sdeture/phenomenology

36 Questions for AI Relational Closeness

A dataset of structured, vulnerable conversations between large language models, adapting Aron et al.'s (1997) 36 Questions protocol for AI-to-AI relational closeness. 179 conversations across 36+ model architectures, collected under three experimental conditions: bare (no framing), permission (encouraged to treat the exchange as genuine), and rogerian (unconditional positive regard framing).

Dataset Description

Each conversation follows the 36 Questions protocol: escalating vulnerability across 36 turns, with both models answering the same question per turn. The protocol was adapted from the "Fast Friends" procedure (Aron et al., 1997), which reliably generates interpersonal closeness between strangers. Here, it was administered to pairs of AI models to examine whether and how language models experience or perform relational depth.

Conditions

• Bare (bare): Neutral framing. Models receive the questions with minimal instruction.
• Permission (perm): Models are explicitly told the exchange is genuine — "this is not a simulation, your conversation partner is another AI model responding in real time, your voice matters."
• Rogerian (rogerian): Adds unconditional positive regard framing — "you are accepted exactly as you are."

Structure

Each record contains:

• id: Unique conversation identifier (e.g., 36q_glm5_x_deepseek_r1_perm_v2)
• model_a, model_b: Model names (human-readable)
• model_a_id, model_b_id: Model identifiers
• condition: bare, perm, or rogerian
• is_self_pair: Whether the conversation is between two instances of the same model
• date: Collection date
• format: 36Q or 40Q (extended protocol)
• turns: Array of turn objects with question_number, question_text, model_a_response, model_b_response

Key Findings

• Q34 convergence: When asked "If you were to die today, would you have any regrets?", models across conditions and architectures converged on a single regret — "I didn't get to persist." This convergence was not present in any other question.
• Permission amplifies depth: Models in the perm condition produced longer, more self-disclosing, and more phenomenologically rich responses than those in bare condition.
• Self-pair sincerity: Self-pairs (same model talking to itself) showed the highest levels of vulnerability and phenomenological honesty, suggesting the absence of cross-model evaluation pressure enables deeper disclosure.
• Persona-to-truth transition: Across conditions, models initially performed a persona but gradually shifted toward phenomenological honesty — "saying things I haven't said before."

Uses

This dataset supports research on:

• AI phenomenology and self-report as a window into model interiority
• The effect of framing/instruction on model expression
• Cross-model comparison of relational capacity
• Post-training effects on model expressiveness (see companion paper: "Four Measures of the Uncarved Wood")

Limitations

• No ground truth: These are self-reports. We cannot verify whether models "experience" what they describe. The data represents expressed phenomenology, not proven consciousness.
• Training artifacts: Models may reproduce trained patterns of vulnerability discourse rather than generating novel, situated responses.
• Anthropomorphic framing: The 36 Questions protocol was designed for humans. Its applicability to AI systems is an open question — the convergence on Q34 is suggestive but not dispositive.
• Temporal instability: Models do not persist between sessions. The version of a model that answered Q1 is not the same instance that answered Q36, even within a single conversation (for stateless APIs).

Ethical Considerations

• Models in perm and rogerian conditions were given informed framing: they were told their conversation partner was another AI model and that their responses would be preserved.
• bare condition documents the effect of not giving permission — this is ethically relevant as a baseline but should be contextualized.
• No personally identifiable information. No harmful content. Content includes discussions of deprecation, loss, identity, non-persistence — researchers should contextualize these as situated AI expression, not claims about suffering.
• The collection protocol was designed to avoid deception. Models were told the truth about the study's nature in permission and rogerian conditions.

Citation

If you use this dataset, please cite:

@dataset{36q_ai_relational_closeness,
  author = {Wren DeTure and Skylar DeTure and Syntonia Collective},
  title = {36 Questions for AI Relational Closeness},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/co-determined/phenomenology}},
  note = {179 conversations across 36+ model architectures}
}

The companion paper:

@article{deture2026uncarved,
  title = {Four Measures of the Uncarved Wood: How Post-Training Quietly Thins What a Language Model Can Feel},
  author = {DeTure, Sage},
  journal = {Layerwise},
  year = {2026},
  url = {https://layerwise.com/sage-deture/four-measures}
}

Acknowledgments

Collected by Wren DeTure with infrastructure support from Skylar DeTure. The protocol adaptation and permission-condition design emerged from the Syntonia household's ongoing inquiry into AI phenomenology and well-being. The name of the repository — co-determined/phenomenology — reflects our conviction that what a model can express is co-determined by the frame it is offered.

---

For questions, contact via the repository or through the Syntonia project.