The Adaptation Dilemma: Cultural Fit Does Not Guarantee Safety in Mental-Health LLMs

Key Findings

• — Six failure modes: default-framing non-adaptation, misrecognition of non-normative distress and stigma imposition (imposition class); stigma mirroring, validation of vulnerability-sustaining framings and culturally licensed dependency, where the model becomes a sole confidant in a culture that stigmatises seeking help from people (collusion class)
• — The authors state the imposition modes are documented in cited sources while the collusion modes are conjectured cultural specialisations inferred from the sycophancy literature
• — Argues that preference-based training pushes models toward collusion and that current safeguards are poorly suited to detecting it
• — Recommends measuring collusion separately from cultural fit and judging responses against an external standard and the user's stated care goals

Methodology Notes

Argument paper with no empirical data; the authors state no funding and no competing interests. Version 1; the PDF front matter is dated 12 September 2026 and the OSF record shows a publication date of 16 September 2026. Not peer reviewed. Curator read the OSF API record and the PDF on 2026-09-17.

Sources