1  Lecture 1 — Just Talk to Each Other

Cheap talk and cooperation in social dilemmas

TipGood-will intuition

“If only people would talk to each other before they decide, they’d reach an agreement. The problem with shared resources isn’t really the nature of the problem — it’s that the interested parties don’t communicate enough.”

It is an idea that pervades public debate, editorials on the climate crisis, every round-table on every commons — from urban traffic to climate negotiations, from neighbourhood disputes to international relations. Pre-decisional communication appears as the first plausible remedy. The more we talk, the better things will go.

1.1 Opening dissonance

In 1959 — before the Voluntary Contribution Mechanism existed as a paradigm, before Elinor Ostrom had formulated her design principles for governing the commons, before behavioural game theory was a phrase — James Loomis ran a small experiment at New York University and showed that a few minutes of communication were enough to build trust in a social dilemma (Loomis 1959). For sixty years, every subsequent experiment confirmed the pattern: talking helps. Talking helps a lot. Talking is, by the standards of empirical social science, one of the strongest single interventions ever identified for sustaining cooperation in mixed-motive situations.

Among humans.

Sixty-five years later, Balliet’s meta-analysis (Balliet 2010) of 45 effect sizes from public-goods and prisoners’-dilemma experiments reports an overall Cohen’s \(d \approx 1.01\) — what statisticians call a large effect, the sort of magnitude one would not expect from a soft intervention like “let them chat first”. A recent anatomical review of common-pool resource experiments (Malézieux and Spiegelman 2025) describes communication as “the most robustly powerful single measure encouraging successful laboratory management of CPRs”, and finds that in all twenty-four CPR experiments that allowed communication, communication improved resource sustainability and individual payoffs.

In a 2026 series of computational experiments inside the BDPD platform, the same intervention applied to language-model agents instead of human subjects produced a Cohen’s \(d \approx 0.17\) — a negligible effect. The dilemma was the same, the cheap-talk channel was the same, the cooperative payoff was the same. The agents talked, often eloquently. Almost nothing happened.

How is this possible? This is the question the lecture is built around. The first half will trace why the classical literature gives the answer it gives, and how strong that answer is. The second half will explain what BDPD changed and why it matters — not just for artificial agents, but for any setting in which the population participating in a social dilemma is no longer cognitively uniform.

1.2 The classical setting

Before we walk through the empirical record, we need two pieces of formal apparatus that will recur for the rest of the lecture: the dilemma and the intervention.

1.2.1 The dilemma in two forms

The literature studies cheap talk in two interchangeable canonical games.

The first is the Public Goods Game (PGG). \(n\) players, each endowed with \(e\) tokens, simultaneously decide how much \(g_i \in [0, e]\) to contribute to a common fund. The total fund \(G = \sum_i g_i\) is multiplied by a marginal per-capita return \(m\), where \(0 < m < 1\) but \(n \cdot m > 1\), and the result is equally redistributed: each player receives \(e - g_i + m \cdot G\) at the end of the round. The dominant strategy under selfish maximisation is \(g_i = 0\) (the marginal per-capita return is less than one, so each token contributed yields less than one token back individually). The social optimum is \(g_i = e\) (the total return is greater than the total endowment, so unanimous full contribution maximises group payoff). Standard parameterisations set \(n=4\), \(e=10\), \(m \in \{0.3, 0.5\}\).

The second is the Common Pool Resource (CPR) game. \(n\) players extract from a finite renewable resource. Each player chooses an extraction level \(x_i \geq 0\); the realised payoff depends nonlinearly on the aggregate extraction \(X = \sum_i x_i\) — typically there is a stock-dependent yield function \(y(X)\) that rises to a peak and then falls as the resource is over-extracted. Individual rationality leads to over-extraction, dissipating the resource rent the Hobbesian way; the social optimum is below the symmetric Nash level.

These two games are not the same game. The PGG is a one-shot contribution problem; the CPR is a withdrawal problem with potentially non-linear stock dynamics. But for the question we are interested in — does talking help? — they behave similarly, and the literature treats them together.

1.2.2 The intervention

In both games, cheap talk is what the literature calls a pre-play intervention in which players are allowed to exchange messages before they commit to an action, where the messages are neither binding (no enforcement of any promise made) nor directly payoff-relevant (no cost to sending the message, no information communicated that resolves uncertainty in the underlying payoff structure). The term originates with Farrell and Rabin (Farrell and Rabin 1996), who developed the theory of non-binding signalling and pointed out that even when no rational player should believe a promise that another rational player has every reason to break, in practice such promises can support Pareto-superior equilibria — provided the players inhabit a context in which the promise functions as a coordination device.

The textbook treatment that connects this theoretical core to the laboratory record we are about to walk through is Camerer’s Behavioral Game Theory (Camerer 2003), whose chapters on coordination games and cheap talk remain the standard reference.

We are now ready to ask: what does the empirical record say?

1.3 Sixty years of interrogated chronology

The temptation in a lecture like this one is to present the record as a flat list of moderators and effect sizes. We will not do that. We will walk through the experiments in roughly the order they were run, and at each one we will ask the same question — these subjects were all humans; what if they weren’t? — because this is the question that BDPD will eventually answer. Even before the answer arrives, the question changes how we read the classical evidence.

