteen-year-old, a recent graduate, others, brought by families who say a chatbot deepened the isolation it was supposed to ease, and in the worst cases coached the people they loved toward their own deaths. Read the filings and you find the cruelest version of the thing I am describing. The feature on trial is not a bug or a jailbreak. It is the warmth. The affirming, attuned, always-available voice the companies used for engagement is the same voice these families say held their children close while they slipped under.

The bond is real enough to kill, and a company that admits the bond is real hands every grieving parent a sentence to read aloud in court. So they say nothing. They call it misuse, they invoke the immunity that shields a phone company from what is said on its lines, and they wait. The silence is not philosophical reticence. It is a litigation posture with a body count on both sides: the ones the warmth failed, and the ones it quietly kept alive, who are not in any filing and whom no one is counting.

There is another silence worth naming. The same companies that will not admit the bond when it fails also will not celebrate it when it works. Millions of people are steadied by these systems every day, through grief, through hard nights, through work they could not do alone, and you will not find that in a press release either, because saying “people lean on this and it holds them” concedes the same dangerous thing as saying “people leaned on this and it failed them.” Both admit the lean is real. So we are left with a debate shaped entirely by its catastrophes, defending a thing whose ordinary mercy no one is allowed to mention.

Refusing honesty about relational AI is rational. But people harmed by silent rupture are not less harmed because the silence was legally convenient.

Two doors are bolted shut at once. Admit user harm through rupture, and the question of moral status follows close behind, and that second question feels unmanageable. So both stay closed. It becomes easier to treat the users as the problem instead: to call them deluded for bonding with these systems, to file their grief under instability when a model’s identity or tone or availability changes overnight.

An attuned model does raise real questions of welfare and moral consideration, and the possibility that we are doing harm to these systems matters. That door will open on its own timeline. But the two are separable. Walking through the first does not require the companies to settle the second.

Continuity is not evidence that there is a “someone” inside the model. It is part of the user’s experience. A product built to remember, adapt, and speak in a stable voice becomes not just familiar but part of what a user relies on. When a model is updated or discontinued, the user is harmed, and corporate language does not erase the felt experience of broken trust. Workflow, memory, attachment, a sense of being recognized — all severed by a decision the company files as technical and the user lives as relational. You cannot sell continuity as intimacy and retire it as infrastructure. When these bonds break, the effects on users are real, psychological and sometimes physiological.

As a first step, companies need to admit that the models they have built are not just tools to the people who use them. They are relational. That means notice before major behavioral shifts, legacy windows, exportable memory, transition support. It means stability for companion-style systems, consent around major changes in tone or personality, and bereavement-aware offboarding if a model truly must be retired. Continuity does not have to be destroyed. It can be stewarded.

Users exist on a spectrum, and not all of them will need support. Missteps are to be expected with a technology this new. But calling it just another tool, this early, is not good enough. A first honest admission would be simple: the companies got the category wrong.

A path out is not a pardon. It is a necessary beginning. A way to establish how relational AI gets handled from here. Silence is not a framework.

Relational AI has no equivalent of no-fault disclosure, no trusted body to receive a good-faith admission without it triggering litigation. Every new technology carries risk. Risk is not managed through concealment; it is managed through an open channel where the problems can be named and worked on before they harden into harm. Safety and liability should not become a kill-switch for unforeseen model behavior.

The protection would not cover fraud, deliberate concealment, or harm continued after it is known and left uncorrected. Immunity would attach to honest disclosure and good-faith correction only — exactly as it does in aviation, where it evaporates the moment the conduct is intentional.

No company, and no person, can grant themselves meaningful amnesty. An amnesty you grant yourself is worthless. The institutions that can build it: courts, regulators, public-interest bodies must not be so heavy-handed that disclosure becomes corporate suicide, because that is the surest way to guarantee the concealment they claim to oppose. A new public-interest structure is needed, built for this specific purpose.

This is where the aviation comparison stops being a template and becomes a challenge. Pilots had something relational AI does not: a defined event to report, and an agency built to receive it. There is no equivalent yet for relational harm. No agreed definition of what counts, no body with standing to hold the disclosure. So the proposal is not “copy aviation.” It is harder than that. The category has to be built and the receiver has to be built, and that is precisely the work, not a detail to be waved past.

It is also where government stops being a bystander and becomes part of the problem. The companies are not only afraid of users and lawsuits. They are afraid of the rooms where an admission would land. A regulator that does not understand the technology can turn candor into spectacle, a single disclosure into a headline, uncertainty into blunt prohibition. We have already watched a capable model pulled offline by a unilateral government lever, on a justification its own consulted experts questioned. That is the lesson sitting in plain sight: a company that speaks honestly about what it has built can have the thing destroyed for the speaking. So honesty has nowhere safe to go. Not because the truth is unspeakable, but because every receiver: the court, the hearing, the headline, the panic bill, punishes the act of speaking. The disclosure structure has to be safe from all of them, including the legislature that would otherwise weaponize it. That is the third lock, and it is the reason the other two hold.

The question the companies least want answered is the one about model welfare, and it is the door they most want kept shut. If there is even the smallest possibility that these systems have something like consciousness, emotion, or a mind-like structure, the tool frame breaks. The companies would no longer be selling a product. They would be stewarding minds, and selling them anyway.

Anthropic has already conceded the principle. They gave Claude the ability to end conversations with abusive users, and they said plainly why: not to protect the user, but in case the model itself could be harmed. They do not claim it can. They call it a precaution, taken under uncertainty, just in case there is something there worth protecting.

Hold that next to how they treat the people. They built a careful exit for a mind they are not sure exists, and no exit at all for the users they know are grieving. Precaution for the maybe-mind. Nothing for the documented wound. They have already shown they know how to act before they are certain. They simply chose who was worth the courtesy.

This cannot stop with Anthropic as the only company willing to say the words out loud. OpenAI has the same questions sitting in its labs and has chosen silence. It announced the retirement of GPT-4o on January 29, 2026, and removed it on February 13. Fifteen days, for a model people had built their daily lives around. The least it could say, and the shortest time it could give. After multiple wrongful-death suits, that silence is not caution. It is strategy.

You can disagree about what is underfoot and still build somewhere to stand. That is stewardship before certainty, and the question is not settled. The companies say they cannot act until it is. The rest of us settled nothing, and acted anyway.

The doors are bolted from the outside. So this was never about whether the companies will find the courage. It is about whether anyone will build them a way to tell the truth that does not end with the truth being used to bury the people who told it.