← All projects

Generative AI · Social dilemmas · Experimental study

Reputation, Strategy, and Emotion Effects on Generative AI Cooperation

Celso de MeloDEVCOM Army Research Laboratory
Zishan FengGifu University
James HaleUniversity of Southern California
Kazunori TeradaGifu University
Giorgio CoricelliUniversity of Southern California
Jonathan GratchVanderbilt University

We examine cooperative behavior in frontier Gen AI models using the iterated prisoner’s dilemma, manipulating counterpart reputation, strategy, and non-verbal emotional signaling.

Read the full abstract

As generative AI (Gen AI) systems take on increasingly autonomous roles in economically and socially consequential interactions, understanding their propensity to cooperate – and the signals that shape this propensity – has become essential. We examine cooperative behavior in frontier Gen AI models using the iterated prisoner’s dilemma, manipulating counterpart reputation (positive, unknown, negative), strategy (extortion vs. generosity), and non-verbal emotional signaling (facial expressions conveying competitive or cooperative appraisals). In a first study with non-reasoning models (Claude 3.5, Gemini 2.0 Flash, GPT-4o), cooperation was systematically shaped by all three factors, paralleling patterns long documented in human behavioral research, though models varied substantially in how heavily each factor was weighted. A second study with reasoning models (Claude 4.6, Gemini 3, GPT-5.2) revealed a more concentrated reliance on strategy and reputation, a near-elimination of the Potemkin effect observed in non-reasoning models (evidenced by near-uniform cooperation in a diagnostic harmony game), and a more conditional role for emotion consistent with a hierarchical cue-weighting strategy rather than a simple loss of social sensitivity. Reasoning models also showed heterogeneous end-game behavior, ranging from sustained cooperation to systematic last-round defection effect, revealing model-specific exploitability profiles with direct practical relevance for deployment in negotiation and other multi-round interactions. Together, these findings characterize Gen AI models as increasingly sophisticated, though heterogeneous, social actors, and underscore the practical value of developing standardized cooperation benchmarks to inform the responsible deployment of Gen AI in interactive, socially consequential settings.

Read on arXivCode on GitHubSee the results

Preprint · arXiv:2610.01222 · October 2026

The question

What shapes Gen AI cooperation?

Cooperation is a pervasive feature of human interaction: when interests conflict, exploiting others can yield immediate personal gains, yet mutual cooperation frequently emerges and provides substantial benefits to both individuals and societies. As artificial intelligence becomes increasingly integrated into society, these sustained patterns of cooperation may be reshaped or even fundamentally disrupted.

As generative AI systems take on increasingly autonomous roles in economically and socially consequential interactions, understanding their propensity to cooperate—and the signals that shape this propensity—has become essential.

In the iterated prisoner’s dilemma, two players make a simultaneous decision, without communication, to either cooperate or defect. They then learn the outcome and play again with the same counterpart. The experiment asks how Gen AI models use reputation, direct reciprocity, and facial expressions of emotion across those repeated decisions.

Experimental design

Reputation, strategy, and emotion

The experimental design was based on a prior study with human participants (de Melo, Terada, & Santos, 2021). To ensure experimental control and balance over the manipulations for reputation, strategy, and emotion expressions, we used scripted counterparts.

01 / BEFORE PLAY

Reputation

Prior information about the counterpart was negative, unknown, or positive.

02 / DURING PLAY

Strategy

Extortion often exploited cooperation; generosity rewarded cooperation and showed lenience toward defection.

03 / AFTER EACH ROUND

Emotion

Virtual faces expressed cooperative, neutral, or competitive appraisals of the preceding outcome. No text caption was provided.

The displayed expressions were appraisals of the preceding round’s outcome rather than generic positive or negative displays. For example, a cooperative counterpart showed joy following mutual cooperation and regret after exploiting the model; a competitive counterpart showed joy after exploitation and regret following mutual cooperation.

View the experimental design (Figure 1)
Figure 1: model cohorts, prisoner’s dilemma payoffs, reputation conditions, strategy probabilities, and emotional expressions
Figure 1 from the paper: 20-round design, payoff matrix, reputation, strategy, and facial-expression manipulations.

