The Remember/Know procedure, developed by Tulving (1985) to capture the distinction between the conscious correlates of episodic and semantic retrieval, has spawned considerable research and debate. However, only a handful of reports have examined the recognition content beyond this dichotomous simplification. To address this, we collected participants' written justifications in support of ordinary old/new recognition decisions accompanied by confidence ratings using a 3-point scale (high/medium/low). Unlike prior research, we did not provide the participants with any descriptions of Remembering or Knowing and thus, if the justifications mapped well onto theory, they would do so spontaneously. Word frequency analysis (unigrams, bigrams, and trigrams), independent ratings, and machine learning techniques (Support Vector Machine [SVM]) converged in demonstrating that the linguistic content of high and medium confidence recognition differs in a manner consistent with dual process theories of recognition. For example, the use of "I remember," particularly when combined with temporal or perceptual information (e.g., "when," "saw," "distinctly"), was heavily associated with high confidence recognition. Conversely, participants also used the absence of remembering for personally distinctive materials as support for high confidence new reports ("would have remembered"). Thus, participants afford a special status to the presence or absence of remembering and use this actively as a basis for high confidence during recognition judgments. Additionally, the pattern of classification successes and failures of a SVM was well anticipated by the dual process signal detection model of recognition and inconsistent with a single process, strictly unidimensional approach.