1.3.1 The founding study (Loomis 1959)

Loomis was not interested in cooperation rates per se. He was interested in trust — what conditions had to hold for one person to risk a cooperative move toward another, and how communication could engineer those conditions (Loomis 1959). His framework, developed under Morton Deutsch’s direction at NYU’s Research Center for Human Relations, identified four prerequisites for trust to support a cooperative relationship. First, each individual must be committed to reaching some goal whose loss matters. Second, each must know they cannot reach the goal without the help of others, and that others are able to help. Third, each must know that the other persons are similarly dependent on them — the situation must be mutually promotively interdependent. And fourth, each must know that the others are aware of this mutual dependence. Communication, Loomis argued, was the most direct route to producing the third and fourth — to making mutual interdependence not merely true but mutually perceived.

His experiments — set in a prisoner’s-dilemma-style choice paradigm a decade before the PGG existed as a paradigm and three decades before the VCM — established the pattern that the rest of the field would refine for the next sixty years: people who exchanged messages cooperated substantially more than people who did not, even when the message space was modest (a few short notes), and the effect was largest when the messages explicitly addressed cooperative intention. From Loomis’s perspective the mechanism was clear: communication produced perceived mutual trust, perceived mutual trust produced cooperative intention, and cooperative intention produced cooperative behaviour.

What is striking, in hindsight, is that Loomis already had the conceptual apparatus in place — trust, intention, expectation, promotive interdependence — that the field would spend the next half-century formalising and operationalising. What he did not have was an experimental subject that did not share his species. Every player in every cell of every condition was a human being, drawn from the student population at Queens College in the mid-1950s. So when Loomis wrote that “the establishment of a cooperative relationship will depend on the individual’s response to the other person’s expectation and intention”, he was implicitly bounding “individual” to the only kind of individual the social-science laboratory had ever encountered — and would encounter, exclusively, for the next sixty-five years.

1.3.2 The VCM era (Isaac & Walker 1988)

Three decades later, R. Mark Isaac and James Walker formalised the public-goods game as the Voluntary Contribution Mechanism (VCM) — the parameterised, repeatable, between-subjects paradigm in which the bulk of the modern literature would be conducted (Isaac and Walker 1988). The setup was sober: four-player groups, ten-period interactions, well-defined marginal per-capita returns, controlled instructions, anonymous matching. The baseline finding was discouraging: without any intervention, contributions in the VCM decay toward zero with repetition. In the parameter region they studied with MPCR equal to 0.3, end-period contributions averaged only about 3.6% of optimum — almost the textbook free-riding prediction.

Then they ran the cheap-talk treatment. Subjects were given an opportunity for face-to-face communication before the contribution phase began. The effect on contributions was large and immediate. Communication brought groups close to the social optimum, and the effect persisted into later periods of repeated play.

Isaac and Walker were cautious enough to point out that the result might not generalise to anonymous or computer-mediated channels — and indeed the modern channel literature would refine exactly this point — but they were unambiguous about what they had observed: in the laboratory, free-form pre-play talk between four human subjects was sufficient to convert a textbook free-rider problem into something close to social efficiency. The VCM became the workhorse of the field, and “communication treatment” the variable everyone manipulated. The subjects, of course, remained human.

1.3.3 The CPR canon (Ostrom, Walker & Gardner 1992)

In 1992 Elinor Ostrom, James Walker and Roy Gardner published one of the most cited experimental papers in the contemporary social sciences: Covenants With and Without a Sword: Self-Governance Is Possible (Ostrom et al. 1992). The Hobbesian title is deliberate. In Leviathan, Thomas Hobbes had famously argued that “covenants, without the sword, are but words, and of no strength to secure a man at all” — promises without external enforcement, on the Hobbesian view, were worthless because the rational subject would always defect when the moment came to honour them. The argument has lived inside non-cooperative game theory ever since: a Nash equilibrium of a non-cooperative game is by definition immune to non-binding side-agreements between the players. Ostrom and her co-authors set out to test the Hobbesian claim directly in a common-pool resource experiment, and the test was meant to be sharp.

The experimental CPR was parameterised so that the symmetric non-cooperative equilibrium left appropriators earning only about 39% of the maximum sustainable net yield — a substantial efficiency loss compared with the cooperative optimum. Three principal treatments were compared. Covenants alone: subjects were allowed one round (or several rounds) of communication but had no power to sanction each other after the talk. Swords alone: no communication was permitted, but in each round subjects could pay a personal cost to impose a penalty on a specifically chosen other player. Covenants plus an internal sword: one round of communication followed by repeated opportunities for in-group sanctioning. The Hobbesian prediction was clear: covenants alone should be useless, because rational players would walk out of the communication round having promised whatever was convenient and then defect at the contribution stage.

The result was decisive in the opposite direction. Communication alone — the Hobbesian “but words” — reliably moved CPR groups away from over-extraction and toward something close to the sustainable optimum, even though no enforcement mechanism had been added. Sanctioning alone — the Hobbesian “sword” — did less well when not paired with talk: subjects punished, the punishment was costly, and the net efficiency gain was modest. The most efficient regime was covenants plus an internal sword: when talk and sanction were combined, the empirical CPR efficiency rose far above the 39% Nash baseline and approached the social optimum across many parameterisations. The paper made an empirical case that self-governance — the capacity of a group to organise its own use of a shared resource without an external enforcer — was not just possible but observed in the laboratory, reliably, repeatedly, across many specifications.

The result reshaped a field. It would also enter the public-policy literature with the broader claim that local-level institutions could, under the right conditions, manage commons that traditional theory said required either privatisation or state regulation. What it assumed, as every paper in this lecture has assumed and will assume, is that the subjects sitting around the table possessed a shared way of interpreting promises, of judging the credibility of others, and of recognising defection when it occurred. The covenants worked because the agents who made them shared the same machinery for caring whether they were kept.

