Introduction

Schizophrenia is a long-term mental health condition characterized by “positive” symptoms such as delusions and hallucinations, “negative” symptoms such as lack of motivation and reduced emotional expression, as well as cognitive impairment and social disconnection1. While positive symptoms can vary in intensity and may even disappear for extended periods, negative symptoms tend to be more stable and persistent. These long-lasting symptoms are largely responsible for the difficulties people with schizophrenia experience in social, family, and work life2.

Schizophrenia affects about 23 million people worldwide, or roughly 1 in 345 individuals (0.29%), with a rate of 1 in 233 (0.43%) among adults3. Symptoms usually begin in late adolescence or early adulthood, often appearing earlier in men than in women. Research from multiple fields - including psychophysiology, genetics, and neurobiology - indicates that schizophrenia may result from several different biological and environmental pathways that converge to produce the clinical picture of the disorder4. It is a highly heritable condition, influenced by many genes, with genetic factors accounting for around 80% of the risk5. However, psychosocial factors also contribute to the development of schizophrenia. Evidence shows that complications during brain development6, chronic stress - particularly related to childhood trauma7, migration or belonging to an ethnic minority group8 and substance abuse during adolescence9 can all increase the likelihood of developing the disorder.

Because schizophrenia presents a wide range of symptoms and levels of impairment that affect nearly every aspect of a person’s life, it is essential to achieve an early and accurate diagnosis that can identify markers of vulnerability and illness severity10. In this context, experimental approaches based on game theory have shown promise for improving diagnosis and tailoring psychosocial interventions to individual needs11. Consistent with the move toward describing mental disorders through observable behaviors rather than categorical diagnoses, Robson et al.12 propose using economic games as a more valid and objective “snapshot” of social functioning than traditional interviews or self-report questionnaires.

In this study, we use the Prisoner’s Dilemma (PD) game to analyze patterns of cooperative behavior in patients with schizophrenia, with the aim of characterizing behavioral differences associated with the disorder. The PD game provides an ideal framework for studying social interaction, as subjects must consider interdependency and coordinate their actions while trying to anticipate other people’s intentions. In repeated versions of the game, players must not only recognize that mutual cooperation yields the highest overall benefit but also understand the evolving dynamics and potential long-term consequences of their actions. Given that schizophrenia is characterized by reduced social functioning and difficulties in understanding others’ mental states13, the PD game offers a valuable paradigm for assessing adaptive social abilities and advancing in the development of behavioral diagnostic tools.

Experimental research has shown that cooperation in the PD increases when the likelihood of future interactions with the same partner is higher14,15. Accordingly, our study includes both one-shot PD games with randomly assigned partners and repeated PD (RPD) games with fixed partners. With this design, our objective is to examine how individuals with schizophrenia adjust their cooperative behavior to the likelihood of future interaction, and to compare these patterns with those observed in non-clinical control participants.

Only a few studies have employed the Prisoner’s Dilemma (PD) paradigm to compare the behavior of individuals with schizophrenia to that of non-clinical controls. Currie et al.16 examined loss aversion using a repeated PD (RPD) task in which 20 patients with schizophrenia and 16 control participants played against a pre-programmed computer under two conditions: a “gain frame,” where all outcomes resulted in monetary gains, and a “loss frame,” where money could be both won and lost. They found no significant differences in cooperative behavior between groups; however, unlike controls, patients with schizophrenia showed no loss aversion, suggesting a reduced adaptive response to contextual framing.

Using a similar approach, Bitsch et al.13 implemented a modified repeated PD game in which participants—patients with schizophrenia and controls—interacted with three pre-programmed partners that followed fixed strategies: cooperative, competitive, or random. Their findings showed that patients adjusted their behavior less frequently in response to partners with differing cooperative tendencies, indicating an impaired ability to update mental representations of others’ intentions. Interestingly, patients with schizophrenia cooperated more than controls when interacting with a competitive partner. A limitation of this study is that participants were told they were playing with human partners, when in fact, they interacted with computer programs.

In contrast to Bitsch et al. and Currie et al.13,16, we conducted a controlled laboratory experiment with a larger clinical sample. Participants − 44 patients with schizophrenia and 59 non-clinical controls - played against an artificial agent (AA) trained on previous experimental data to simulate human decision-making in the same games, rather than following a fixed strategy. This setup created stronger incentives for participants to continuously adapt their beliefs and behavior based on the evolving history of play and the type of game. Consequently, our experimental design provides a more sensitive framework for detecting behavioral markers associated with schizophrenia. Our artificial agent serves as a logistical tool that enables the study of repeated interactions in the Prisoner’s Dilemma game under both “stranger” and “partner” matching protocols. Importantly, this approach allows us to implement such interactions in the hospital, without the constraints imposed by time synchronization. Additionally, we also collected data on participants’ beliefs and emotions to better understand the factors underlying potential differences in cooperative behavior between patients with schizophrenia and control participants.

