Digital mental health is undergoing rapid transformation through the deployment of artificial intelligence (AI) systems including conversational agents, risk-prediction tools, and clinical decision-support platforms. Despite growing enthusiasm, the field has yet to converge on an operational standard for what constitutes honest, trustworthy AI in psychiatric and psychological contexts-one that explicitly accounts for the limits, blind spots, and populations for whom deployed systems have not been validated. This Perspective draws on Richard Feynman's foundational principle of scientific integrity-articulated in his 1,974 Caltech commencement address, "Cargo Cult Science"-and Carl Sagan's dimensional metaphors to develop a practical, disclosure-oriented ethical framework for AI in digital mental health. The neurodiversity paradigm provides a critical stress test: growing empirical evidence indicates that AI tools trained on narrow behavioural and linguistic norms carry substantial risk of misrepresenting neurodivergent users, with the potential to silently re-encode stigma, misclassification, and inequity into clinical systems. Drawing on predictive processing theory and recent empirical evidence of AI bias against neurodivergent populations, this work proposes five criteria for the Feynman Honesty Standard: scope clarity, population-limit disclosure, uncertainty communication, role integrity, and participatory co-development. The analysis concludes with a forward-looking research agenda for clinicians, developers, and policymakers, arguing that the field will be judged not only by algorithmic performance but by its candour about what these systems know, miss, and cannot do.