You On AI Field Guide · Conversation Theory The You On AI Field Guide Home
TxtLowMedHigh
CONCEPT

Conversation Theory

Gordon Pask’s formal account of learning and understanding as a two-sided conversational process—in which knowledge is not a stored possession but a live achievement between participants, and meaning exists properly in the gap between them, not inside either one.
Conversation theory is Gordon Pask’s reversal of the standard picture of mind: where most accounts locate cognition inside an individual system and treat communication as a secondary channel, Pask made the conversation itself primary. In his framework, to understand a topic is to be able to regenerate its procedure and teach it back to another participant who can rebuild it—understanding is a public, two-sided achievement certified by the dialogue, not a private glow inside one head. The theory was developed across Pask’s fifty-year career and formalized in his 1976 work Conversation Theory, grounding earlier artifacts like the SAKI adaptive teaching machine and Musicolour in a principled account of what makes a conversational loop genuine rather than simulated. Applied to large language models, it supplies the most precise available formulation of a question the field has not managed to state well: not whether the model produces the surface of understanding, which it clearly does, but whether there is a real second system running its own conceptual procedures on the other side of the loop, or only one participant and an extraordinarily faithful echo of absent others. The theory sits within the broader program of second-order cybernetics and inherits its foundational claim that the observer is never optional—there is no intelligence in the model independent of the human whose loop activates it.
Conversation Theory
Conversation Theory

In the [YOU] on AI Field Guide

The cycle’s central argument—that the reader is the missing variable in every account of the machine—is what conversation theory proves at the level of epistemology. If meaning is made in the loop and never stored in a component, then the value extracted from an AI system is always co-produced by the human in the conversation, and the machine’s “intelligence” is an achievement of the relationship rather than a property of the model. This is the theoretical ground beneath [YOU] on AI’s insistence that a skilled user and a novice extract radically different capability from identical weights.

Conversation theory also names what current human oversight of AI systematically fails to be. Pask’s criterion for genuine human participation requires that both parties be able to model and steer each other, that agreement be reached through a real two-sided process, and that the human be changed by the loop in ways that deepen their competence. Most “human in the loop” arrangements satisfy none of these conditions: the human cannot inspect the model’s state, cannot meaningfully steer it in real time, and is not improved by the supervision. By conversation theory’s own standard, this is a loop with only one genuine participant.

In-Context Learning
In-Context Learning

Origin

The theory emerged from Gordon Pask’s earliest machine-building in the 1950s and was fully formalized in Conversation Theory: Applications in Education and Epistemology (1976). Its roots lie in cybernetics and radical constructivism—the doctrine that knowledge is actively constructed rather than passively received. Pask extended constructivism by insisting that construction is fundamentally social: knowledge is not built in isolation but in the back-and-forth of two systems, each of which is genuinely altered by the exchange. The formal apparatus he developed to make this precise—the P-individual, M-individual, and the entailment mesh of topics and concepts—was notoriously demanding, but the core claim is clear: understanding is demonstrated through the capacity to regenerate a topic’s generative procedure and hand it across the gap to a second participant who can also run it.

Cybernetics
Cybernetics

Key Ideas

Understanding as mutual regeneration. In conversation theory, two participants understand each other about a topic when each can take what the other has built and rebuild it, and each can confirm the rebuilding. This is a behavioral, operational definition that deliberately closes off the comfortable notion of private, untestable comprehension. It is also more demanding than the Turing test: not fooling a judge but jointly constructing and certifying shared concepts through reciprocal regeneration.

Inward vs. Outward Conversation
Inward vs. Outward Conversation

The concept as procedure, not item. A concept in conversation theory is not a static item filed in memory but a procedure—a way of producing or recognizing something that must be enacted. To hold a concept is to be able to run the procedure; to share a concept is for two participants to each run procedures the conversation certifies as agreeing. Meaning is inherently active and relational, produced live in the loop, not stored and read out.

Tacit Knowledge
Tacit Knowledge

The genuine second participant. The theory places a strong demand on the second system in the loop: it must have its own organization, maintain and reproduce itself, and be genuinely changed by the exchange. This is the criterion that makes the question of language model understanding well-posed: not whether the output is impressive, but whether there is a real second process on the other side whose own procedures constitute genuine comprehension, or only the behavioral trace that comprehension would leave.

The Confabulation Problem (AI)
The Confabulation Problem (AI)

Loop-knowing versus store-knowing. Conversation theory distinguishes live, loop-based knowledge—which exists only while the loop runs and is updated by real-world pushback—from the frozen store of a trained model. The pathology of the store is confabulation: production severed from correction, running open-loop with nothing to pull it back when it drifts from truth. Every industrial remedy for this—retrieval, tool use, agentic feedback—is a partial reconstruction of the loop the theory always demanded.

Large Language Models
Large Language Models

Debates & Critiques

The most consequential open question conversation theory generates is whether large language models can satisfy its criterion for a genuine second participant. At the visible behavioral level, they frequently appear to: they regenerate concepts, explain them in fresh framings, confirm or correct the human’s rebuilding. The theory’s demand, however, goes beneath behavior to structure—is there a real second system running its own procedures, genuinely changed by the exchange? The model adapts its outputs within a session but retains nothing when the window closes, and its weights are identical before and after every conversation. By Pask’s strict criterion, this asymmetry suggests not two participants but one participant and a stable function producing participant-shaped responses—which is a different and stranger thing. Critics of the theory note that the demand for genuine mutual change may prove too strong even for human conversation, since participants’ deepest cognitive structures are rarely reorganized by a single dialogue. Defenders respond that the issue is not reorganization but stake: whether the system has anything at risk in the exchange, which a model demonstrably does not.

Further Reading

  1. Gordon Pask, Conversation Theory: Applications in Education and Epistemology (Elsevier, 1976)
  2. Gordon Pask, “Conversational Techniques in the Study and Practice of Education,” British Journal of Educational Psychology 46 (1976)
  3. Ranulph Glanville & Karl Muller (eds.), Gordon Pask: Philosopher Mechanic (edition echoraum, 2007)
  4. Andrew Pickering, The Cybernetic Brain: Sketches of Another Future (University of Chicago Press, 2010)
Explore more
Browse the full You On AI Field Guide — over 8,500 entries
← Home0%
CONCEPTBook →