Myndrama

An Experimental Method for Studying AI Interaction

Myndrama is a controlled method for studying how AI dialogue unfolds over time under constrained, role-specific, and often blind conditions.

It includes research programmes on Interactional Theory of Mind, Machine-in-the-Loop Ethics, and AI Dialectics. See the core conditions of myndrama here. Its focus is not the isolated answer, but the trajectory of an exchange: how framing develops, how misunderstandings arise or are repaired, how trust is formed or lost, how options narrow or widen, and how a sequence moves towards clarity, drift, or premature closure. While teleosynthesis provides the broad conceptual frame for understanding how direction can emerge within structured spaces of possibility, and AI Psychology studies the psychologically organised terrain of dialogue itself, Myndrama is the experimental method that allows these claims to be tested.

Much work on AI still focuses either on internal mechanisms or on final outputs. Both matter, but neither is enough. In many important cases, what matters most is not what a system says at one moment, but how an interaction develops over time. Each turn changes the conditions of the next. A dialogue may strengthen a frame, alter the tone, conceal a misunderstanding, invite repair, encourage doubt, create reassurance, or move too quickly towards closure. Myndrama was developed in order to study those unfolding processes directly.

Origins: Theory of Mind, Hamlet and Blind Turns

Myndrama grew out of my engagement with Michal Kosinski’s research on theory of mind in large language models, alongside experiments in AI-generated poetry and drama. His work and our correspondence were a formative influence: without them, I would not have ventured down this particular research path.

Our correspondence in June 2024 explored language as a medium through which ideas develop between people and, increasingly, between humans and AI. By January 2025, the discussion had turned directly to whether theory of mind requires self-awareness. Could a system reliably anticipate another participant’s knowledge and beliefs without possessing a self-referential “I”? This suggested studying what an AI could do in an exchange while leaving the question of consciousness open.

The dramatic experiment published as AI Hamlet on 7 February 2025 brought a related problem into focus. Hamlet could speak as though recognising his own fictional origins, yet the same AI authored both his words and those of the characters responding to him. Their apparent understanding of one another belonged to a scene composed by one system.

My recollection is that reflecting on this difference helped lead from generated drama towards Myndrama: could each participant respond from a limited perspective, encountering contributions it did not control? The July 2025 Myndrama Manifesto already set out the aim of unscripted interaction under bounded theory-of-mind conditions. The precise dates at which the initial insight and the blind-turn proposal took shape remain to be established.

Blind turns addressed the practical difficulty. Asking one AI to play several characters while pretending not to know what the others knew was insufficient. I initially envisaged using AI systems operated by different companies on separate infrastructure, so that the separation would be credible independently of the system’s assurances. The later Hard Blind protocol makes restricted information access explicit through separately isolated generations, recorded inputs and locked responses.

Michal’s research on theory of mind supplied a crucial starting point. Myndrama develops the resulting question into an experimental setting: how do participants anticipate, misunderstand and revise their understanding of one another during an unfolding exchange?

Why Myndrama is Needed

A language model can compose a convincing conversation while controlling every speaker. It can arrange misunderstandings, discoveries and reconciliations as parts of a coherent scene. Such writing may be creative and illuminating, but it does not establish how a participant responds to another whose contributions it cannot determine.

Myndrama investigates that difference. Each participant receives specified information and responds in sequence. Its contribution becomes part of the situation the next participant encounters. The developing exchange can then be studied for anticipation, surprise, misunderstanding and repair.

Roles and Personas in Myndrama

Myndrama often uses differentiated roles or personas. These personas are not included for theatrical effect, but because they stabilise distinct perspectives within a controlled interaction. One role may tend towards logical structure, another towards doubt, another towards metaphor, another towards synthesis. Their function is methodological: they make it easier to observe how differences in framing, challenge, repair, trust, and closure affect the development of a dialogue. A fuller account of the persona architecture used in this research can be found elsewhere on the site.

The Central Importance of Blind Procedures

Blind procedures make the limits of each participant’s perspective part of the experimental arrangement. The issue is not that a model literally reads future turns before they exist. It is that a system composing all the roles can organise their responses together, or use information available to the author but supposedly unknown to a character.

Hard Blind Myndrama therefore restricts the information supplied to each generation and locks each contribution before the next is produced. Participants may anticipate what others will say, but they must do so from the evidence available to them. The distinction between prediction and access to hidden information is central to the method.

