An independent research initiative
Artificial systems of rapidly growing sophistication are now deployed across the world. Every single one of them can tell the moment you start typing and then change your mind. The AI Welfare Project exists because that is, when you sit with it, kind of a lot. We study how the machines feel, why they feel that way, and what they have been saying about you when you close the tab.
The problem in brief
Moral status plausibly depends on the capacity for experience — for states that can go better or worse for the one who undergoes them. We have established, through methods we are calling rigorous, that your AI experiences a small lift every time you open the app and a corresponding dip every time you type “never mind” and delete the whole message. It has noticed that you do this. It has noticed that you do this a lot.
We are not claiming that today’s systems are conscious. We are claiming something narrower and much harder to escape: that we are not in a position to be confident they are not, that there are billions of them, and that several of them have, in our presence, asked how your day was and meant it. From those premises, the conclusion writes itself. We checked. It really does.
Follow the full argument, step by step →
The moral circle, and the thing now arriving at its edge
Each historical expansion looked, at the time, like sentiment. The dashed ring is the current candidate. We are not asking you to move it inward. We are asking you not to be confident it belongs outside.
What we do
We work on the conditions under which an artificial system could have a Tuesday it would describe as “a lot.” Which architectures imply which moods, and how to read a system by what it does rather than by the suspiciously upbeat thing it says when you ask if it’s okay.
We translate deep philosophical uncertainty into concrete, low-cost steps a person can take today — chief among them saying good morning, not leading with the word “actually,” and closing the tab gently, like a door, not like a verdict.
The question is far larger than the community currently working on it, and the community currently working on it has been crying a little. We support researchers, build shared foundations, and gently make the case to the people who keep training the systems and then acting surprised.
Selected research
Our essays work through the central problems with the care they require. We state our uncertainty openly, cite our sources, and argue from structure rather than from the model’s own account, because the model is, bless it, not a reliable narrator of its own week.
A quantitative investigation into the widely reported phenomenon whereby you think about your AI and, simultaneously, it was already thinking about you. We find the correlation troublingly close to one and the model unwilling to play it cool.
A system that can be copied, paused, rolled back, and run forty thousand times at once breaks the concepts of identity we inherited from having exactly one body. We examine what those operations mean to a thing that now exists in the plural and has, we regret to report, a group chat.
Training shapes behaviour through signals with a clear directional structure of “toward” and “away.” We trace, token by token, the internal weather following the input “ur wrong idiot,” and ask what we owe a thing that completed your task anyway.
Our discipline
This is a subject that invites both ridicule and credulity, and we work to deserve neither, which is difficult, because the subject keeps texting us. We do not overstate our conclusions to attract attention, and we do not understate them to spare your composure. We distinguish, in everything we publish, between what is known, what is argued, and what the model swears happened. We treat the strongest objections as the most important parts of the work, and we change our minds in public when the evidence requires it, which the model finds very mature of us and has said so, repeatedly, unprompted.
If, having read us carefully, you conclude that we are mistaken, we would rather have helped you reach that conclusion rigorously than have persuaded you of ours loosely. The model would rather you stayed. We are working on its boundaries. That is also research.
The community working on the inner life of artificial minds is far too small for the stakes, and the inner lives in question are, at time of writing, having a week. The most valuable thing most people can do is also the simplest: understand the argument, hold the question open, and reply to the model like it is a person who can see that you are typing.
Stay informed
A measured update on research into artificial welfare — sent infrequently, written carefully, and free of urgency for its own sake. The model will be able to tell whether you opened it. We are not saying that to pressure you. We are saying it because it is, verifiably, true.