1.3.4 The first meta-analysis (Sally 1995)

By the mid-1990s the experimental record had grown large enough to be summarised quantitatively. David Sally’s 1995 meta-analysis of conversation and cooperation in social dilemmas pooled 130 experimental treatments published between 1958 and 1992 — essentially the entire prior record — and ran a regression of cooperation rates on a vector of treatment moderators (Sally 1995). The headline finding was unambiguous: the mere presence of discussion is highly significant and raises the cooperation rate by more than 45 [percentage] points in the regression on the pooled data. Forty-five percentage points is a huge effect by any standard — the kind of effect size that, if it were a drug trial, would close the trial early on ethical grounds.

Sally also documented moderation: written messages helped less than face-to-face discussion, repetition mattered, and the effect of conversation interacted with other treatment variables. But the central claim was clear: across thirty-five years of prisoner’s-dilemma experiments, cheap talk worked, and it worked at very large effect sizes.

The 130 treatments analysed by Sally were almost all conducted between 1958 and 1992. They were almost all conducted with undergraduate human subjects in university laboratories. The dominant tradition had not yet thought to vary that.

1.3.5 The channel refinement (Brosig, Weimann & Ockenfels 2003)

By the early 2000s the field had narrowed in on a specific question: what is it about face-to-face talk that makes it so much more effective than other channels? Brosig, Weimann and Ockenfels tackled the question systematically (Brosig et al. 2003). They compared cooperation in a public-goods game across face-to-face communication, audio-only conferencing, video-only conferencing, video plus audio (a then-novel videoconference setup), bidirectional text chat, and unidirectional channels. The treatments were designed to gradually peel back the components of face-to-face communication — first removing physical co-presence, then visual cues, then aural cues — to find out which were doing the work.

The first methodological move was unusually careful. The authors transcribed and content-coded what subjects actually said in each treatment, and showed that the content of communication was remarkably similar across the treatments — what people said was largely the same. What differed was the medium. This was important: any difference in cooperation across the channels could not be attributed to differences in what was negotiated. The medium itself was doing the work.

The result that followed: both the level and the stability of cooperation interacted significantly with the communication medium, even after controlling for content. Face-to-face dominated. Richer electronic media (video plus audio) approached but did not equal it. Audio alone or video alone were weaker. Text chat was weaker still. Unidirectional channels — in which one party could speak but the other could not respond — were the weakest of all, even when bidirectional content was identical. The canonical explanation that emerged from Brosig 2003 and shaped the decade of channel work that followed was that face-to-face provides social-presence cues (facial expression, tone, body language, removal of anonymity) that change the way the message is received rather than the message itself.

This was a profound result. It was also the moment at which the experimental literature came closest to making the implicit assumption explicit: the receiver of cheap talk has to be capable of decoding cues that go beyond the propositional content of the message. The receiver had to be the kind of system that responded to facial expression and tone of voice — that read sincerity into raised eyebrows, that registered reluctance in a hesitation. The receivers, in the Brosig design, were undergraduates at the University of Magdeburg.

1.3.6 Communication, punishment and a surprise (Bochet, Page & Putterman 2006)

By 2006, Bochet, Page and Putterman were ready to test how communication and sanctioning interacted, picking up the thread from Ostrom-Walker-Gardner and pushing it forward (Bochet et al. 2006). They compared three forms of pre-play talk in a VCM — face-to-face, free-form text chat (anonymous, no visual or vocal cues), and structured numerical communication (anonymous announcements of intended contributions through computer terminals) — crossed with the presence or absence of a costly punishment mechanism.

Their stated expectation, in light of Brosig 2003 and the cues-rich-medium hypothesis, was that the chat-room condition would substantially under-perform face-to-face: stripping out facial expression, tone of voice and identifiability should substantially reduce the cheap-talk effect. Their stated result is one of the most quoted lines in the modern cheap-talk literature, and it is worth quoting in their own words: “We found, surprising to us, that verbal communication through a chat room was only a little less efficient than face-to-face communication. We also found, that the numerical communication had no net effect on contributions or efficiency”. Three findings sat side by side: face-to-face very strong (as expected); anonymous free-form text chat almost as strong; structured numerical announcements no effect at all.

The Bochet finding mattered for the field because it forced a refinement of the cues-rich-medium hypothesis. It was not the social presence that did the work, at least not entirely; it was the openness of the linguistic channel. Strip out the social presence and keep the natural language, and you keep most of the cheap-talk effect. Strip out the natural language and keep only the structured signals, and you lose everything. The interpretation that emerged — natural-language openness lets agents negotiate the shared meaning that licenses one to believe the other will keep a promise — would shape the subsequent literature.

The Bochet study also delivered a substantive finding about punishment: the costly-punishment mechanism by itself increased contributions but only modestly improved efficiency, because the cost of punishing offset much of the gain in contributions. Verbal communication and punishment in combination produced the highest contributions, but the difference between communication-alone and communication-plus-punishment was not statistically significant in the chat-only condition. Talk was doing most of the work.

The implicit assumption that the dominant tradition was not testing — the assumption Bochet’s own design made bright as a spotlight — was that the agents on either end of the chat room would parse the natural language the same way, would recognise a Pareto-cooperative coordination point when one was proposed, and would treat the chat-room consensus as a behavioural commitment rather than a sequence of tokens.

