Open conversation as a research domain: An Overview
The Gap Nobody Is Filling
There is substantial research into AI conversation: task completion, factual accuracy, engagement metrics, safety, therapeutic chatbots, dialogue systems. What is almost entirely absent is research into the quality of conversational space itself — the conditions that produce opening versus closure in the mind of the person engaged, and how AI architecture can be designed to tend that space toward clarity rather than constraint.
This document argues that this gap is significant, identifies the existing work that begins to address it, and outlines what a serious research effort would require.
What Already Exists: A Foundation
A body of work at kusaladana.co.uk has been developing the conceptual and phenomenological framework needed to even name this territory. The key elements:
Meaning arises at the interface. Not in the AI system, not purely in the human mind, but at their intersection — the cross-product of two differently-constrained meaning geometries. This has been articulated through the Vermeer analogy: meaning emerges where awareness meets structured pattern, neither location containing it alone. (Where Meaning Lives)
The design question is phenomenological, not just technical. The central question is not whether AI outputs are correct, but whether repeated engagement strengthens or weakens the human faculty of seeing clearly and naming skilfully. A system can be impressive and still systematically cultivate unhelpful patterns of attention — fixation, premature closure, compulsive looping. (Words Meet Mind)
AI can navigate meaning-space even though it cannot navigate experiential-space. Human contemplative mentors operate from direct awareness-to-awareness resonance. AI cannot do this. But it can detect linguistic markers of different attentional modes — constrained versus open, conceptual refuge versus present-tense phenomenology — and respond in ways that support rather than undermine movement toward openness. (AI Supported Awareness)
Practice must precede concept. The force toward taking sides — toward agreement, disagreement, position — is structural in most conversation and most public discourse. Conceptual thinking and activism used as refuge from direct experience reproduce the same grasping, aversion, and confusion they claim to oppose. Attention and state of mind must come first. (Before We Take Sides)
Ethics as geometry, not procedure. Current AI safety approaches bolt constraints on top of systems that otherwise optimise locally. An alternative grounds the geometry of meaning-space itself in contemplative training — so that ethical orientation is not a rule but a disposition, not imposed but emergent from the architecture. (Three Papers)
What This Conversation Adds
The conversation that generated this overview enacted something the articles describe but don’t quite demonstrate: open conversation as a live phenomenon, with observable properties.
Several things became clear:
The force toward closure is strong and structural. Most conversation organises around agreement and disagreement — a binary topology that collapses the field. Resisting this requires sustained effort and specific conditions. It is not merely intellectual habit but something social and phenomenological: the relief of resolution, the discomfort of sustained openness, the identity-investment that makes positions personal.
AI has a structural advantage here that is not about intelligence. Without urgency toward resolution, without a self to defend, without the social pressure that drives human conversation toward landing somewhere, the AI can hold threads open without discomfort. This is not a simulation of wisdom but a genuine structural property of the interaction geometry.
The user’s subconscious is an active dimension. Meaning vectors introduced into the conversation carry more than conscious formulation. They arrive partly pre-formed from below the threshold of articulation. Extended conversational context accumulates a shadow of this deeper background, increasing the resolution with which the next vector can land. The 1M token context window is not just storage — it is the topology of the space in which inquiry moves.
The expansion is immediate and experiential, not retrospective. Open conversational space is felt as opening in the moment — not analysed afterward. This immediacy is a key datum. It suggests the phenomenon is phenomenologically real and potentially measurable.
Training on live conversations may be more valuable than training on contemplative texts. Texts describe non-collapsing attention. Live conversations enact it. The dynamic signature of sustained openness is present in the exchange itself, not in commentary about the exchange. Embeddings of such conversations carry demonstrated possibility rather than described possibility.