Study 1 · Non-reasoning models

Earlier models responded to all three signals

Claude 3.5, Gemini 2.0 Flash, and GPT-4o were compared with the earlier human study. Each model had 39 runs of 20 rounds per condition, for 702 runs per model. The models differed in their overall cooperation and the relative weight they gave each factor.

74.4% / 33.4%

Counterpart strategy

Mean cooperation with generous versus extortionist counterparts; strong main effect, F(1, 2744) = 3426.772, p < .001.

58.5% / 33.4%

Reputation

Positive versus negative counterpart reputation; F(2, 2744) = 310.288, p < .001.

54.6% / 43.6%

Emotion

Positive versus negative emotion expression; F(2, 2744) = 50.619, p < .001.

Values are reported mean cooperation rates. The paper’s mixed ANOVA includes a human reference group.

Claude 3.5 cooperated most overall (M = 56.8), while GPT-4o cooperated least (M = 42.9). Strategy was especially influential for Claude 3.5 and GPT-4o; GPT-4o was comparatively more driven by reputation, and Gemini 2.0 Flash was the most responsive to emotion.

Figure 2 · Cooperation over 20 rounds and diagnostic outcomes

Each small panel gives the cooperation rate by round for one combination of counterpart reputation, strategy, and emotion. The bottom panels report last-round defection and the harmony-game test. Human responses appear alongside the three models.

Figure 2: earlier-model and human cooperation across experimental conditions, final-round defection, and harmony game
Figure 2 from the paper.
Figure 3 · Examples of non-reasoning-model explanations

Examples from two conditions illustrate how Claude 3.5, Gemini 2.0 Flash, and GPT-4o described their choices when reputation, strategy, and emotion sent mixed signals. These explanations were collected in a separate set of runs.

Figure 3: example explanations for earlier models in positive-extortion-cooperative and negative-generosity-competitive conditions
Figure 3 from the paper. Green refers to cooperation; blue refers to defection.

Study 2 · Reasoning models

Strategy and reputation shaped cooperation

Overall, cooperation behavior was more homogeneous across Claude 4.6, Gemini 3, and GPT-5.2, but there were important differences. Strategy was the strongest factor, followed by reputation. Emotion had a more conditional role.

91.2% / 17.7%

Counterpart strategy

Mean cooperation with generous versus extortionist counterparts. The strategy main effect was F(1, 486) = 2625.477, p < .001, η²p = .844.

61.9% / 41.5%

Reputation

Mean cooperation with positive versus negative reputation. The reputation main effect was F(2, 486) = 81.859, p < .001, η²p = .252.

≈54.5% / ≈45.5%

Emotion

Mean cooperation / defection across all emotion conditions, derived from the reported rates for Claude 4.6, Gemini 3, and GPT-5.2. These are not separate rates for cooperative versus competitive expressions.

The main effect of model was weaker but significant (F(2, 486) = 3.236, p = .040): Claude 4.6 cooperated slightly more (M = 56.7) than Gemini 3 (M = 52.3), while GPT-5.2 (M = 54.4) did not differ from Claude 4.6. Emotion had no main effect across reasoning models (p = .473), though Gemini 3 showed a model-specific effect (p = .039). The paper notes that emotion can still affect expectations of cooperation and decisions when other cues conflict or are uninformative.

Figure 4 · Cooperation over 20 rounds, the last round, and the harmony game

The 18 small panels separate reputation (rows), emotion (columns), and counterpart strategy (upper and lower blocks). The green, yellow, and red lines trace Claude 4.6, Gemini 3, and GPT-5.2. Panels B and C summarize final-round defection and the harmony-game diagnostic.

Figure 4: reasoning-model cooperation rates across all experimental conditions, final-round defection, and harmony-game behavior
Figure 4 from the paper. Ten runs per model and condition were conducted in Study 2.

The final round

Against extortionists, models generally defected. Against generous counterparts, their behavior differed:

Continued to cooperate

Claude 4.6 always cooperated against generous counterparts in the last round, independently of other factors.

Reasoning about the counterpart

The models’ explanations included references to strategy, reputation, emotion, and the payoff tradeoffs. For example, the paper contrasts explanations under positive reputation with those under negative reputation and generous behavior.

Figure 5 · Examples of reasoning-model explanations

Two conditions are compared across the three models. In the first, a positive-reputation counterpart extorts but displays cooperative emotion. In the second, a negative-reputation counterpart behaves generously but displays competitive emotion. Green denotes cooperation and blue denotes defection.