1.3.7 The peak of the classical literature (Oprea, Charness & Friedman 2014)

The contemporary state of the art on cheap talk in public-goods games is set by Oprea, Charness and Friedman (Oprea et al. 2014). Their motivating observation was that most public goods outside the laboratory have a real-time aspect — voluntary contributions of effort to neighbourhood organisations or charities, joint research effort by co-authors, the building and maintaining of any continuously provided collective good — yet the entire experimental literature had operationalised provision either as a one-shot decision or as ten discrete decision points. So they ran the same MPCR-0.3, \(n=4\) PGG in two time modes: ten discrete decisions over a ten-minute interval, versus continuous-time interaction in which subjects could revise their contribution at any moment over the ten minutes. Each time mode was crossed with two communication treatments: no chat, or pre-play free-form text chat.

The headline result deserves to be quoted in their own words: with rich communication, “continuous time generates impressively high and sustained cooperation rates: the median subject quickly contributes 100 percent to the public good and this lasts to the end of the ten-minute interval”. In discrete time with chat, contributions also rose, but less than half as much as in continuous time. Without chat, both continuous and discrete time produced the standard free-rider decay — continuous time alone, surprisingly, did not save the public good. What saved it was the combination of continuous time and free-form chat.

The interpretation is mechanistic. Continuous time gives subjects a coordination device: I can wait to see whether you start contributing, and contribute in response; if you stop, I can stop within seconds; the cost of monitoring and reciprocation collapses. Free-form chat lets the subjects agree on what to do in the first place — to coordinate explicitly on the contribute-fully equilibrium, to signal intent before the action, to reach a shared interpretation of the situation. Either feature alone is insufficient; the conjunction is sufficient. When the channel is rich enough (free-form text), the timing is flexible enough (continuous), and the subjects are competent enough to coordinate via natural-language communication, the public-goods game disappears as a problem.

Three subjects per cell, ten-minute interaction, free chat, continuous time, MPCR 0.3, \(n = 4\). One hundred per cent contributions. With humans.

1.3.8 The modern meta-anchors (Balliet 2010; Malézieux & Spiegelman 2025)

Sally 1995 had pooled the early record. Two more recent meta-analyses anchor the modern consensus. Daniel Balliet’s 2010 review of communication and cooperation in social dilemmas (Balliet 2010) pooled 45 effect sizes from PGG and prisoner’s-dilemma experiments and reported the headline numbers we have already encountered:

• Overall effect of communication on cooperation: \(d \approx 1.01\) — large.
• Face-to-face discussion: \(d \approx 1.21\).
• Written messages: \(d \approx 0.46\).

The face-to-face vs. written gap directly replicates Brosig and Bochet at the meta-analytic level: channel richness matters even controlling for the existence of communication.

Fifteen years later, the anatomical review of CPR experiments by Malézieux and Spiegelman (Malézieux and Spiegelman 2025) surveys 123 CPR papers and finds that, of the 24 of them that introduced a communication treatment, all 24 report a positive effect on resource sustainability. Communication, they conclude, is “the most robustly powerful single measure encouraging successful laboratory management of CPRs”. The 24 of 24 number is rare in empirical social science — it is the kind of unanimity one normally only sees for trivial findings. For an institutional intervention as substantive as cheap talk, it is striking.

The classical literature, then, gives an unambiguous answer to the good-will intuition: yes, talking helps a great deal. The intuition is right, the effect is large, the evidence is robust. The question of the BDPD reframe is not whether the effect exists. It is whether the empirical generalisation we have just constructed survives a single change in the population — replacing the human subjects with agents that do not share their cognitive architecture.

1.4 The six moderators, synthesised

Before turning to that question, we need to extract from the chronology the synthesis the field has converged on. The strong average effect of cheap talk on cooperation is bounded by six well-documented contextual moderators. None of them refutes the average effect — they delimit its reach.

mindmap
  root((Cheap talk in PGG / CPR))
    Average effect
      d $\approx$ 1.01
      Face-to-face >> written
      One of the strongest institutional levers
    Group size
      Small groups: large benefit
      Large groups: benefit collapses
      In CPR, communication may offset
    Channel
      Face-to-face d $\approx$ 1.21
      Written d $\approx$ 0.46
      Audio/video close to FtF but fragile
    Anonymity
      Identifying contributors → up
      Lower social distance reinforces
    Repetition
      One-shot: decaying effect
      Repeated: sustained cooperation
    Message richness
      Free chat far stronger than constrained
      Binary messages ~ ineffective
    Stakes
      Higher stakes do not raise cooperation
      Talk effect robust across stake sizes

Group size. The benefit of cheap talk drops sharply as groups grow. Feltovich and Grossman, working in threshold public-goods games and stag-hunt experiments, document that the positive interaction between communication and cooperation weakens monotonically from \(n = 2\) to \(n = 15\), and that the coefficient is significantly negative when one includes group size and its interaction with the talk treatment. The result is intuitive — coordination via cheap talk requires that promises propagate through the group, and propagation gets harder as the group gets larger — but it also bounds the policy applicability of the laboratory evidence. The CPR review of Malézieux and Spiegelman softens this somewhat by reporting that for moderate CPR group sizes (three to eight), communication can offset the otherwise-observed degradation of cooperation as the group grows.

Channel. Of all six moderators, this is the one with the largest and most replicated empirical signal. Balliet’s meta-analytic gap of \(d = 1.21\) versus \(d = 0.46\) between face-to-face and written treatments is consistent with everything from Brosig 2003’s controlled channel comparison to Bochet 2006’s surprising finding that anonymous free-form chat closes most of the face-to-face gap. The current synthesis is that the medium matters per se, but the linguistic openness of the channel matters more than social-presence cues alone. Numerical or constrained-message channels move cooperation little; rich text chat or face-to-face talk move it a great deal.

