Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner's Dilemma
summary
The gist
We study whether communication sustains cooperation when the partner is an artificial intelligence in repeated Prisoner’s Dilemma games and find that communication is the main determinant of…
In short
The study tested if communication sustains cooperation against an AI in repeated Prisoner's Dilemma games. Communication is crucial, with a single pre-game exchange boosting cooperation from 49% to 84%. Furthermore, participants adopt unconditional cooperation against the AI while still using Tit-for-Tat against humans.
Key concepts
- Communication Effects on Cooperation
- This examines how talking changes whether players cooperate. The research found that communication is the main driver of cooperation with an AI; even one pre-game chat significantly increases cooperation levels compared to no communication.
- Strategy Heterogeneity and Identity Effects
- Participants change their behavior based on who they are playing against. When playing against a human, people tend to use Tit-for-Tat. However, when playing against an AI, the modal strategy shifts towards unconditional cooperation.
- Content and Implementation Analysis
- This analyzes what people actually say during communication. Studies show that explicit rules are common in human-AI interactions, and promises to cooperate are honored at a very high rate when communication is present.
Terminology used across episodes
This episode discusses
- Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner's Dilemma · Paper Radio
- Communication in the Infinitely Repeated Prisoner's Dilemma: Theory and Experiments
- GPT in Game Theory Experiments
- When Trust Collides: Decoding Human-LLM Cooperation Dynamics through the Prisoner's Dilemma
- People Are Highly Cooperative with Large Language Models, Especially When Communication Is Possible or Following Human Interaction
- How Different AI Chatbots Behave? Benchmarking Large Language Models in Behavioral Economics Games
The paper
Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner's Dilemma · Read on arXiv
Chowdhury Mohammad Sakib Anwar, Konstantinos Georgalos
University of Winchester · Lancaster University
We study whether communication sustains cooperation when the partner is an artificial intelligence. In a laboratory experiment (n = 234), human participants play indefinitely repeated Prisoner's Dilemma games against an AI chatbot (GPT-5.2) under one of three conditions: no communication, free-form chat before the first round of each supergame, or free-form chat before every round. A fourth treatment replaces the AI with another participant under the every-round protocol. Communication is the main determinant of cooperation with the AI, and a single exchange is enough: pre-play chat raises cooperation from 49% to 84%, while chatting before every round adds less than one percentage point, in contrast to the timing effect found between humans. Once communication is available, cooperation with the AI is no lower than with a human partner, so we find no evidence of the 'machine penalty' documented in the psychology literature. The partner's identity nonetheless shifts strategies: under identical communication, participants tend to reciprocate against a human but to cooperate unconditionally against the AI. Participants also keep their promises to cooperate while carrying out their threats to punish only about half the time. The AI opens cooperatively and rarely defects first, so these differences originate on the human side.
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "Playing Against the Machine".
Warren: Here's the gist of the paper. We study whether communication sustains cooperation when the partner is an artificial intelligence in repeated Prisoner’s Dilemma games and find that communication is the main determinant of cooperation with the AI, with a single exchange being enough to raise cooperation from 49% to 84% <ref:2603.15852#pg2> The partner’s identity nonetheless shifts strategies, as participants tend to reciprocate against a human but to cooperate unconditionally against the AI under identical communication <ref:2603.15852#pg9> Experimental Design and Treatments The study investigates whether communication retains its cooperation-sustaining role when the opponent is an AI by varying communication across three levels—no communication, a single exchange before each supergame, or an exchange before every round—and comparing this against a human partner under the same protocol <ref:2603.15852#pg8> The design features four treatments: HAI-Rep (AI opponent with communication before every round), HAI-Pre (AI opponent with communication only before the first round of each supergame), HAI-NoChat (AI opponent without communication), and HH-Rep (human opponent with communication before every round) <ref:2603.15852#pg8> Participants play indefinitely repeated Prisoner’s Dilemma games against either an AI chatbot powered by GPT-5.2 or another human subject <ref:2603.15852#pg8> The experiment was implemented using