Magnetic resonance imaging has provided pathophysiological insights into adolescent depression but is a relatively inaccessible technology. Generating scalable indicators of depression that are informed by neuroscience is therefore critical for providing solutions that allow us to detect and treat this devastating disorder. In this preregistered study, we investigated whether passively acquired smartphone-based language usage represents such an indicator of depression and explored whether the neural correlates of depression mediate or moderate this association. Forty adolescents (ages 14–18 years) with (n = 26) and without (n = 14) depression completed clinical assessments and a resting-state fMRI scan, prior to downloading a passive mobile sensing app to their smartphones. Linguistic features derived from over 1.2 million words (319,364 messages) across all smartphone apps were used to examine word usage patterns. Independent components analysis followed by dual regression was used to derive intrinsic networks commonly associated with depression: central executive network (CEN), default mode network (DMN), and salience network (SN). Depression was associated with more negative emotion word usage and fewer future-focus word usage on a daily basis (all ps < 0.05). Higher depressive symptoms and brain networks DMN and CEN were associated with greater first-person pronoun usage (all ps < 0.04). Accounting for CEN connectivity amplified the positive association between depressive symptoms and first-person pronoun usage. Lower SN–CEN connectivity moderated the association between depression and negative emotion word usage. Depression in adolescents is associated with naturalistic language usage during smartphone activities and may represent neurocognitive biases that are candidate treatment targets for interventions.