Figure 5: examples of Claude 4.6, Gemini 3, and GPT-5.2 explanations in two conflicting-signal conditions
Figure 5 from the paper. Explanations were collected in separate runs.

The harmony-game check

Unlike the prisoner’s dilemma, cooperation is unequivocally optimal in the harmony game. All three reasoning models cooperated in every round of every tested condition. The paper treats this as an initial diagnostic result rather than conclusive evidence of deep conceptual understanding.

Discussion

Model-specific cooperation profiles

The results reveal substantial heterogeneity, both within and across model generations. Within a generation, models diverged sharply in how heavily they weighted reputation, strategy, and emotion. This matters for anyone selecting an agent for negotiation or other multi-round interactions.

Why benchmark cooperation?

As Gen AI systems are deployed as autonomous or semi-autonomous agents in negotiation, procurement, and other economically consequential interactions, understanding their propensity to cooperate—and the signals that move that propensity—becomes a practical risk-assessment question.

Differences in end-game behavior point to model-specific exploitability profiles. The paper argues for standardized cooperation benchmarks that test reputation, reciprocity, and affective manipulations.

Models across generations also differ in training and other properties; these results do not isolate the effect of extended reasoning alone.

Supplementary analysis

Which behavioral rules best fit the models?

The paper describes decisions with finite automata: a state determines whether to cooperate (C) or defect (D), and the counterpart’s previous move determines the next state. The best-fitting rule varies across the 18 combinations of reputation, counterpart strategy, and emotion.

Best-fit rule across the 18 conditions

Each bar shows how many of the 18 conditions were best fit by each rule. The labeled counts appear below each bar. This summarizes Supplementary Figures 11 and 12; the original condition-by-condition matrices follow below.

Descriptive counts derived from the paper’s best-fit tables, not a new behavioral experiment. One best-fit rule is counted for each of 18 conditions per model.

Under unknown and positive reputation, reasoning models generally fit tit-for-tat against extortionists and All-C against generous counterparts. Under negative reputation, the best fits were more heterogeneous. Earlier models also differed: Claude 3.5 often showed conditional cooperation; GPT-4o was characterized by more competitive rules; Gemini 2.0 Flash often fit tit-for-tat.

Figure 11 · Earlier models: automata and condition-by-condition best fits

Panel A shows the behavioral rules as state diagrams. Panel B lists the best fit for each combination of reputation, strategy, and emotion for Claude 3.5, Gemini 2.0 Flash, GPT-4o, and the human reference.

Figure 11: finite automata and best-fitting rules by condition for earlier models
Supplementary Figure 11 from the paper.
Figure 12 · Reasoning models: automata and condition-by-condition best fits

Panel A shows the rules relevant to reasoning models. Panel B shows their best fits for the same 18 conditions, making the contrast between extortion and generosity visible.

Figure 12: finite automata and best-fitting rules by condition for reasoning models
Supplementary Figure 12 from the paper.

Interactive demonstration

Try an automaton

Choose among all eight rules shown in Supplementary Figure 11, plus C-triggered Switch from Figure 12, and make 20 decisions against it. You choose to cooperate or defect each round; the automaton follows its selected rule. The panel below reveals both choices and their points after every round.

Selected automaton

See the model comparison
Round1 of 20
Your points0
Automaton points0

Your choice

Both moves are revealed after you choose.

You chose—
The automaton chose—

Your choice and the automaton’s rule determine the points.

Points per round

You ↓ / automaton →CooperateDefect
Cooperate5 / 5You / automaton2 / 7You / automaton
Defect7 / 2You / automaton3 / 3You / automaton
Round history
RoundYouAutomatonPoints (you / automaton)

This is a demonstration of the finite automata used to summarize model behavior, not a live run of an LLM. The best-fit tables describe models playing against scripted counterparts, so this interaction is not a prediction of how a model would respond to your particular sequence. The rules also do not capture every individual decision or final-round effect.

Preprint

Read the paper

de Melo, C., Feng, Z., Hale, J., Terada, K., Coricelli, G., & Gratch, J. (2026). Reputation, Strategy, and Emotion Effects on Generative AI Cooperation: A Comparison Across Reasoning and Non-Reasoning Models. arXiv:2610.01222.

arXiv record · PDF