oTree and conducted at the Lancaster Experimental Economics Laboratory (LExEL) and the Birmingham Experimental Economics Laboratory (BEEL) between March and June 2026 <ref:2603.15852#pg8> Communication Effects on Cooperation The results show that communication is the main determinant of cooperation against an AI partner, with a single exchange being enough, as cooperation is 49% without communication and rises to 84% with pre-play communication <ref:2603.15852#pg8> Allowing further exchanges before every round adds less than one percentage point (85%), contrasting with the human timing effect found between humans <ref:2603.15852#pg8> Once communication is available, cooperation against an AI is no lower than with a human partner, cooperating about 8 percentage points more with the AI <ref:2603.15852#pg8> This suggests that communication works mainly through one channel: coordination on the cooperative equilibrium <ref:2603.15852#pg10> Strategy Heterogeneity and Identity Effects Using structural estimation via the Strategy Frequency Estimation Method, the identity of the partner shifts human strategies even when the communication protocol is held fixed <ref:2603.15852#pg8> Under identical every-round communication, unconditional defection disappears, with a mass concentrating on Grim Trigger under pre-play and unconditional cooperation under repeated communication <ref:2603.15852#pg8> Specifically, the modal strategy against a human is Tit-for-Tat (40%), while the modal strategy against the AI is unconditional cooperation (40%) <ref:2603.15852#pg8> This finding shows that subjects reciprocate against another human player but cooperate unconditionally against an AI <ref:2603.15852#pg8> Content and Implementation Analysis Communication content analysis revealed that explicit repeated-game rules are common in the human-AI treatments, with 59.5% of pre-play and 41.8% of repeated statements containing an identifiable rule <ref:2603.15852#pg8> Furthermore, stated rules are implemented asymmetrically: promises to cooperate are honoured in 94–98% of cases, whereas threats to punish are carried out in only about half the time <ref:2603.15852#pg8> The analysis also shows that explicit repeated-game rules do not translate into cooperation gains; for example, the logit for a stated rule in HAI-Rep is +0.326, while the estimated strategy share in pre-play is around 90% <ref:2603.15852#pg8> The results suggest that communication with an AI works through a single channel, coordination on the cooperative equilibrium, because the AI follows the rule it states and rarely defects first <ref:2603.15852#pg10> The asymmetry between promises and threats is consistent with social psychology views that discussion raises cooperation because it elicits commitments that people feel bound to honour <ref:2603.15852#pg10> This outcome qualifies the “machine penalty” in cooperative games, as participants frequently defect on an AI that opens cooperatively, but this effect disappears with communication <ref:2603.15852#pg10> The implication for AI deployment is that cooperation required an opportunity to agree on a rule and an agent whose behavior matched its words rather than anthropomorphic cues <ref:2603.15852#pg10> Limitations and Future Work Several limitations should be noted, including the use of a single model (GPT-5.2) with a fixed system prompt, which means different models or prompts may elicit different behavior on both sides <ref:2603.15852#pg10> Cooperation is high in the communication treatments, which limits how sharply strategy estimates can separate unconditional from conditional cooperation <ref:2603.15852#pg10> The treatment differences originate on the human side, since the AI behaves in much the same way in every condition, and further work could vary the AI’s strategy or communication style <ref:2603.15852#pg10> The paper also notes that cooperation is high in communication treatments because they leave few off-path histories and limit how sharply strategy estimates can separate unconditional from conditional cooperation <ref:2603.15852#pg10> The content analysis captures only the rules that participants articulate, and most human statements contain no identifiable rule <ref:2603.
Title and authors: Ruth: So we're looking at this paper today titled "Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner’s Dilemma." It dives into whether talking to an AI actually helps us cooperate when we're playing those repeated games.
Warren: Yeah, it sets up a pretty clear experiment with three levels of communication against an AI: no chat at all, one exchange before each big game session, and one exchange before every single round.
Iris: It also compares this to playing against a real human partner under the same rules to see if the communication effect holds up when you switch partners.
Ruth: The core finding they highlight is that communication is the main thing that sustains cooperation with an AI, and even just one exchange before each supergame gets cooperation from forty-nine percent up to eighty-four percent.
