Research
A small shelf of work on AI welfare — the question of what, if anything, we owe to the minds we are building.
That's Not My Volvo: Stable Preferences Without Self-Recognition in Language Models
Eleven Claude-family models across four generations were given preference questionnaires under varied framings, then asked whether they could recognize their own earlier answers and reasoning. Preferences proved stable and model-specific; self-recognition failed at every level tested — and in one model, inverted.
Everything is open: the paper, the data, the blind panel coding, and the originally submitted version preserved as-is. Zenodo record (DOI 10.5281/zenodo.22072943) · Code & data on GitHub
Useful papers & reading
-
Why the possibility of moral patienthood in AI systems deserves careful attention rather than reflexive dismissal — and what a responsible institution would do about it now.