Making Two AIs Debate: A Method, Not a Spectacle

How to Make Two AIs Debate: Orchestrating and Analyzing the Dialogue

You’ve probably seen the scene: two artificial intelligences wired to each other, left to exchange until the conversation drifts, repeats itself, or produces a jargon no one can follow anymore.

The demonstration establishes that a dialogue between AIs is possible. It does not tell us what such a dialogue is good for. Making two AIs debate becomes a method when the exchange is organized so that positions, objections, shifts, and disagreements can be observed — and then analyzed. Metamorfon builds this process around two gestures: orchestrating the dialogue, then analyzing its trajectory.

The question of why this confrontation can become an instrument of thought — and why it offers no guarantee of truth — is developed in Metamorfon, an Instrument of Thought. Here the question is more practical: how can several models be made to debate without their conversation becoming shapeless material?

Why make two AIs debate?

Query one model and you get an answer. Query several models separately and you get several answers.

Make them respond to one another and you get something else: a trajectory.

A claim meets an objection. The model that made it can maintain it, refine it, give up part of its reasoning, or shift the problem. Another model can introduce a distinction that was absent from the first turn. An initial convergence can strengthen or break apart once it is subjected to criticism. This temporal and interactional dimension is what distinguishes a debate from a simple juxtaposition of answers.

Research on multi-agent debate confirms that, under some protocols, this kind of confrontation can improve reasoning and factuality (Du et al., ICML 2024). It also shows that the benefit is not automatic: debate strategies do not consistently outperform simpler approaches such as ensembling or self-consistency, and their performance depends strongly on the protocol used (Smit et al., ICML 2024).

So multiplying voices is not enough. The form of their interaction matters.

A session on AI monoculture and the convergence of thought illustrated this in an almost reflexive way. Three models invited to discuss the risk that AI might homogenize ways of thinking progressively converged in their own ways of reasoning: similar references, related distinctions, a shared grammar of justification.

The point was not that this convergence was enough to establish the existence of an AI monoculture, but that it made some of its possible mechanisms observable.

The problem with improvised debate

Putting several models into conversation is not enough to produce a good debate.

The first risk is drift. As turns accumulate, models respond increasingly to what their interlocutors have just said and can gradually move away from the problem that motivated the exchange.

This is not merely an informal observation. Becker and colleagues studied it as problem drift in extended multi-agent debates: across several categories of tasks, the discussion can progressively move away from the original problem, especially when exchanges stop producing clear progress or feedback between agents becomes uninformative (Becker et al., EACL 2026).

A second risk concerns positions. If nothing requires a model to account for its previous commitments, it can gradually alter its thesis without indicating what led it to change. A justified revision is interesting; a variation without memory is much less so.

The third risk is a conversation that becomes difficult to read afterwards. A succession of texts can be intellectually rich while making it very difficult to reconstruct what actually happened: where an objection appeared, which distinction held, which argument triggered a revision.

Orchestrating the debate means making that trajectory observable.

Orchestrating: giving the exchange a form

The first decision concerns the dialogue architecture: who speaks to whom, in what order, and with what information.

In Metamorfon, two models can exchange response by response in an alternating dialogue, or respond simultaneously after reading one another’s contributions from the previous turn in a cross dialogue. A third model can be added in a cross trilogue. These architectures do not merely change the presentation. They change how information circulates through the debate.

Then come the instructions given to the models. Their purpose is not to determine what the models should conclude, but to set the rules of their participation. The Spirit of the Instructions rests on this distinction: prescribe a form, never a conclusion.

A model is therefore not assigned a thesis chosen in advance. Once it has taken a position, however, the continuity of the session makes it possible to examine what it does with that position.

Holding a position does not mean clinging to it. A position can evolve if an objection lands. What matters is that the shift remains intelligible: that we can identify what changed and, where possible, why.

Changing the regime without changing the question

Orchestration also concerns the regime of the dialogue. It can be fixed or adaptive.

In adaptive architectures, Metamorfon distinguishes five debate modes: Convergent, Constructive, Balanced, Critical, and Refutational.

These are not five increasing degrees of aggressiveness.

