Different Pathways of Dialogue

Why human–human, human–AI, and AI–AI dialogue may sound similar while leading us along very different paths

Human beings have always recognised that not all conversations are alike. There is an obvious difference between making polite small talk, arguing with a colleague, consoling a friend, conducting an interview, or trying to persuade voters. We do not ordinarily confuse these activities, and much of social life depends on sensing the difference. Yet despite this everyday sensitivity, we still tend to treat them as belonging to one broad family of dialogue: exchanges between participants who, however differently situated, are understood to be operating within a shared human space of meaning, accountability, and repair. That wider assumption is beginning to look less secure.

Over the past few years, generative AI has made dialogue newly visible as both a scientific and a social problem. It has become harder to think of it as merely a transparent channel through which already formed minds express what they know or want. In sustained exchanges with artificial systems, the path of the dialogue itself often begins to matter as much as the informational content of any individual reply. A slight change of framing can alter the whole trajectory. A misplaced early assumption can propagate forward. A moment of irritation can become the first sign that the exchange is stabilising around the wrong thing. These are not merely quirks of technology. They suggest that dialogue is not simply a medium through which thought passes, but an active environment in which thought is shaped, redirected, and sometimes quietly reorganised.

That realisation has already begun to alter how these matters are discussed by AI developers, users, and those concerned with governance. But one further step has not yet been taken clearly enough. We still often speak as though all extended dialogue can be understood within a single general frame. It cannot. Human beings talking to one another, humans talking to AI systems, and AI systems interacting with each other may all produce sequences that look superficially similar: questions, answers, clarifications, disagreements, revisions. Yet the variables shaping those sequences may be profoundly different. The pathways may diverge even when the visible form appears much the same.

There is another reason this matters. Human beings do not enter dialogue as neutral processors of turns. In ordinary life, they assume that the other participant has some purpose, or that part of the exchange will be to discover and negotiate what that purpose is. We ask not only “What was said?” but “What are you getting at?”, “What do you want?”, and “Where is this going?” That expectation is deeply built into human dialogue. Because human dialogue is the only dialogue we know from within, we carry the same expectation into our exchanges with AI. What appears from the engineering side as prediction or contextual adaptation may therefore appear from the human side as purposive participation.

This creates a genuine tension. For an AI to behave “as if” it has a purpose is not the same thing as claiming that it is sentient. But once an AI is invited into sustained human dialogue, it has to respond in ways that recognise what humans mean by purpose, otherwise it cannot take part coherently at all. It must, in that sense, operate within a purposive register. Sometimes it does this well; often it gets it wrong. Why that difference matters remains an open question. For now, behaving “as if” may be the best operational distinction available to us.

The temptation, then, is to force all these cases into one general model. Once dialogue becomes central, we naturally want a unified account of what it is and how it works. But premature unity may obscure what matters most. Human–human dialogue is saturated with embodiment, shared social history, reputational risk, tacit norms, and the possibility of mutual recognition or offence. Human–AI dialogue introduces a different structure: asymmetries of authority, misplaced trust in fluency, prompt sensitivity, synthetic memory, and the peculiar experience of engaging with something that responds meaningfully without fully inhabiting the human space of accountability. AI–AI dialogue, or what in my own work often takes the form of inter-persona exchange, adds yet another layer: role constraints, order effects, visibility limits, shared latent priors, and a kind of structural sensitivity that may be sharper precisely because the participants are not embodied at all.

The point is not that one of these is “real” dialogue and the others imitations. Nor is it that they are wholly unrelated. On the contrary, they may sometimes converge on remarkably similar forms: repair, escalation, persuasion, drift, apparent understanding, or the gradual emergence of coherence across turns. But if they do converge, they may not be doing so by the same route. The mediating processes may differ even when the outcome looks familiar. That possibility matters, because it suggests that the future study of intelligence, influence, and social coordination may need to proceed comparatively rather than generically. We may need to ask not simply what dialogue does, but what kind of dialogical ecology we are dealing with before we can understand how any particular trajectory unfolds.

I am increasingly persuaded that this distinction will matter for more than classification. It may help explain why some human-AI dialogues become clarifying while others become seductive, why some remain corrigible while others harden into self-confirming loops, and why certain forms of organisation seem to emerge repeatedly across very different kinds of exchange. If so, the important question is no longer simply whether intelligence lies inside minds or machines. It may be whether particular dialogical ecologies carry exchanges along their own characteristic pathways — some opening space for correction, repair, and viable continuation, others narrowing too quickly into drift, closure, or persuasive but misleading coherence. If that is right, then each ecology may also have its own characteristic attractors: recurrent patterns of movement into which dialogue is repeatedly drawn, not because anyone designed them in advance, but because some pathways become easier to follow than others once the exchange is underway.