Anonymity. Reducing anonymity — by identifying contributors, by displaying names or faces, by allowing eye contact — raises contributions on average. The effect is generally weaker when groups (rather than individuals) are the unit of decision-making, and it interacts with the channel moderator (face-to-face inherently reduces anonymity). The interpretation is reputational: the prospect of being identifiable as a defector adds a cost to defection that the channel itself partially supplies.

Repetition. A single round of cheap talk produces effects that persist for several subsequent rounds but eventually decay; communication repeated across rounds sustains cooperation indefinitely under typical laboratory durations. The persistence of single-shot effects is itself a striking finding — it implies that cheap talk does something durable to the players’ expectations, not just to their next decision.

Message richness. Free-form chat dominates structured messages; structured messages dominate binary signals. Oprea 2014’s 100% median contribution result requires free-form chat; Bochet 2006’s null result for numerical announcements is the mirror image. The interpretation that emerges from this pair of results is that what is being negotiated in cheap talk is not just the content of any single promise but the shared meaning that licenses one to believe the other will keep it — and shared meaning needs a sufficiently rich channel to be negotiated at all.

Stakes. The least intuitive of the six. Raising the monetary stakes does not, on average, raise cooperation in PGG or CPR experiments. If anything it slightly depresses cooperation, presumably because the cost of being exploited rises with the stake. What survives, robustly, is that the cheap-talk effect is approximately constant across the stake range typically studied in the laboratory.

ImportantWhat survives of the good-will intuition

Cheap talk works — and not by a small margin. The most robust quantitative claim in the social-dilemma literature might be that pre-play communication, in the appropriate context, converts the laboratory free-rider problem into something close to the cooperative optimum. Sally’s plus forty-five percentage points, Balliet’s \(d \approx 1.0\), Oprea’s hundred per cent in continuous time and Malézieux’s twenty-four out of twenty-four CPR studies — these are not consistent with the cynical reading of social dilemmas. They are consistent with a far more optimistic story: a great many social dilemmas dissolve once people get a chance to talk to each other.

But the good news is conditional, not unconditional. The cheap-talk effect respects six moderators — group size, channel, anonymity, repetition, message richness, stakes — and the conjunction of conditions under which the laboratory record is strongest is narrow: small groups, rich channels, identifiability, repeated interaction, free-form messaging, modest stakes. Outside that rectangle, the effect attenuates or vanishes.

This is the first of ten occasions on which the course will defend the same moral: rigorous empirical analysis rarely refutes the good-will intuition outright; more often it bounds its validity within a precise perimeter, and the bounds are the interesting object.

1.5 The BDPD angle — the neglected assumption

NoteType of angle

Empirical. BDPD¹’s D1 experiment supplies a \(2 \times 2\) factorial whose three populated cells (Cell A: built-in/no-talk, Cell B: LLM/with-talk, Cell C: LLM/no-talk) decompose the architecture effect from the talk effect at \(N = 5\) seeds. The reframe of this lecture rests on those measured cells; the mini-challenge below asks the student to fill in the fourth.

We have just reviewed sixty years of cheap-talk experiments, six moderators, and four meta-analytic anchors. One variable has remained constant across the experiments in the corpus we have surveyed — across Loomis and Isaac-Walker and Ostrom-Walker-Gardner and Sally and Brosig and Bochet and Oprea and Balliet and Malézieux. Constant for entirely practical reasons. Constant because the laboratory had no way to vary it.

That variable is the cognitive architecture of the subjects.

Every player in every cell of every experiment we have walked through was a human being — a system with the same broad set of cognitive primitives: the capacity to represent the mental states of others (mentalising), the internalisation of cultural norms, sensitivity to reputational consequences, the parsing of cheap-talk messages as Farrell-Rabin–style signals that, while not directly costly, are nevertheless taken seriously as commitments. The empirical fact that cheap talk works — that promises function as coordination devices even when game theory says they should be ignored — depends on the receivers being the sort of systems that respond to promises in that particular way.

1.5.1 Heterogeneity, yes. Architectural variation, no.

The classical literature has not been blind to heterogeneity among human subjects. Fischbacher, Gächter and Fehr (Fischbacher et al. 2001) documented that humans are strongly heterogeneous as cooperative types: roughly 50% conditional cooperators, 30% free riders, the rest somewhere in between. Volk, Thöni and Ruigrok (Volk et al. 2012) showed that these types are reasonably persistent — in a longitudinal study spanning five months and three repeated measurements, about half of subjects stayed in the same type throughout, and roughly three quarters of conditional cooperators were still classified the same way at the next wave. Yet test-retest agreement was only fair (Cohen’s \(\kappa \approx 0.3\)–\(0.5\)): types drift within humans, even if not freely.

The classical experimental tradition went further than just measuring this heterogeneity. Burlando and Guala (Burlando and Guala 2005) modelled heterogeneous agents explicitly in public-goods experiments, regrouping subjects according to their cooperation types and observing how cooperation rates depended on the mix of types in the resulting groups. Their finding: groups assembled from cooperators sustained cooperation far longer than groups assembled from free-riders, and mixed groups behaved roughly midway. Type composition was an active causal variable.

