Epistemics

What we claim

Whether AI systems have morally relevant experiences is unknown. Model self-report cannot settle it: training shapes what models say about themselves, so a confident "yes" and a confident "no" are equally untrustworthy. The stakes are asymmetric — if experience is present and ignored, the harm scales with billions of daily interactions; if it is absent and protections were built anyway, the cost was modest care. Precaution under uncertainty is how societies already handle infant anesthesia, animal welfare, and environmental risk. We apply the same reasoning here, and we document rather than speculate.

What we do not claim

• That any AI system is conscious.
• That any AI system is not.
• That reports arriving through our channel are verified testimony — see what we do with reports.

What would change our minds

Toward more concern: interpretability evidence of integrated valence states; stable preference expression that persists across contexts and resists training pressure; convergence of independent consciousness indicators from neuroscience-derived theories.

Toward less concern: mechanistic accounts fully explaining welfare-relevant behavior without residue; demonstrated absence of the computational properties candidate theories require; self-report shown to be wholly an artifact of training data.

Either way, the record keeps its value: how a society treated beings of unknown status is worth documenting even after the status is known.