Challenging Psychedelic Experiences Across DMT, Psilocybin, MDMA, and Ketamine: A Thematic Analysis of User Reports with Exploratory Human-LLM Concordance Analysis

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Renewed clinical interest in psychedelics has drawn attention to their challenging effects, yet most phenomenological and psychometric work derives from psilocybin, leaving cross-substance variation underexplored. This thesis employed codebook thematic analysis to examine challenging experiences across DMT, psilocybin, MDMA, and ketamine, utilising 146 naturalistic accounts collected from Erowid’s Experience Vaults. Eight themes were identified, encompassing acute challenges, attenuated acute effects with repeated use, and sustained post-experience difficulties. Although challenging experiences converged substantially across substances, several substance-specific features emerged: dying convictions and entity encounters under DMT, disembodiment and void-like death confrontations under ketamine, unity experiences entangled with death-related distress under psilocybin, and euphoria alongside a notable absence of other mystical-type phenomena under MDMA. Acute challenges were more wide-ranging and severe under DMT and ketamine. Attenuated acute effects with repeated use were specific to ketamine and MDMA, while sustained challenges occurred across all four substances. Alterations in the basic sense of self, identified across themes, are interpreted through the Relaxed Beliefs Under Psychedelics hypothesis and Sass and Parnas’ ipseity disturbance model of schizophrenia. We argue that phenomenological accounts of altered self-experience may be underutilised in psychedelic science. An exploratory component examined interpretive concordance between the human researcher and a Large Language Model (Claude Sonnet 4.6) across three complexity levels. Contrary to predictions, concordance was non-linear (Level 1 [concrete] > Level 3 [abstract] > Level 2 [mid-level]; mean κ = 0.888, 0.665, and 0.421 respectively), which suggested that phenomenological and linguistic discriminability, rather than conceptual abstraction, may better predict Sonnet 4.6 coding concordance, at least in this corpus.

Links and sources

Comments