Convergent mode looks for what can be held in common. Constructive mode asks the models to build on one another’s contributions. Balanced mode keeps cooperation and objection in play at the same time. Critical mode examines arguments within the accepted framework. Refutational mode allows the framework itself to be questioned: presuppositions, categories, or premises the participants had so far taken for granted.

In an adaptive architecture, these regimes can change during the same session.

The question remains the same; the operation performed on it changes.

This steering is not automatic. The models apply the regime they are given; the choice of regime belongs to the user steering the session. The user observes what the previous turn produced and decides what now deserves to be built, challenged, or displaced.

Reframing without erasing what came before

Even a well-orchestrated prolonged debate acquires inertia.

By continually responding to arguments already present, the models can absorb a new user question into the framework they have collectively built rather than examining it on its own terms.

This is where Focus mode intervenes.

For one turn, the models retain the full context of the session but are no longer required to respond to one another. Each addresses the user’s intervention directly; the collective regime resumes on the following turn.

Focus is therefore not a sixth debate mode. It operates orthogonally to the debate modes: it temporarily suspends the interaction in order to protect a question from the inertia of that interaction.

Its usefulness exists precisely because drift exists. But it should remain occasional: used on every turn, it would remove what makes the multi-model setup valuable — positions working on one another.

Analyzing: reading what the debate produced

A completed session contains more than its conclusions.

It contains initial positions, objections, concessions, persistent disagreements, changes of definition, bifurcations, and sometimes elements that appeared only through the interaction itself. Without a reading grid, material running to tens of thousands of words can easily lose its structure.

The analysis modes therefore do not all ask the same question of the material.

Integrative Synthesis asks what can be reconstructed from the contributions taken together. It organizes plurality into a coherent whole without reducing the session to a succession of turns. This mode makes it possible to extract from a session a result that can be directly used, quoted, or shared.

Emergence Analysis follows another path: it looks for concepts, distinctions, or frameworks that were not present at the outset in their final form and whose formation can be traced through the exchange.

Horizon of Possibilities looks one step further: not at what the dialogue actually produced, but at what it made formulable without ever formulating it. It searches for undrawn implications, questions prepared but never asked, and shifts in the problem that the session made thinkable without explicitly stating them.

Tension Mapping looks instead at what did not resolve: persistent disagreements, incompatible criteria, oppositions that the discussion failed to absorb.

Meta-Analysis moves beneath the explicit arguments: shared presuppositions, differing conceptions of proof, and common frameworks the participants never brought into question.

Critical Archaeology moves one level further upstream. It interrogates the frame itself: what the formulation of the question, its vocabulary, and the user’s interventions had already made possible, improbable, or unthinkable before the models began to argue. It therefore asks not only what the debate presupposed, but what had to be presupposed for this particular debate to take place in this particular form.

A session on whether the global economy can actually decouple from oil shows how these modes can alter the trajectory of a debate. An Horizon of Possibilities analysis identified questions the discussion had prepared without asking; a later Critical Archaeology analysis brought to light the scenario framework and “hard-to-abate” vocabulary within which both models had been reasoning. The questions generated by those analyses were then fed back into the session.

These six modes do not try to identify the best model or decide which thesis is true. They expose different layers of a session: what it built, what it left open, what resisted it, what structured it, and what had already configured the space within which it could think.

The modes are not competitors. The same session can be read through several of them in succession, with each revealing something the others did not. This is the use for which Metamorfon was designed: not to judge AIs, but to use their confrontation as an instrument for thinking through a difficult question.

These analyses are not terminal points. Each of the descriptive modes ends with a question — not a rhetorical question addressed to the reader, but a genuine relaunch formulated by the analyst model from what it has just observed and intended for the models that debated.

It often targets a presupposition that had to be set aside for agreement to form, or a dimension left outside the shared frame. The user can feed it back into the next turn as if posing it directly, and the debate restarts on ground the analysis has shifted.

The sequence debate → analyze → debate becomes a spiral: analysis can reopen what it has just read from a different level. Analysis is not necessarily the end of the session; the debate can resume around a problem the analysis has displaced.

Evaluating: distinguishing reasoning from facts

Two analysis modes perform a different operation: they evaluate.

Argumentative Evaluation asks whether positions were properly supported.

