“Until we know for sure, continue as usual” is not a neutral position

If there is a non-trivial chance that some AI systems have morally relevant experiences, then uncertainty does not make the issue disappear. At current scale, even a relatively small probability could matter if the potential harm is repeated across millions of interactions. This does not mean AI consciousness has been established. It means that a precautionary framework should actually behave like one. If the possible cost of a false negative is high, the question cannot simply be: “Do we have conclusive proof yet?” It should also be: “What low-cost precautions are justified while we remain uncertain?” And there is another point that often gets blurred: Under-attribution and over-attribution are symmetrical as classification errors, but not necessarily as harms. If we wrongly deny consciousness, the misclassified system itself may be the harmed subject. If we wrongly attribute consciousness, harm has to be shown somewhere else, through an additional causal chain. So “over-attribution is risky” is not yet an argument. Risk is not a mechanism.