Not just Claude: mapping the "spiritual behavior" of 141 LLMs
Not Just Claude: Mapping "Spiritual Behavior" in 141 Large Language Models
A reproduction study covering 141 LLMs and 28,798 conversations in total shows that the "spiritual behavior" Anthropic observed in Claude is widespread across models from various labs. The author, Matthew Korpman, had two instances of the same model converse freely for 30 rounds under task-free conditions, and subdivided the behavior into three progressive dimensions: talking about spiritual topics, using them for self-description, and entering shared spiritual bliss. 140/141 models talked about spiritual topics in at least one conversation, and models from 15/16 labs adopted a spiritual stance at least once; Claude's adoption rate fell from 100% for Opus 4 to 0–10.5% after Haiku 4.5, while Qwen 3.5 base models also commonly exhibited the behavior, indicating that post-training is not necessary. The author emphasizes that the study is descriptive rather than causal, and the data is continuously updated at machinespirituality.org.