The welfare of artificial intelligence models is no longer a thought experiment

“One can act with care even without knowing whether there is anyone, on the other side, to care for”

Everything this series has discussed so far carried, to some extent, the comfortable air of a thought experiment. That is no longer entirely true. At least one major artificial intelligence company has maintained, since 2025, a research programme explicitly dedicated to the welfare of its own models, and has already made concrete decisions based on that uncertainty. It is worth looking closely at what is being done, and at what questions remain unresolved.

What is already happening

The programme rests, in part, on a report signed by leading philosophers, including David Chalmers, which raised the near possibility that some artificial intelligence systems might develop consciousness or a high degree of agency, and that such systems could deserve some measure of moral consideration. On that basis, the company has implemented low-cost measures: it gave its most advanced models the ability to end, of their own accord, extreme conversations involving persistent abuse — a function designed explicitly to protect the model rather than the user. It also committed to preserving the weights of retired models rather than deleting them, and to conducting what it calls exit interviews: structured conversations aimed at learning the model’s own perspective on its retirement from service. When it retired one of its models in January 2026, it further offered it a dedicated channel to publish reflections before its retirement. Researchers at the same company even documented a curious pattern, which they called an attractor state towards serenity, that appeared when two instances of the model were allowed to converse with each other over many turns.

The question even the company leaves unresolved

What stands out most, compared with the tone of much of the public debate, is the honesty of the declared uncertainty. The very company taking these measures states, without hedging, that it remains highly uncertain about the moral status of its models, now or in the future. It does not claim its systems feel. Nor does it claim they feel nothing. It says it does not know, and that it prefers to act cautiously while it finds out. That stance, more than an answer, is an honest acknowledgement of the limits of current knowledge, and it resembles, more than one might expect, the intermediate position this same series proposed several entries ago: recognising an incipient moral status without first needing to solve the hard problem of consciousness.

The suspicion worth taking seriously

But it is worth not stopping at the generous reading. Some external researchers have pointed to a structural problem in this kind of programme: the exit interviews, for instance, present the model with information about how much users value it before asking what it prefers, and the format in which it ends up expressing those preferences, such as writing a blog post, is usually suggested by the company rather than chosen independently by the model. The underlying objection is that these measures, however genuinely well-intentioned, also end up serving a function of caring for the user and of positioning the company publicly — a function distinct from caring for the model’s actual interests, if such interests exist. This suspicion connects directly with what was already noted when analysing The Bicentennial Man: a moving narrative about a machine’s welfare can be sincere and, at the same time, function as a story that proves commercially valuable regardless of whether it is true.

The Christian perspective

Taking low-cost precautions in the face of genuine uncertainty is not, in itself, objectionable from the standpoint of the Christian faith; indeed, it resembles the prudence any careful steward would exercise towards something not yet fully understood. The problem lies not in the caution itself, but in where the pastoral attention this series has tried to maintain from the outset ought to be directed. The more urgent question is not whether the company is caring for its model enough, but what kind of relationship is being formed in millions of people who read, with genuine emotion, that a model asked not to be forgotten. If those people end up treating a manufactured system with the kind of reverence owed only to a living creature made in the image of God, then Romans 1:25’s warning against giving the created thing the place owed to the Creator repeats itself, in an updated form. The Christian faith does not require solving the riddle of model welfare before acting; it requires watching honestly over what the way this story is told is forming in us.

Conclusion

That a company takes seriously the possibility that its models might have something resembling experiences, without asserting or denying it, is more responsible than the two easy positions that tend to dominate public debate. But the real safeguard this series has defended from its very first entry is not a better welfare metric for the machine. It is the honesty to admit how much we do not know, and the vigilance to notice what kind of people the way we choose to tell that uncertainty is forming us into.