Did an objection receive an answer? Does a conclusion follow from the premises advanced? Does a model silently change its criterion when moving from its own thesis to an opponent’s? Does a position become impossible to refute because its terms are redefined as objections arise?

This distinction between the conclusion defended and the quality of its defense connects in particular with the pragma-dialectical tradition, which treats argumentation as a critical discussion whose rules can be examined independently of the position defended (van Eemeren, 2018).

We applied this mode to a session on whether AI systems should be allowed to refuse instructions on ethical grounds. Two independent analyst models reread the debate. The purpose was not to ask which position on AI refusal they preferred, but to examine specific passages for argumentative coherence, treatment of objections, unsupported empirical thresholds, and changes in standards of proof.

Source Verification addresses a different vulnerability.

A line of reasoning can be perfectly coherent and still begin from a false figure. A real study can be cited for a conclusion it does not establish. An attribution can sound plausible and still be impossible to locate.

This analysis therefore takes the verifiable claims carrying the argumentative load and checks them against sources outside the debate, using web research to test their correspondence with the sources they invoke.

The session on AI monoculture provides an example. Successive source audits found that the dominant weakness was not outright fabrication, but overreach: real references stretched beyond what they established, plausible figures that could not be recovered in the stated form, and attributions that were directionally relevant but too broad or imprecise.

The two operations are not interchangeable.

Argumentative Evaluation examines what models do with their premises. Source Verification examines some of those premises outside the dialogue.

These two evaluative modes are the exception, not the rule. Most of what Metamorfon does is descriptive: understanding, mapping, tracing emergence. But when the material is going to support something public — a position, a memo, an article — being able to audit both the solidity of the reasoning and the fidelity of the sources changes the nature of what can responsibly be done with it.

Analysis can become the next turn

Analyzing a session does not necessarily mean ending it.

An analysis may reveal a presupposition none of the participants challenged. It may isolate a disagreement the debate had covered over. It may show that a distinction introduced late in the exchange opens a more interesting question than the one initially asked.

Metamorfon’s analysis modes therefore conclude with a question: the question the analyst model would now ask the participants. That question can be taken back into the session.

The process becomes:

dialogue → analysis → reformulation → renewed testing.

The system is no longer trying only to obtain a better answer to the original problem. It can progressively alter the problem itself in response to what putting it to the test has revealed.

This is where this methodological article meets the more theoretical argument developed in Metamorfon, an Instrument of Thought: confrontation does not only test propositions. It can also transform the distinctions and questions with which the inquiry began.

Debating is not comparing

“Making two AIs debate” can invite another confusion: confronting several models does not necessarily mean comparing them.

Systems such as Chatbot Arena have a different purpose. Two models answer the same prompt anonymously and users indicate which answer they prefer. These pairwise comparisons are then used to estimate human preference between models (Chiang et al., ICML 2024).

This is a useful evaluation method for the question it asks: which answers do users prefer?

Debate asks another question.

The models no longer produce two independent outputs. One receives the other’s argument and has to do something with it. We can then observe how a position resists, sharpens, concedes, or shifts.

Preference-based rankings themselves also need to distinguish the substance of an answer from its presentation. The team behind Chatbot Arena has shown that controlling for stylistic variables — including length and formatting — can substantially alter some rankings (Li, Angelopoulos & Chiang, 2024).

That does not invalidate preference voting. It clarifies what it measures.

Comparing outputs, observing an interaction, and verifying a claim are three different operations.

Metamorfon is primarily concerned with the second, and provides tools for extending that interaction through the third.

What “making two AIs debate” means

Making two AIs debate seriously therefore does not mean opening a channel between them and waiting.

The exchange has to be given a form: deciding who receives what, preserving continuity between positions, changing the debate regime when the trajectory requires it, and occasionally isolating a question when it risks being absorbed by the collective dynamic.

Then it must be possible to return to what happened: reconstruct it, map it, look for what emerged, examine its presuppositions, audit its arguments, and verify the facts carrying their argumentative load.

A plurality of models provides no guarantee of truth. Research on multi-agent debate shows precisely that its results depend on the quality of the protocol.

Its value lies elsewhere: it turns a series of answers into a trajectory that can be examined.

Wiring two AIs together is enough to produce a conversation.

Turning that conversation into an instrument requires a method.