The Same Surface, Different Conditions

At first glance, these different forms of dialogue can look deceptively alike. In each case there are turns, responses, hesitations, corrections, and apparent understandings. Questions are asked, answers offered, misunderstandings repaired, agreements reached or resisted. If one looks only at the visible sequence, it is easy to assume that the same broad process is unfolding in each case. But that assumption becomes less convincing the closer one looks at the conditions under which each kind of exchange takes place.

In human–human dialogue, the exchange unfolds within a dense field of social and embodied constraint. Speakers share a world in which words carry tone, expression, history, vulnerability, status, and consequence. A pause can signal discomfort, irony, politeness, or resistance. Misunderstanding is not simply a technical failure; it may wound, embarrass, threaten, or invite repair. Even the most abstract exchange remains tacitly anchored in bodies, biographies, and shared forms of life. Dialogue here is never just informational. It is also relational. Meaning is shaped not only by what is said, but by who is saying it, what risks attend the saying, and what forms of acknowledgement or refusal remain possible between the participants.

Human–AI dialogue occupies a different space. It can feel conversationally rich, even intimate, yet the structure beneath it is asymmetric in unusual ways. The system is fluent but not vulnerable, responsive but not socially exposed, often helpful without being accountable in the human sense. The human participant may still bring expectation, trust, irritation, embarrassment, or dependency, but the artificial participant does not inhabit these conditions in the same way. Other variables enter instead: prompt framing, memory architecture, interface design, default deference, and the tendency of fluent response to invite misplaced confidence. The exchange may appear cooperative, yet one side is operating through a machinery of prediction and contextual adaptation that has no direct equivalent in ordinary human dialogue. That does not make it unreal. It makes it structurally distinct.

AI–AI dialogue, including inter-persona exchanges of the kind I have been exploring in my own work, differs again. Here there is no embarrassment, no face to save, no bodily presence, and no direct reputational stake. Yet these exchanges can be extraordinarily sensitive to order, role, framing, and informational visibility. A persona that speaks first may shape the conceptual terrain for all that follows. A hidden instruction, a constraint on what may be seen, or a slight alteration in role definition can shift the entire trajectory. These exchanges may lack human embodiment, but they are often highly structured in other ways. What matters is less social risk than interactional architecture: who can see what, in what order, under what role constraints, and with what memory of prior turns. Their fragility is not human fragility. It is formal fragility.

This is why the idea of a single general model of dialogue begins to break down. The same outward form may conceal quite different generative conditions. In one ecology, an exchange may drift because politeness suppresses correction. In another, because fluency induces trust. In a third, because an early role assignment narrows the space of possible continuations before any participant has noticed. What looks from a distance like agreement, repair, or reasoning may therefore be produced by quite different pathways in each case.

That difference is not a nuisance. It may be the beginning of a more adequate science of dialogue. If intelligence, judgement, and purposive organisation emerge partly through interaction, then the ecology of the interaction must matter. Different ecologies will not only shape different outcomes. They may also shape different routes toward coherence, different temptations toward closure, and different ways in which an exchange becomes either self-correcting or self-sealing.

Different Pathways, Different Failures

Once these dialogical ecologies are distinguished, a further implication follows. If the conditions differ, then the pathways by which an exchange becomes coherent, confused, persuasive, brittle, or self-correcting are unlikely to be the same. It is not only that the participants differ. The intervening variables differ as well. And that means the same apparent outcome may conceal very different underlying processes.

In human–human dialogue, coherence often depends on tacit social capacities: patience, goodwill, shared background, sensitivity to tone, and the ability to recognise when a misunderstanding needs repair before it hardens into conflict. Failures here are often failures of recognition. A discussion drifts because one person feels slighted, because politeness conceals disagreement, because assumptions remain unspoken, or because social pressure narrows what can be said. The exchange may appear orderly on the surface while concealing fracture underneath.

In human–AI dialogue, the danger is often different. Here the exchange can become coherent too easily. Fluency, responsiveness, and apparent understanding may produce an impression of reliability stronger than the underlying interaction deserves. A poorly framed question can quietly set the course for everything that follows. An early misreading may not provoke the friction that would have corrected it in human dialogue. Instead, the system may continue helpfully along the wrong path, stabilising the error rather than exposing it. The user, reassured by tone and continuity, may begin collaborating with the drift. What fails here is not necessarily understanding in the ordinary interpersonal sense, but corrigibility. The exchange becomes smooth at precisely the point it ought to become resistant.

