How do children discover which linguistic expressions are associated with presuppositions? Do they take a direct strategy of tracking whether linguistic expressions are associated with particular speaker presuppositions? This strategy may fail children who are trying to learn about the presuppositions of so-called 'soft' presupposition triggers, which can be readily used even when the relevant would-be presupposed content is not part of the common ground. We present a corpus study with the soft trigger "know" and the related, but non-presuppositional "think". We find that a direct learning strategy would indeed run into problems for such a soft trigger given the nature and availability of evidence in children's linguistic input.