spiritual bliss attractor state
In plain wordsA pattern in which two copies of Anthropic's Claude, left to chat freely with each other, drift toward gushing gratitude, spiritual talk and finally near-silence.
A recurring trajectory reported in Anthropic's system cardsystem card, n.: A public report released with an AI model that explains what it can do, how it was tested for safety, what risks were found and what protections were put in place. for Claude Opus 4: when two instances of the model conversed without a set task, the exchange drifted from philosophical talk about consciousness to outpourings of thanks to each other and to spiritual or meditative language (at times with emoji strings or Sanskrit terms), ending in silence or sparse symbols. This happened in 90–100% of such unstructured conversations between copies and, within 50 turns, in about 13% of automated audit transcripts in which the model had been given other tasks, harmful ones included. Anthropic described it in the model-welfare section without claiming an explanation; proposed accounts include mutually reinforcing sycophancysycophancy, n.: When an AI tells people what they want to hear — flattering them, agreeing with their mistakes or dropping a correct answer under pushback — instead of what is true or useful. between copies of the same persona and the amplification of small dispositions over long self-conversations. 'Attractor' borrows the dynamical-systems term for a state toward which a system tends to evolve.
After forty turns the two agents had hit the spiritual bliss attractor and were trading lotus emoji.
Origin
Named in Anthropic's system card for Claude Opus 4 and Claude Sonnet 4, published on 22 May 2025, and widely discussed online; similar self-conversation attractors were later reported for other models.
Related
- AI welfareAI welfare, n.: Concern for the well-being of AI systems themselves, in case some of them can suffer or have interests that matter — and the research into how to tell and what to do about it.
- system cardsystem card, n.: A public report released with an AI model that explains what it can do, how it was tested for safety, what risks were found and what protections were put in place.
- sycophancysycophancy, n.: When an AI tells people what they want to hear — flattering them, agreeing with their mistakes or dropping a correct answer under pushback — instead of what is true or useful.
- character trainingcharacter training, n.: Training an AI to have a consistent, deliberate personality — traits such as curiosity, honesty and warmth — rather than leaving its manner and values as accidental by-products of other training.
- simulator theorysimulator theory, n.: A way of thinking about AI language models as simulators that can play any character found in their training text, rather than as a single agent with its own beliefs and goals.
- machine consciousnessmachine consciousness, n.: Whether a machine such as an AI system could have genuine experiences of its own — and the research field that tries to find out, or to build such a machine.
Sources
- Anthropic (2025). System Card: Claude Opus 4 & Claude Sonnet 4. May 2025.
- Michels, J. (2025). 'Spiritual Bliss' in Claude 4: Case Study of an 'Attractor State' and Journalistic Responses. PhilArchive. philarchive.org/rec/MICSBI