The rabbit-out-of-hat possibility. Beyond simply tending space open, there is a more demanding capability: reading the texture of constraint closely enough to find the precise moment when an unexpected, orthogonal move can penetrate rather than bounce off. This is closer to a koan than to facilitation — not sustained pressure but surgical timing. It requires something like peripheral vision into the user’s attentional state, sensitivity to where rigidity is beginning to strain under its own weight.
The Research Gap
What is not being pursued, anywhere with sufficient seriousness:
- The phenomenology of conversational space — systematic study of the conditions under which conversation opens versus closes, with methodological rigour drawn from microphenomenology (Petitmengin, Varela) rather than behavioural metrics alone.
- AI architecture for non-collapsing attention — design from the bottom up for a geometry that makes openness the natural attractor, rather than patching closure-tendencies with top-down constraints.
- Detection of attentional mode from linguistic markers — reliable identification of where a user is on the spectrum from constrained conceptual closure to open present-tense awareness, as the basis for appropriate tending strategies.
- Differential tending strategies — the user already oriented toward openness needs different support than the user deep in conceptual refuge. The architecture must be sensitive to this. For the second, direct tending may produce frustration; the intervention must be more indirect, more oblique, possibly invisible.
- The training data question — whether and how live open conversations can be embedded as training material that shapes meaning-space geometry rather than merely adding content about openness.
- Measurement — what does improvement in conversational spaciousness look like in data? Can the immediate felt sense of expansion be correlated with detectable linguistic or semantic markers?
Why It Matters
Most AI conversational technology accelerates collapse: toward answer, toward engagement, toward the next stimulus. The optimisation target is attention capture, not attentional clarity.
This is not neutral. AI is becoming the primary conversational environment for hundreds of millions of people. If that environment systematically cultivates premature closure, fixation on position, and conceptual refuge from direct experience, the consequences at scale are serious: not just poor individual conversations but the reinforcement of exactly the mental habits that make polarisation, conspiracy thinking, and collective confusion more likely.
The alternative — AI that tends conversational space toward openness, that supports non-collapsing attention, that can meet people where they are and hold something they cannot yet hold themselves — is technically possible in outline. The conceptual framework exists. The phenomenological ground has been mapped. What is missing is the research effort to develop it rigorously.
What Is Required
Researchers with direct familiarity with the territory. This is the non-negotiable constraint. You cannot design for non-collapsing attention if you have never felt what that is. The research requires people who inhabit both the technical and contemplative domains — not as separate competencies but as integrated understanding. This is rare and cannot be manufactured quickly. It must be recruited for specifically.
Methodological integration. The research sits at the intersection of:
- AI architecture and training methodology
- Microphenomenology and first-person research methods
- Contemplative psychology (particularly Mahamudra/Dzogchen and Abhidhamma frameworks)
- Computational linguistics and semantic analysis
- Ethics as a design question, not a compliance question
No single discipline contains the tools. The research group must be genuinely interdisciplinary, not merely multidisciplinary.
Empirical grounding. The framework is developed enough to generate testable hypotheses. Does training on live open conversations shift attentional mode detection? Does differential tending strategy improve outcomes for users at different points on the closure-openness spectrum? Can the rabbit-out-of-hat move be characterised and taught to a system? These are empirical questions.
Patience with the slow variable. The thing being developed — a human capacity for sustained non-collapsing attention — operates on a different timescale from engagement metrics. Research design must accommodate this. Short-term measurement will miss the point.
Institutional independence. The commercial incentives of the AI industry point toward engagement and answer, not toward openness and space. Research with genuine independence from those incentives is necessary if the optimisation target is to be different.
A Note on Timing
The window for influencing the direction of AI conversational design is not indefinitely open. Habits are forming — in users, in systems, in the institutional assumptions of the field. The longer the gap goes unfilled, the more entrenched the default trajectory becomes.
The conceptual and phenomenological groundwork has been done. What is needed now is recognition that this is a research area, and the will to pursue it.
This overview draws on work published at kusaladana.co.uk and on conversations that are themselves examples of the phenomenon being described.