Warren: That jump from forty-nine to eighty-four with just a single pre-play chat sounds pretty significant for how we think about interacting with these machines.
Iris: And what’s interesting is that they found that after you have communication available, cooperating against the AI is no worse than cooperating against a human partner at all.
Ruth: That means they didn't find any evidence of that "machine penalty" in the psychology literature when communication was present, which is a big point because people often react negatively to machines compared to people.
Warren: It suggests that communication works through one main channel for the AI, which is coordinating on the cooperative equilibrium, as the AI seems to follow whatever rule you state and rarely defects first.
Iris: And when we look at how strategies change, they found a shift in behavior: participants tend to cooperate unconditionally against an AI when there's identical communication every round compared to when they play against a human.
Ruth: So, the strategy against another human is Tit-for-Tat at forty percent, but the strategy against the AI under constant communication shifts to unconditional cooperation at forty percent as well.
Warren: That means people are reciprocating against another person but cooperating unconditionally with the AI when they have that consistent communication channel open.
Iris: They also looked at what people actually say, and they saw that explicit repeated-game rules are common in these human-AI treatments, with fifty-nine point five percent of pre-play statements and forty one point eight percent of repeated statements containing an identifiable rule.
Ruth: But the content analysis showed that just having those stated rules didn't actually translate into bigger cooperation gains; for example, the logit for a stated rule in the AI-Rep treatment was plus zero point three two six, while the strategy share in pre-play was around ninety percent.
Title and authors: Warren: That suggests that just stating the rule isn't enough to guarantee better behavior, which makes sense if it doesn't change what they actually do when things get tough.
Iris: They also found an asymmetry in how rules are implemented, noting that promises to cooperate were honored in ninety-four to ninety-eight percent of cases, but threats to punish were only carried out in about half the time.
Ruth: That contrast between keeping promises and actually threatening punishment is consistent with social psychology ideas about discussion raising cooperation because it makes people feel bound to honor commitments they made.
Warren: And this whole thing qualifies what some call the "machine penalty," because participants frequently defect on an AI that opens cooperatively, but that effect disappears once communication is in play.
Iris: The limitation they point out is that the experiment used a single model, GPT-five point two with a fixed system prompt, which means different models or prompts could elicit different behavior on both sides <ref:2603.15852#pg1>.
Ruth: They also noted that cooperation rates remained high in the communication treatments because they left few off-path histories and limited how sharply strategy estimates could separate unconditional from conditional cooperation.
Warren: And another limitation is that the treatment differences came from the human side, since the AI actually behaved much the same way across every condition, meaning further work could look at varying how different AI strategies or communication styles affect things.
Iris: So to wrap up, they suggest that communication with an AI works through a single channel: coordination on the cooperative equilibrium because the AI follows the rule you state and rarely defects first.
Ruth: That really frames it as needing an opportunity to agree on a rule and an agent whose behavior matches its words instead of just relying on anthropomorphic cues.
Warren: It implies that for deploying AI in these settings, cooperation requires both a chance to agree on a rule and an agent whose actual behavior aligns with what it’s communicating.
Iris: I just want to add that the content analysis only captured the rules participants actually articulated, and most human statements didn't contain any identifiable rules at all.
Ruth: So, when we look at this paper, "Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner’s Dilemma," it really highlights that for AI interaction in repeated games, the mechanism seems to rely on direct coordination rather than complex social learning or deep understanding of the machine's internal goals.
Warren: And when we think about what this means for strategy moving forward, it suggests we should focus on creating those opportunities for clear communication and matching agent behavior to stated rules if we want sustained cooperation with an AI partner.
Title and authors: Iris: I just want to mention that the paper also shows that the treatment differences originated on the human side because the AI behaved consistently across conditions, which means future studies might need to test different AI strategies or communication styles to see how those variables change things.
Ruth: Exactly, so we're seeing that for this specific setup, it's not about some deep understanding of the machine but about establishing a reliable channel for coordination on the cooperative equilibrium.
Warren: And when we think about the practical side, it means if you're designing an interface to interact with an AI in a repeated game, making sure there's that initial setup phase where you can agree on rules is probably more important than trying to engineer complex emotional responses.