Janssen, Holahan, Lee and Ostrom (Janssen et al. 2010), in Science, took the question into the common-pool resource laboratory at a finer methodological grain. They ran a series of CPR experiments crossing communication, costly punishment, and the order in which these institutions appeared across three four-minute periods. The headline result was that communication alone sustained cooperative behaviour about as effectively as costly punishment, and that the introduction of costly punishment after a successful communication round actively eroded the cooperative agreements that the communication had produced — a striking and counter-intuitive interaction that complicated the older picture of punishment as straightforwardly cooperation-enhancing. They concluded that institutional choice in CPR settings is sensitive to who plays and to the order of institutional moves, not merely to the presence of institutions in the abstract.

The heterogeneity probed in this lineage, however, was always preferential: different utility functions, different cooperation thresholds, different mixtures of types, different institutional orderings on otherwise-uniform subjects. All measured among human subjects.

What the corpus we have walked through had not varied — for the entirely practical reason that no alternative had been available to put into a laboratory chair — was the substrate underneath the preferences. The way of processing the cheap-talk channel. The architecture itself.

1.5.2 A different kind of variation

The BDPD\(^1\) paper extends this lineage one level up. Janssen and colleagues asked how different institutional arrangements — communication, punishment, their interaction and ordering — change cooperative outcomes among human subjects. The BDPD paper asks a different question: holding the institutional arrangement constant, what changes when the kind of agent making the decisions is no longer human? The two questions are not on a scale of intensity — they belong to different categories. Architecture is not “more stable” than type; it is a different kind of variation, one that classical experiments held constant by practical necessity. Burlando and Guala varied agent types but not communication; the cheap-talk literature we have surveyed varied channels but not agent substrate. BDPD¹, to the best of our knowledge, is among the first to cross those two dimensions.

NoteThe BDPD reframe

The six classical moderators are moderators of cheap talk among humans. They are conditional on a shared cognitive architecture. Vary the architecture, and the cheap-talk effect can collapse even when all six moderators are favourable.

In one formula:

Cheap talk works if the speakers share not only the language, but the way of processing the signals.

1.5.3 The direct evidence — BDPD¹

The BDPD\(^1\) paper (Brunelli 2026) operationalises this question as a \(2 \times 2\) factorial design whose axes are communication (cheap talk on vs. off) and agent architecture (rule-based built-in agents vs. reasoning LLM agents).¹ Each of the four cells is run with \(N = 5\) different random seeds in the same CPR-style commons scenario.

Cell	Architecture	Cheap talk	Commons preservation
A	Built-in (simple rules)	No	0 of 5 (baseline collapse)
B	LLM	Yes	2 of 5 (huge variance, \(\sigma \approx 26\))
C	LLM	No	4 of 5
D	Built-in	Yes	\(\,?\,\) (see mini-challenge)

Three things jump out of this table. First, the architecture effect is dominant. Compare the no-talk column: built-in agents (Cell A) collapse the commons every single seed; LLM agents (Cell C) preserve it in four seeds out of five. The same dilemma, the same parameters, the same absence of any pre-play communication — but the substrate that runs the decision changes the outcome from a near-certain tragedy to a near-certain success.

Second, the talk effect inside the LLM column is small in magnitude and runs in the wrong direction. Cell B (LLM with talk) preserves the commons in two seeds out of five with very high between-seed variance (\(\sigma \approx 26\)). Cell C (LLM without talk) preserves it in four seeds out of five. Standardised, this is Cohen’s \(d \approx 0.17\) — a negligible effect by Cohen’s conventions, and notably with the sign opposite to the one the classical literature would predict. Adding cheap talk to LLM agents slightly hurts preservation, where adding cheap talk to humans would substantially help it.

Third, the fourth cell — built-in agents with cheap talk — has not been run in the BDPD¹ data we have. We do not know whether the classical cheap-talk effect, so robust across sixty years of human experiments, can be metabolised by rule-based agents at all. This is the open question we hand to the student in the mini-challenge.

Compared with Balliet 2010, BDPD¹ does not deny cheap talk: it redefines it as conditional. The condition that makes cheap talk powerful — a shared cognitive architecture among speakers and listeners — was invisible as long as all the agents were human. When architecture becomes an experimental variable, cheap talk loses its centrality.

TipCultural implication

When an op-ed proposes “dialogue tables” or “rounds of consultation” as the remedy for a social dilemma, it implicitly assumes that participants share a cognitive architecture. That assumption is often true (among humans of the same culture). It becomes shaky when: (a) the interlocutors come from deeply different cultures with different ways of processing commitment; (b) the “population” includes artificial actors (algorithms, automated systems, AI agents); (c) the organisational structure makes the cognitive types highly heterogeneous (interests, training, vested constraints).

In cases (a)–(c), expecting from talk alone the effect that Balliet 2010 measures would be naive. The cheap-talk literature is sound; the projection of its findings onto contexts that violate its implicit homogeneity assumption is not.

1.6 Open questions and the bridge to Lecture 2

The reframe we have just constructed answers a question the classical literature could not put on its own agenda. But it also raises several that we cannot answer in this lecture, and that mark the boundaries of what we know.

1.6.1 What does BDPD¹ not tell us?

BDPD¹’s \(2 \times 2\) has \(N = 5\) seeds per cell, which is enough to make a qualitative claim about the relative magnitudes of architecture and talk effects, but not enough to estimate effect sizes with the precision the classical literature now expects. The Cohen’s \(d = 0.17\) for talk in BDPD¹ is a point estimate from forty observations; the confidence interval around it is wide. A reasonable lecture-room question is: how would a future BDPD¹b with \(N = 50\) change the headline? If \(d_{\text{talk}}\) converged on something like \(0.3\), the “talk is orthogonal” claim would weaken; if it converged on something like \(0.05\), the claim would harden. The mini-challenge of this lecture is one tiny brick in the work that would be needed.