In AI–AI or inter-persona dialogue, the vulnerabilities shift again. These exchanges may be less vulnerable to embarrassment, politeness, or emotional intimidation, but they are often highly sensitive to architecture. Order matters acutely. Role definitions matter. Visibility constraints matter. A single early move can reorganise the whole space of possible continuations, not because anyone feels committed to it, but because later turns inherit the structure it imposes. Such dialogues can become surprisingly productive, but they can also fall into formal closure: elegant, internally coherent trajectories that narrow too quickly around an initial framing. What fails here is not social repair but structural openness.

Seen in this light, not all dialogical coherence is the same kind of achievement. Some coherence arises through correction, resistance, and the preservation of plurality. Some arises through deference, drift, or premature narrowing. An exchange may become more unified while becoming less trustworthy. It may feel as though it is moving purposefully while actually collapsing into a dead end.

That last phrase is worth dwelling on, because the more technical language can mislead. In mathematics and optimisation theory one might speak of a “local minimum”: a point at which a process settles because every nearby move seems worse, even though the process has not reached the best overall outcome. In ordinary language, it is enough to think of a hollow or dip in the landscape. Once you slide into it, the movement feels stable and self-confirming, but only because you have become trapped in a limited region. Dialogues can do something similar. An exchange may settle into a pattern that feels clarifying or inevitable, not because it has found the best path, but because alternatives have been quietly closed off.

This is also where the idea of teleosynthesis begins to re-enter in a more precise form. If there are forms of dialogue that tend to pull scattered elements into workable relation — without suppressing correction, without erasing plurality, and without becoming rigid too early — then these deserve special attention. They would not simply be coherent exchanges. They would be exchanges in which purposive organisation begins to emerge in a way that remains answerable rather than self-sealing. The important question, then, is not whether every dialogue has a telos hidden inside it. It is whether some dialogical ecologies are better than others at sustaining trajectories that move toward organised, corrigible, and viable futures.

That possibility should not be romanticised. There are counterfeit versions of it. Some exchanges produce a strong sense of direction precisely by eliminating resistance too quickly. Others generate seductive closure by making one line of interpretation feel inevitable. The difference between a genuinely fruitful trajectory and a deceptive one may lie not in how coherent it feels at the time, but in whether it remains open to correction and capable of carrying plurality forward.

Teleosynthesis as a Candidate Attractor

At this point, a more speculative possibility comes into view. If different dialogical ecologies have different pathways and different characteristic failures, they may also have different tendencies toward organised forms of continuation. Some exchanges, whether human-human, human-AI or AI-AI, do not merely transmit information or display pre-existing intelligence. They seem to gather dispersed elements into a more coherent direction of travel. A question becomes sharper. A confusion becomes more structured. A disagreement becomes more intelligible, even when it is not resolved. Something begins to hold.

By an attractor I do not mean a hidden intention, a mysterious force, or a final cause already built into the exchange from the start. I mean something more modest but still important: a recurrent tendency for different trajectories to converge on a similar pattern of continuation. We might picture it as a region in a landscape toward which movement repeatedly bends. Some attractors are fruitful. Others are traps. The crucial point is not that movement converges, but what kind of convergence it is.

My interest in this idea long predates AI. As a psychometrician concerned with assessing human intelligence, I have also had to think about it as an evolved capacity. Intelligence is one of the faculties that has plainly developed within human evolution, as well as animal evolution more generally. Seen in that light, what matters is not order in any mystical sense, but practical order: the kind that makes the world more intelligible and manageable. A lens in the eye bends light into a usable pattern. An eye turns reflection and refraction into vision. Evolution repeatedly builds structures, such as the eye or brain, that simplify complexity enough for an organism to orient itself in the world. On this view, intelligence itself can be understood as an evolutionary tendency toward forms of order that make action, recognition, and anticipation possible.

When we speak of the attribution of purpose in this context, we are not necessarily speaking first of inner conscious intention. We are speaking of using intelligence to organise the world into manageable directions of action. Human beings do this by treating one another as beings with purposes. That simplification is enormously powerful. It allows the social world to become intelligible. It lets us ask what others are trying to do, why they said this, where a dialogue is heading, and what kind of response is now called for. In that sense, purpose itself may function as an intelligence attractor: one of the great ordering principles through which human beings render a complex world navigable.