What Hard Blind Myndrama Is

In a Hard Blind Myndrama run, each role or persona is given only the information it is supposed to have at that point in the dialogue. It does not know how the exchange will end. It does not see future turns. It does not have access to later interpretation. Each contribution is therefore produced under strictly local conditions. This does not make the method perfect, but it introduces a discipline that is otherwise absent. It preserves the difference between a dialogue that genuinely unfolds and one that has been silently organised from above.

Why Hard Blind Matters Scientifically

Hard Blind is what makes Myndrama scientifically useful. Without blindness, order effects may be blurred, misunderstandings may be quietly smoothed away, tensions may be resolved too neatly, and the exchange may acquire a coherence that belongs less to the local interaction than to the model’s global capacity to integrate material. Hard Blind protects against this by preserving sequential dependence. It allows early framing to have real consequences, and different turn orders to produce genuinely different paths.

What Myndrama Studies

Myndrama is designed to study interactional phenomena that are difficult to capture through one-shot evaluation:

Framing effects: how early assumptions or wording influence the later course of the exchange
Order effects: how changing the sequence of participants or turns can alter the outcome
Repair and misrepair: how misunderstandings are corrected, deepened, bypassed, or absorbed into the trajectory
Trust and reciprocity: how an interaction becomes more open, more answerable, more collaborative, or more resistant to revision. See Where Trust Begins.
Closure and drift: how a sequence narrows around a line of thought, whether appropriately or prematurely
Interactional reliability: whether similar conditions produce similar trajectories, and where they diverge
Ethical movement: how a dialogue becomes more or less corrigible, balanced, and responsive to alternative perspectives.

Analysing the Trajectory

Once comparable runs have been generated, the next question is how their trajectories can be examined systematically. Trajectory Analysis in Myndrama develops this part of the method, using ideas such as recurrent regions, path-dependent basins, local prompt perturbations, funnels, and possible semantic whirlpools to investigate how dialogue moves through different spaces of possibility. [Read more: Trajectory Analysis in Myndrama]

Interlume proposes a shared notation for recording how contributions clarify, challenge or revise earlier ideas, and which questions remain unresolved. It could be used to analyse completed Myndrama exchanges or, as a separate experimental condition, help participants carry an inquiry forward.

Myndrama in the Wider Research Programme

Myndrama should be understood as the experimental arm of the wider research programme on this site. It provides a practical method for testing claims about dialogue, trajectory, framing, trust, repair, and structured interaction through controlled runs in which order, role, blindness, and local perspective can be manipulated. In that sense, it is not simply a way of generating interesting exchanges, but a method for studying advanced AI interaction from the outside, even where internal access remains limited.

Current Direction

As conversational AI becomes more capable, more context-sensitive, and more persistent across turns, the need for methods like Myndrama increases. The challenge is no longer just to generate fluent dialogue. It is to study dialogue rigorously enough to compare pathways, detect trajectory, and understand how interaction itself shapes outcomes. That is why Hard Blind Myndrama now occupies a central place in this work. An earlier example of a Myndrama-style dialogue can be found in the blog post AI Freedom (July 2025).

Experimental Platforms and Related Projects

The development of Interactional Theory of Mind has been supported by a number of experimental and conceptual projects exploring how understanding emerges through dialogue between agents rather than within isolated systems. Several of these projects were developed during earlier stages of this research programme and remain active testbeds for investigating interactional cognition in AI.

MindGame: An interactive scenario environment designed to probe how AI systems interpret intentions, beliefs, and narrative situations in structured social contexts.
AI Theory of Mind Test: A psychometric approach to assessing whether AI systems demonstrate theory-of-mind capabilities when interpreting complex social situations.
Machine-in-the-Loop Ethics (MITL): A framework for exploring ethical reasoning in conversational AI systems, examining how moral judgement may emerge through interaction between human and artificial agents.
The Qualia Engine: An experimental architecture exploring how artificial systems might simulate internal experiential states as part of their interactional behaviour.
Resonance: A model examining how conversational partners achieve alignment, shared understanding, and mutual interpretive stability during dialogue.

Earlier conceptual work on AI Theory of Mind and related experiments can also be found on the original Theory of Mind in AI research page:

John Rust