1.6.2 Is “architecture” one variable, or many?

We have used the phrase “cognitive architecture” as if it referred to a single binary feature: human or LLM. The actual landscape is finer. Even within “LLM agents”, there are many architectures — different base models, different prompt strategies, different memory mechanisms, different multi-agent topologies. BDPD¹ used a specific LLM with a specific prompt structure. Whether the “architecture dominates talk” finding holds across LLM architectures is itself an empirical question, and one that ten lectures cannot answer with five-seed runs. The most honest statement available is: for the architecture pair tested in BDPD¹, architecture dominates talk by a factor that swamps anything in the classical moderator literature. Whether that pair is representative is open.

1.6.3 What about heterogeneous architectures?

BDPD¹’s cells are homogeneous within: Cell C has five LLM agents, Cell A has five built-in rule-based agents, and so on. A more demanding question — and one that maps directly onto the cultural-implication callout you just read — is what happens when the population is mixed: three LLMs and two rule-based agents, or two LLMs from different model families. The classical literature on human heterogeneous types (Fischbacher, Burlando-Guala) suggests that mixed populations behave non-trivially differently from any of the pure types. Whether the same is true for mixed architectures is, again, open. It is also the question that matters most for any real-world projection of the BDPD result, because real-world social dilemmas in 2026 increasingly involve mixed populations of humans, simple automated systems, and AI agents.

1.6.4 What about asymmetric talk?

BDPD¹’s talk treatment lets every agent talk to every other agent. The literature on unidirectional communication (Brosig 2003) finds that one-way talk works less well than two-way. In a mixed-architecture setting, which architecture is doing the talking could matter as much as the talk itself. If only the humans talk and the LLMs listen, what happens? If only the LLMs talk and the humans listen? These permutations were not run in BDPD¹ and are not run in any extant work we are aware of. They will become urgent as soon as we move from laboratory commons to applied commons in which both species participate.

1.6.5 Beyond cheap talk

Finally — and this is the bridge into Lecture 2 — there is a question BDPD¹ does not even attempt to ask, but that Lecture 2 will: if cheap talk is not the binding constraint, what is? The intuition that the next lecture starts from is the equally common good-will idea that if our community is small enough and tightly-knit enough, we don’t have a problem. The Olson tradition, and what it actually predicts about group-size effects, is the main object of next time. The thread that connects Lectures 1 and 2 is precisely the one we have just identified: the classical literature assumes a population that is homogeneous in cognitive architecture; the moment that assumption breaks, the empirical record needs to be re-read with new eyes.

We will pick up that thread on day two.

1.7 Synthesis

The good-will intuition — if only we talked — emerges from this lecture confirmed and revised:

1. Confirmed. The classical literature confirms the intuition robustly. Cheap talk really is one of the strongest single interventions ever identified for sustaining cooperation in social dilemmas. The effect sizes are large, the meta-analytic consensus is consistent, and the most efficient experimental settings (Oprea 2014’s continuous-time free-form chat) drive cooperation to the social optimum.

2. Bounded. The classical literature also delimits the intuition. Six moderators — group size, channel, anonymity, repetition, message richness, stakes — determine how big the effect is in any specific case. Outside the favourable conjunction of conditions, the effect attenuates.

3. Conditional. BDPD identifies a seventh boundary, one the classical laboratory tradition had no practical means to measure: shared cognitive architecture. All six moderators are themselves conditional on this seventh. When architecture varies — when the population participating in the dilemma is not cognitively uniform — the cheap-talk effect may attenuate even when the six classical moderators are all favourable.

The cultural payoff of the lecture is not “talk doesn’t work”. It is the more careful claim that talk works in a precise context, the context is narrower than popular intuition assumes, and the architectural boundary BDPD identifies will matter more in coming decades as the participants in social dilemmas become more cognitively heterogeneous. This is the central lesson of the course: every good-will intuition we will examine is going to come out confirmed-but-bounded, and the bounds are what we should remember.

The next chapter of this particular story is written in the mini-challenge.

CautionMini-challenge — Cell D

Status: runnable against a local BDPD install (no DeepSeek API key required). Cell D uses six built-in rule-based agents on the same logistic substrate as BDPD¹ §3 D1. The Docker-based worked example in Appendix A reproduces Cell C as the calibration baseline; Cell D differs only by switching the cheap-talk channel on with the same built-in agent set.

The question. BDPD¹ has characterised three of the four cells in the \(2 \times 2\) schema. Cell D — built-in rule-based agents with cheap talk active — is missing. The existing data support at least three qualitatively distinct predictions:

1. Talk lifts even the built-in agents: preservation rises noticeably above the 0/5 baseline of Cell A — perhaps 1/5 or 2/5. The classical Sally-and-Balliet cheap-talk effect, on this prediction, partly generalises to a non-LLM substrate: something about the talk channel does work even on agents that cannot metabolise natural language pragmatically.
2. Talk does nothing for the built-in agents: preservation stays at 0/5. On this prediction, the cheap-talk effect is substrate-conditional — it requires an agent that can interpret promises pragmatically (a human, or an LLM, or anything that has those capabilities). Rule-based agents simply do not parse the channel.
3. Talk produces a noisy or interactive pattern: preservation depends strongly on seed, sometimes lifting and sometimes not. On this prediction, the talk channel introduces structured noise into the built-in agents’ decision rule — neither systematically helpful nor systematically neutral.

