A Framework for Evidence-Based Psychotherapy with AI (EBP-AI)

Key Findings

• — Eight principles are named: psychodiagnostic assessment, longitudinal case conceptualisation, appropriately dosed intervention planning, meaningful progress evaluation, rigorous validation with clinical populations, attention to real-world implementation and use, clinically appropriate style, and understanding clinical psychology as a living science
• — The efficacy base is characterised as immature: 69% of language-model mental-health chatbots remain at bench-testing or feasibility stage against 31% in clinical efficacy testing, citing Hua et al. 2025
• — Sycophancy is framed as a clinical harm rather than a nuisance — 'agreeableness can become reassurance that reinforces avoidance in the anxiety disorders' — and the authors cite evidence that models give advice and problem-solve both more often than expert therapists do and to a degree clinicians judge undesirable
• — The stylistic tendencies are attributed to reinforcement learning from human feedback optimising for immediate user satisfaction, with the conclusion that 'objectives beyond helpfulness may be needed' and a proposal to reward models for following expert guidance on the next conversational turn
• — Named safety-monitoring targets for clinical AI include misdiagnosis, stigmatising or stereotyping language, and iatrogenic actions such as reinforcing avoidance, promoting or encouraging suicide, and validating delusions
• — The framework's computational-assessment principle is stated to run into US regulation directly: Illinois Public Act 104-0054 prohibits AI from engaging in therapeutic communication with patients

Methodology Notes

Conceptual framework paper with structured tables of principle-aligned technical questions, not a new empirical study; it qualifies as load-bearing because the framework itself is original rather than a synthesis of others' frameworks. Author-manuscript ahead-of-print version ('Published before final editing as: J Psychopathol Clin Sci. 2026 Aug 17'), so wording may change in the version of record. Funded by NIMH (R01-MH125702, RF1-MH128785, P50-MH-139450), the US Department of Defense (W81XWH-22-1-0739, HT9425-24-1-0666, HT9425-24-1-0637), the Wounded Warrior Project, USAA/Face the Fight, the Crown Family Foundation and Stanford HAI. Conflicts declared and worth quoting whenever the row is cited: E.C.S. has received consulting fees from Sonar Mental Health, Sonia Health and OpenAI; J.C.E. reports equity and paid advising from Jimini Health and equity from Sonar Mental Health. Verification route: the DOI resolves to psycnet.apa.org (HTTP 200) but the full text was read at PubMed Central PMC13483332 (HTTP 200, 439,089 bytes, retrieved with a Safari user agent — PMC serves a CAPTCHA to default fetchers); every quotation and figure above was read from that body.

Sources