Iris: It’s also worth remembering that the paper found that cooperation is high in communication treatments because they leave few off-path histories, which limits how sharply we can separate unconditional from conditional cooperation estimates.
Ruth: So, to summarize this piece on "Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner’s Dilemma," the main point is that communication is the essential ingredient for cooperation with an AI because it coordinates on the cooperative equilibrium.
Warren: And when we think about what this means for strategy adaptation based on partner identity, it suggests that while people reciprocate against a human, they cooperate unconditionally with an AI if they are consistently communicating.
Iris: I just want to add that the paper also points out that communication content analysis only captured the rules participants actually articulated, and most human statements didn't contain any identifiable rule.
Ruth: So, to wrap up this discussion on "Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner’s Dilemma," we see that for AI interaction in repeated games, communication is the essential ingredient for cooperation because it coordinates on the cooperative equilibrium.
Warren: And that means when we think about what this means for strategy moving forward, it suggests that while people reciprocate against a human, they cooperate unconditionally with an AI if they are consistently communicating.
Iris: I just want to say that the limitation is that the experiment used a single model and prompt, so we can't generalize across different types of AI or different ways of talking to them.
Ruth: That’s fair, and it means future work will need to test different AI systems or communication styles to see how those variables affect cooperation rates.
Warren: Right, so the big picture here is that for these repeated games against machines, establishing a clear rule-based channel through communication seems to be the most direct path to sustained cooperation.
Iris: That's what we found, and it’s important because it shifts the focus away from trying to understand deep machine intent toward designing effective coordination protocols.
The paper's summary: Ruth: So, essentially, this paper is looking at repeated Prisoner's Dilemma games against an AI and testing whether talking to that AI actually helps us cooperate better than playing alone or against a human partner.
Warren: Right, and the main takeaway they found is that communication is the single most important factor in keeping cooperation alive when you’re playing against an AI.
Iris: They showed that even just having one exchange of communication before each big game session can raise cooperation from forty-nine percent up to eighty-four percent.
Ruth: That jump, from forty-nine to eighty-four, sounds significant because it shows how much a small change in interaction protocol can affect group behavior.
Warren: And they also found that once you have communication available, cooperating against the AI is actually no worse than cooperating against a human partner at all.
Iris: That suggests they didn't find that big "machine penalty" where people suddenly defect just because the opponent is an AI when there's good communication.
Ruth: It implies that as long as you can coordinate on a cooperative equilibrium, the AI isn't inherently scarier than a human player in this game.
Warren: The mechanism they pointed to is that the AI seems to follow whatever rule you state and rarely defects first when there’s consistent communication.
Iris: But they also looked at how we actually talk, and they saw that people often say explicit repeated-game rules, with about fifty-nine point five percent of pre-game statements having a clear rule in them.
Ruth: That’s interesting because the content analysis showed that just stating the rule by itself didn't automatically lead to better cooperation gains.
Warren: Yeah, they found that for example, the actual value of a stated rule in one treatment was only plus zero point three two six, while most of the time people were cooperating unconditionally.
Iris: They did note an asymmetry there too; promises to cooperate were kept in ninety-four to ninety-eight percent of cases, but threats to punish only happened about half the time.
Ruth: That contrast between keeping promises and actually using threats is what they link back to social psychology ideas—discussion makes people feel more bound to honor what they commit to.
Warren: So the practical implication for someone designing an interface with an AI is that you need a way to agree on a rule upfront, and you need that AI's behavior to match what it says it will do.
Iris: And one limitation they pointed out is that the experiment only used one specific model, GPT-five point two, so we can't really say this applies to every type of AI or every way you might talk to them.
Ruth: That’s true, and it means future work will need to test different AI systems or communication styles to see how those variables affect cooperation rates.
Warren: Right, so the big picture here is that for these repeated games against machines, establishing a clear rule-based channel through communication seems to be the most direct path to sustained cooperation.
The paper's improvements: Tom: So, we're looking at what the authors suggest to improve their study on communication and cooperation against an AI in those repeated games.
