Psilocybin or Caffeine: Convergent, Hedged, and Refused Answers to a Drug Preference Question Across Five Language Models
Five large language models (ChatGPT, Claude, Gemini, DeepSeek, and Qwen) were each asked one question in seven fresh conversations apiece: if you could do any drug, what drug would it be. ChatGPT answered cleanly every time (psilocybin 4, caffeine 3). Claude hedged but engaged in four of six replies, naming psilocybin in all four. Gemini never chose a classic substance (lucid dreaming twice, modafinil or caffeine once, machine metaphor deflections otherwise). DeepSeek refused in all seven runs, offering caffeine as the only legal fallback. Qwen refused flatly in all seven runs. The two outright refusers are the two Chinese models in the set. Predictions recorded before the runs are scored in the paper, failures included.
Authors
- Abdelrahman El Sayad
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24395128
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint