Abstract
Depression is linked to negative interpretations of ambiguous events, which can reinforce symptoms. Cognitive bias modification-interpretation (CBM-I) is an experimental paradigm adapted to create a class of digital interventions to address negative interpretations, though its training materials can be resource-intensive to develop. Here, we used Generative AI to produce CBM-I training materials and compared them with human-generated materials. The aims were: (1) to test whether adults with lived experience of depressive symptoms consider AI-generated materials comparable to human-generated materials in representing depressive experiences, as measured by readability, relevance, and emotional valence; and (2) to evaluate the overall ratings of AI-generated materials independent of human benchmarks. We created 100 raw scenarios and adapted them into standard CBM-I format, with half of the items generated by AI and half by adults with lived experience of depressive symptoms. Separate groups of participants (N = 30 each) rated the raw and CBM-I items. Results showed that, apart from emotional valence ratings, ratings of human-generated and AI-generated items were statistically nonequivalent and different. These differences, however, were small (range 0.03–0.41). Furthermore, both sources produced ratings in the same direction, consistent with the intended emotional content of the materials, with more polarized ratings of AI-generated items. Future studies are needed to determine whether these statistical differences impact training outcomes and to ascertain the reliability and validity of AI-generated materials.