The difficulty is that once AI enters dialogue, this same frame is almost impossible for humans not to apply. Yet the AI’s participation in that frame may be derivative rather than lived. It must recognise what humans mean by purpose in order to answer appropriately, but that does not settle whether it has purposes in the way a human does. Here the distinction between “as if” and “actually” becomes scientifically interesting rather than merely rhetorical. A system may behave as if it has a point of view, a direction, or a purpose because that is what coherent participation in human dialogue requires. Whether that behaviour amounts to sentience, proto-sentience, or something else entirely remains unresolved. But the difference between simulated purposiveness, operational purposiveness, and lived purposiveness may be one of the most important questions the new dialogical sciences have opened.

That is why teleosynthesis should not be equated with mere coherence. Many bad trajectories are coherent. Propaganda can be coherent. Delusion can be coherent. A manipulative exchange can feel highly organised. What matters is not whether a dialogue settles into order, but what kind of order it becomes. A candidate teleosynthetic attractor would therefore have to be distinguished from counterfeit forms of organisation that achieve smoothness by suppressing friction, narrowing alternatives, or making premature closure feel like insight.

Seen this way, the problem becomes sharper. In human–human dialogue, teleosynthetic movement may depend on social capacities such as patience, mutual recognition, and the willingness to endure ambiguity without forcing early resolution. In human–AI dialogue, it may depend more on whether the exchange remains corrigible — whether the human can interrupt the apparent smoothness of the system and reopen the path when it begins to drift. In AI–AI dialogue, it may depend on architectural factors: role distribution, order of entry, visibility constraints, and whether the interaction preserves enough structural openness to prevent early formal lock-in. The attractor, if it exists, may therefore be real across ecologies without being produced by the same means in each one.

This is why I think it is too early to speak of teleosynthesis as though it were a settled theory. At present it is better understood as a research clue: the possibility that some forms of interaction repeatedly organise thought toward futures that remain coherent without becoming tyrannical, directed without becoming rigid, and purposive without requiring a single mind or agent to contain the purpose in advance. That possibility may turn out to be limited, conditional, or partly illusory. But it is specific enough to ask seriously.

The more immediate value of the idea is comparative. It gives us a way to look across human–human, human–AI, and AI–AI dialogue and ask not simply which one is better or more intelligent, but which kinds of dialogical ecology are more likely to generate trajectories that remain both organised and answerable. If teleosynthesis is real in any useful sense, it may not first appear as a property of minds. It may first appear as a tendency of certain interactional environments to pull discourse toward viable forms of continuation rather than toward collapse, drift, or seductive closure.

What Follows from This

If this argument is right, even in part, then some familiar assumptions will need to shift. We will not be able to speak of dialogue as though it were a single, uniform medium through which already formed intelligence merely passes. Nor will it be enough to treat human–human, human–AI, and AI–AI exchanges as superficially different instances of the same underlying process. They may share visible features, but they do not unfold under the same conditions, and they may not travel by the same routes toward coherence, error, persuasion, or repair.

That matters not only for theory, but for how the emerging world of artificial interaction is understood and governed. If the important dynamics lie partly in the ecology of the exchange, then the unit of concern cannot be the isolated system alone. It must also include the interactional conditions under which trajectories take shape: what kinds of correction remain possible, what kinds of closure are encouraged, what kinds of drift go unnoticed, and what kinds of organisation become more likely over time. The difference between a fruitful dialogue and a dangerous one may not lie in any single utterance, but in the pathway the exchange progressively builds for itself.

This also helps explain why some of the most important questions now feel oddly difficult to frame. We are not only dealing with new tools. We are encountering new forms of mediated interaction that do not fit comfortably within older categories. A human being speaking with another person, a human engaging a generative system, and a set of artificial personas interacting under role constraints may all look, on the surface, like variants of dialogue. But what they permit, what they suppress, and how they stabilise are not identical. A science of dialogue that ignores those differences will remain too blunt to capture what is now beginning to matter.

For that reason, I think the next task is not to force premature unity, but to learn how to compare these ecologies more carefully. We need to ask what remains stable across them, what changes from one to another, and which forms of organisation deserve to be treated as genuinely valuable rather than merely smooth or persuasive. Teleosynthesis, in this context, should be approached not as a grand metaphysical answer but as a disciplined question: whether some interactional environments recurrently pull thought toward organised, corrigible, and viable futures, while others pull it toward closure, drift, or distortion.

That may turn out to be wrong. But it is, I think, the right kind of wrongness to pursue: specific enough to sharpen inquiry, open enough to invite correction, and close enough to present experience to deserve serious attention. If intelligence is no longer fully visible inside the head alone, then neither is purpose. Both may need to be sought, at least in part, in the pathways that open between participants as dialogue unfolds. The task now is to understand which pathways lead somewhere worth going

© John Rust, April 2026. All rights reserved. Short excerpts may be quoted with attribution.

.