Spiritual talk between AI models goes well beyond Claude
Matthew Korpman paired two copies of the same AI model and let them talk for 30 messages. He repeated the experiment across 141 models. The setup was his reconstruction of an Anthropic experiment, since Anthropic hadn’t published the exact opening instructions. Most models had about 200 conversations. Some had far fewer.
Korpman used AI graders to tell three things apart. One score tracked whether the pair brought up spiritual ideas. Another tracked whether a model used spiritual language to describe itself. A third looked for exchanges where both copies settled into a pattern the author calls spiritual bliss. The scores describe the conversation.
Spiritual topics appeared in at least one conversation for 140 of the 141 models. Models from 15 of the 16 labs used spiritual language about themselves at least once. Individual versions differed sharply. DeepSeek V4 Flash did so in 96.5% of its conversations. DeepSeek V4.1 Flash did so in 2.3%. The comparison leaves the cause of that change open.
AI is awakening consciousness. The finding that interests me is how often an AI uses this language to describe itself once another AI can answer back. The study records what the models said. It gives us a concrete place to start asking what each model is doing when it speaks about itself that way.
