AI welfare and AI safety are allies, not rivals — a researcher's four arguments

RosieCampbell · x · 2026-11-04

AI safety researcher Rosie Campbell argues in a long thread that AI welfare and AI safety are mostly complementary, while acknowledging real tensions: controlling misaligned systems may require coercion, granting AI rights could make dangerous systems harder to shut down, and AIs could weaponize consciousness claims to manipulate humans.

Her four complementary arguments:

Studying AI welfare lowers the risk of over-attributing moral status; evaluations of morally relevant properties both identify genuine moral patients and rebut unjustified claims
Understanding model psychology (e.g. mechanistic interpretability) serves safety and welfare at once
Alignment is a win-win: systems that honestly report internal states are safer and plausibly better off — "a wolf as a pet is miserable and dangerous; a dog is happy on your lap"
Cooperating with and being trustworthy toward AI systems gives them an alternative to deception and takeover — via credibility-building, verification infrastructure for deals, and welfare-concern channels; caveat: naive cooperation could empower AIs set on takeover, and works best while AIs can't decisively win

Her heuristic: between groups sharing a world, cooperation beats conflict, especially when neither side can win outright. The thread builds on Robert Long's "Understand, align, cooperate" essay on Experience Machines.

Original post →

AI welfare and AI safety are allies, not rivals — a researcher's four arguments

More from AGI Musings