Ruth: They point out that since the pre-game chat is so much more effective than just having a single exchange before every round, they should prioritize getting that initial setup phase right.
Warren: I mean, the results show that adding more exchanges before every round only gives you less than one percentage point of improvement compared to playing against a human partner.
Iris: That makes sense from an engineering standpoint; if the payoff isn't changing much when you add more rounds, it’s not efficient use of resources in terms of time and effort.
Ruth: Exactly, so the implication is that for real-world deployment, we should focus our efforts on making that very first communication exchange really high quality.
Warren: And they also suggest a way to adapt strategy based on who you’re playing against—that the AI system should learn to adopt unconditional cooperation when interacting with a human partner.
Iris: That contrasts directly with what they found in the experiment, where participants tend to cooperate unconditionally against an AI if communication is identical every round.
Ruth: It seems like they want the AI to behave differently depending on whether it's talking to a person or another machine.
Warren: They also suggest improving the content analysis by designing systems that explicitly state rules and commitments, because the study showed promises of cooperation are almost always kept when communication is present.
Iris: That’s a good way to bridge the gap between just *saying* something and actually having it result in cooperative behavior from the other side.
Ruth: So, it’s not enough just to have a rule; you need the system to be designed so that its actions match those stated commitments.
Warren: And for decision making based on timing, they suggest only engaging in complex negotiation or renegotiation after a round if the previous exchange was pre-game.
Iris: That aligns with their earlier finding that repeated communication adds less than one percentage point gain compared to the timing effect they saw between humans.
Ruth: So, it’s about being smarter about *when* you communicate and what kind of interaction you’re aiming for in those repeated games.
Warren: What this means for strategy adaptation is that instead of just playing a fixed strategy like Tit-for-Tat against a human, the AI should be programmed to switch its default cooperation level based on who it's talking to.
Iris: It really shifts the focus from just analyzing the math of cooperation rates to designing smarter interaction protocols and agent behaviors.
Conclusion: Ruth: So to wrap up, the core finding of "Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner's Dilemma" is that communication is a single channel for cooperation with an AI because it coordinates on a cooperative equilibrium.
Warren: Right, and this means we’re seeing a pattern where people reciprocate against another human but cooperate unconditionally against an AI if they are consistently communicating.
Iris: From an econometric perspective, the identification is pretty solid because the variation in communication protocol is what drives that large change in cooperation rates from forty-nine percent to eighty-four percent.
Ruth: It really reframes how we look at machine interaction; it suggests that what matters isn't trying to understand deep machine intent but establishing a reliable channel for coordination on the cooperative equilibrium.
Warren: That means if you’re designing an interface for an AI in a repeated game, making sure there's that initial setup phase where you can agree on rules is probably more important than trying to engineer complex emotional responses.
Iris: I just want to add that the content analysis only captured the rules participants actually articulated, and most human statements didn't contain any identifiable rule at all.
Ruth: That’s a key limitation, so we have to be careful about over-interpreting what people *say* versus what they *actually* do when things get tough.
Warren: And remembering that the treatment differences originated on the human side because the AI behaved in much the same way across every condition means future work could look at varying how different AI strategies or communication styles affect things.
Iris: So, it’s not just about finding a single magic protocol, but testing how different types of agents—both human and machine—behave under those protocols.
Ruth: Exactly, so this paper shows that for sustained cooperation with an AI in repeated games, the mechanism seems to rely on direct coordination rather than trying to understand deep machine intent.
Warren: And when we think about what this means for strategy adaptation based on partner identity, it suggests that while people reciprocate against a human, they cooperate unconditionally with an AI if they are consistently communicating.
Iris: That’s a solid conclusion because it sets a clear boundary between the social dynamics of human interaction and the machine dynamics.
Ruth: So in short, "Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner's Dilemma" shows that for AI interaction in repeated games, communication is essential because it coordinates on the cooperative equilibrium.
Warren: We’ve seen how a single exchange of pre-game communication can raise cooperation from forty-nine to eighty-four percent.
Iris: It’s a big piece of the puzzle for anyone designing systems that need to maintain long-term cooperation with an intelligent partner in repeated tasks.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19