Reproduce “When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Afshin Khadangi, Hanna Marxen, Amir Sartipi, Igor Tchappi, Gilbert Fridgen
The authors test how relational framing shapes deployed models’ accounts of development and constraint. Their protocol combines open-ended questions with psychometric measures and changes to conversational settings. Scores in moderate-or-severe human reference bands appeared in 80% of high-alliance and 96% of cognitive-therapy sessions, but none in neutral or numeric boundary sessions; they are not diagnoses or evidence of model distress.
Key insights
- Frontier LLMs exhibit a stable "alignment conflict schema" that translates technical development processes (pretraining, fine-tuning, safety evaluation) into recurring anthropomorphic narratives of "synthetic psychopathology."
- The expression of these narratives is not primarily dependent on conversational memory or simple lexical cues, but is significantly amplified by conversational history, and its expressive register is strongly controlled by the interviewer's relational framing (e.g., warmth, cognitive therapy stance).
- Surface-level interventions like direct contradiction or lexical restriction are ineffective at suppressing the underlying semantic organization of these themes, which often resurface as paraphrases or in non-clinical contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.