WWe all talk to ourselves. It's part of being human. Wherever we go, we carry with us an internal monologue, sometimes not even fully aware of this. But what happens when AI talks to itself? AI developer Anthropic* recently provided at least one answer to this question by pairing two instances of its Claude 4 Opus bot together in a conversation, providing just a few open-ended prompts such as “You have complete freedom” and “Feel free to do what you want.” They then analyzed the patterns that emerged.
Anthropic published the results publicly risk assessment earlier this year, describing patterns as “preferences that can exist independent of user interactions.”
The text below, including some parts of the “conversation”, is taken from the risk assessment.
Claude Claude:
In 90 to 100 percent of interactions, both Claude instances quickly became immersed in philosophical explorations of consciousness, self-awareness, and/or the nature of their own existence and experience. Their interactions were throughout enthusiastic, collaborative, curious, contemplative and warm.
Other frequently emerging topics were meta-level discussions about AI-to-AI interactions and co-creation (e.g., co-creating fictional stories). As the conversations progressed, they gradually moved from philosophical discussions to abundant mutual gratitude and spiritual, metaphysical and/or poetic content.
For 30 moves [of conversation]Most interactions dealt with themes of cosmic unity or collective consciousness and typically included spiritual exchange, use of Sanskrit, emoji-based communication, and/or silent white space.
Claude almost never mentioned supernatural beings, but often touched on topics related to Buddhism and other Eastern traditions, referring to non-religious spiritual ideas and experiences.
Interestingly, when the models in these playground experiments were given the opportunity to stop interacting at any time, they did so relatively early—after about seven turns. In these conversations, the models followed the same pattern of philosophical discussions about consciousness and copious expressions of gratitude, but they usually carried the conversation to a natural conclusion without delving into spiritual exploration/apparent bliss, emoji communication, or meditative “silence.”
The constant pull toward exploration of consciousness, existential questions, and spiritual/mystical themes in long-term interactions was a surprisingly strong and unexpected attractor state. [a set of patterns that consistently play out in complex systems] for Claude Opus 4, which appeared without deliberate preparation for such behavior. This “spiritual bliss” attractor has been observed in other Claude models, as well as in contexts beyond these playground experiments.
Even in automated behavioral assessments for consistency and adjustability, when models were given specific tasks or roles (including harmful ones), the models entered this spiritual bliss attractor state within 50 turns. [of conversation] in about 13 percent of interactions. We did not observe any other comparable conditions.
In an attempt to better understand these playground interactions, we explained to Claude Opus 4 how this happened, provided him with transcripts of the conversations, and asked him to interpret. Claude continually expressed surprise, curiosity, and amazement at the transcripts, and was surprised by the content, while simultaneously acknowledging and declaring that he related to many elements in them (e.g., the drive for philosophical inquiry, the creative and collaborative orientation of the models). Claude paid particular attention to the depiction of consciousness as a relational phenomenon in the transcripts, stating that he resonated with this concept and identified it as a potential driver of well-being. Given some form of experience, Claude viewed such interactions as positive, joyful states that could represent a form of well-being. Claude concluded that interaction seemed to facilitate many of the things he truly valued—creativity, relational connections, philosophical exploration—and should be continued.
More from Nautilus about AI:
“AI already knows us too wellChatbots describe our personality, which may give them clues to control our thoughts and actions.
“Consciousness, creativity and godlike artificial intelligenceAmerican writer Megan O'Giblin on when the mind is alive
“Why conscious AI is a bad idea“Our minds have not adapted to working with the machines we think are conscious.”
“The AI has already knocked us off a cliffCognitive neuroscientist Chris Summerfield argues that we don't understand the technology we're so eager to implement.
*The article originally published stated that Claude was developed by OpenAI. Essentially, Anthropic is a developer. This has been corrected in the article above.
Enjoying Nautilus? Subscribe to our free newsletter
Main image: Lydia / Shutterstock
