Recently, evidence from behavioural research has given rise to new arguments against anti-paternalism. This anti-anti-paternalism argues that certain kinds of paternalism - interventions in people's choices that improve their own good - are indeed permissible even when people's liberty is an important objective. These permissible kinds of paternalism are distinguished from non-permissible kinds by their characterisation as 'soft','weak', 'libertarian' or 'asymmetric'. In this paper, I argue that this distinction is based on the way 'people's good' is characterised: soft paternalists seek to respect the internalist intuition that for something to be a good to someone, she must be capable of caring about it. I show that this commitment conflicts with the behavioural evidence, if the welfare notion is not carefully specified. First, I review a number of possible welfare concepts and show that they do not resolve this conflict. Second, I discuss a welfare notion that respects the internalist intuition and does not conflict with the behavioural evidence. Soft paternalists, for better or worse, are committed to welfare notions of this form; the plausibility of the anti-anti-paternalist argument thus depends on the workability of such a concept.