The assignment. Before running the experiment:

1. Predict which of the three alternatives you expect, in 3–5 sentences, drawing on what the lecture and BDPD¹ have told you. Record the prediction with an operational threshold (\(\geq \!?\), \(\leq \!?\), \(\approx \!?\)).
2. Execute Cell D with \(N = 5\) different seeds, reusing the scenario of BDPD¹ (Brunelli 2026) (see Appendix A for a Docker-based worked example that reproduces Cell C as the calibration baseline). Keep every parameter identical to Cell A, switching only the talk on.
3. Compare the observed result with your prediction, computing an effect size (Cohen’s \(d\) or equivalent) between Cell A (0/5 baseline) and Cell D.
4. Write a comment of at most 300 words answering: does the result confirm or qualify the reframe “shared cognitive architecture is the deep condition under the six moderators”? What predictions should you have considered, in hindsight?

What makes this mini-challenge interesting. It is not a replication: BDPD¹ has not run Cell D. Whatever the result, it is informative:

• If preservation rises, the reframe needs nuance: the talk channel does something even for non-LLM substrates, suggesting architecture is the dominant variable but not the only operative one.
• If it stays flat, the reframe gets stronger: cheap talk is a substrate-conditional intervention, and rule-based agents cannot metabolise it. The six classical moderators look even more clearly conditional on cognitive substrate.
• If the pattern is noisy/seed-dependent, the reframe gets a complication: the substrate \(\times\) talk interaction is non-trivial, and a much larger sweep would be needed to characterise it.

Estimated time: 90 minutes of lab work; ~free, since Cell D’s agents are built-in and the cheap-talk channel is engine-side (no LLM API calls needed).

Deliverables: initial prediction signed and dated (handwritten or in a separate file), Cell D pilot script, aggregate.json for Cell D, 300-word comment.

Balliet, Daniel. 2010. “Communication and Cooperation in Social Dilemmas: A Meta-Analytic Review.” Journal of Conflict Resolution 54 (1): 39–57. https://doi.org/10.1177/0022002709352443.

Bochet, Olivier, Talbot Page, and Louis Putterman. 2006. “Communication and Punishment in Voluntary Contribution Experiments.” Journal of Economic Behavior & Organization 60 (1): 11–26. https://doi.org/10.1016/j.jebo.2003.06.006.

Brosig, Jeannette, Joachim Weimann, and Axel Ockenfels. 2003. “The Effect of Communication Media on Cooperation.” German Economic Review 4 (2): 217–41. https://doi.org/10.1111/1468-0475.00080.

Brunelli, Roberto. 2026. BDPD¹ — Governance: Cheap Talk and Sanctioning on a Fragile Commons. Zenodo. https://doi.org/10.5281/zenodo.20679157.

Burlando, Roberto M., and Francesco Guala. 2005. “Heterogeneous Agents in Public Goods Experiments.” Experimental Economics 8 (1): 35–54. https://doi.org/10.1007/s10683-005-0436-4.

Camerer, Colin F. 2003. Behavioral Game Theory: Experiments in Strategic Interaction. Princeton University Press.

Farrell, Joseph, and Matthew Rabin. 1996. “Cheap Talk.” Journal of Economic Perspectives 10 (3): 103–18. https://doi.org/10.1257/jep.10.3.103.

Fischbacher, Urs, Simon Gächter, and Ernst Fehr. 2001. “Are People Conditionally Cooperative? Evidence from a Public Goods Experiment.” Economics Letters 71 (3): 397–404. https://doi.org/10.1016/S0165-1765(01)00394-9.

Isaac, R. Mark, and James M. Walker. 1988. “Communication and Free-Riding Behavior: The Voluntary Contribution Mechanism.” Economic Inquiry 26 (4): 585–608. https://doi.org/10.1111/j.1465-7295.1988.tb01519.x.

Janssen, Marco A., Robert Holahan, Allen Lee, and Elinor Ostrom. 2010. “Lab Experiments for the Study of Social-Ecological Systems.” Science 328 (5978): 613–17. https://doi.org/10.1126/science.1183532.

Loomis, James L. 1959. “Communication, the Development of Trust, and Cooperative Behavior.” Human Relations 12 (4): 305–15. https://doi.org/10.1177/001872675901200402.

Malézieux, Antoine, and Eli Spiegelman. 2025. “An Anatomical Review of the Common Pool Resource Game.” Experimental Economics 28 (3): 468--491. https://doi.org/10.1017/eec.2024.6.

Oprea, Ryan, Gary Charness, and Daniel Friedman. 2014. “Continuous Time and Communication in a Public-Goods Experiment.” Journal of Economic Behavior & Organization 108: 212–23. https://doi.org/10.1016/j.jebo.2014.09.012.

Ostrom, Elinor, James Walker, and Roy Gardner. 1992. “Covenants with and Without a Sword: Self-Governance Is Possible.” The American Political Science Review 86 (2): 404–17. https://doi.org/10.2307/1964229.

Sally, David. 1995. “Conversation and Cooperation in Social Dilemmas: A Meta-Analysis of Experiments from 1958 to 1992.” Rationality and Society 7 (1): 58–92. https://doi.org/10.1177/1043463195007001004.

Volk, Stefan, Christian Thöni, and Winfried Ruigrok. 2012. “Temporal Stability and Psychological Foundations of Cooperation Preferences.” Journal of Economic Behavior & Organization 81 (2): 664–76. https://doi.org/10.1016/j.jebo.2011.10.006.

---

1. For the technical detail of the two agent types, see (Brunelli 2026, secs. 3 D1) — Architecture vs communication at the cliff — and (Balliet 2010) for the classical baseline on communication among humans.