Another relevant study by Ellett et al.17 examined the behavior of patients with schizophrenia experiencing active persecutory paranoia in a one-shot PD game, finding that these individuals cooperated less than controls. The authors interpreted this reduced cooperation as resulting from heightened distrust toward other players. In contrast, our sample consisted of patients with schizophrenia who were clinically stable and did not exhibit current persecutory paranoia.

Other studies have employed a variety of experimental games to investigate the behavior of patients with schizophrenia in strategic contexts. Across these studies, patients consistently exhibit lower behavioral adaptation than non-clinical controls, although the direction of differences in cooperative behavior varies depending on the specific game design and implementation. For example, using a public goods game (analogous to a group Prisoner’s Dilemma), Chung et al.18 found that fear of losing money and greed for relative gains led controls to free ride, whereas patients with schizophrenia behaved in a non-strategic manner, showing low sensitivity to losses and resulting in highly cooperative behavior. In contrast, Hanssen et al.19, employing a different public goods design, reported that while patients’ responses to social enforcement did not differ from controls, they were initially less cooperative in group interactions. Consistent with this pattern, Purushothaman et al.20, using a repeated stag-hunt game, found that patients with schizophrenia shifted from high-risk to low-risk options at lower payoffs than controls, indicating impaired cooperation in risky situations.

Based on the literature reviewed above, we hypothesize that patients with schizophrenia will exhibit less adaptive behavior in response to changes in the PD framework. Specifically, the difference in cooperative behavior between the one-shot and repeated Prisoner’s Dilemma is expected to be smaller for patients with schizophrenia than for controls. Consistent with the difficulties in understanding others’ intentions reported by Bitsch et al.13, we also anticipate that patients’ predictions of their counterpart’s behavior in the PD game will be less accurate than those of controls. Additionally, our study contributes to ongoing discussions regarding differences in cooperative behavior between patients with schizophrenia and non-clinical participants.

The remainder of the paper is structured as follows: the next section presents the results, followed by a discussion and concluding remarks. The methods are detailed at the end, and experimental instructions, along with supplementary statistical analyses, are provided in the “Supplementary Information” file.

Results

Individual cooperation dynamics

Figure 1 illustrates the evolution of cooperative behavior in the two samples (44 clinical patients vs. 59 non-clinical participants). In the sequence of one-shot PD games, the cooperation rate among controls is initially low (10–20%) and declines rapidly, falling below 5% after round 5 and reaching 0% by the final round. In contrast, patients with schizophrenia maintain substantially higher cooperation rates, remaining above 30% until round 6 and showing a gradual downward trend; their cooperation rate is still 34% in round 10. A Pearson’s χ² test indicates that, across all rounds of the one-shot PD game, patients with schizophrenia exhibit significantly higher cooperation than controls (see Table SI5.1 in the Supplementary Information). In the first PD game, the 35-percentage-point difference in initial cooperation yields a large effect size (Cohen’s d = 0.81). Overall, a stark difference emerges between the two samples: while only 2% of the non-clinical participants cooperate more than twice in the sequence of ten one-shot PD games, 68% of the schizophrenia patients exhibit this behavior.

Result 1

Across all one-shot Prisoner’s Dilemma games, the clinical sample exhibited a significantly higher rate of cooperation than the control group.

Fig. 1
Fig. 1
Full size image

Mean individual cooperation rates by task, round, and treatment. Shaded areas represent 95% confidence intervals around the mean.

As shown in Fig. 1, in the repeated PD games (RPD1 to RPD3), the cooperative behavior of the non-clinical sample differs markedly from that observed in the one-shot PD. Controls’ average cooperation at the start of each RPD ranges from 30% to 40%, higher than in the one-shot PD. The differences are significant for RPD2 and RPD3 compared to the one-shot PD, based on McNemar tests with Bonferroni-corrected p-values of 0.014 and 0.013, respectively (RPD1: p = 0.087). Cooperation remains high through round 8 of each RPD, after which it declines sharply to below 10%.

In contrast, the behavior of patients with schizophrenia does not differ across tasks and rounds. For this group, the cooperation rate in the first round of each RPD task does not differ significantly from that in the first round of the one-shot PD, based on McNemar tests with Bonferroni correction. Average cooperation is approximately 33% and does not decline toward the end of the repeated games, contrary to the typical “end-of-the-world” effect. Notably, patients with schizophrenia cooperate significantly more than controls in the final round of all RPD tasks (Pearson’s χ² test, p < 0.001). As hypothesized:

Result 2

While controls adjust their behavior according to the type of PD game (one-shot or repeated), the cooperation rate of patients with schizophrenia remains similar across games and rounds.

Beliefs

Figure 2 shows the percentage of players who believe their partner will cooperate in each round (first-order belief, that is, the answer to question 1 reported in Section SI2 of the Supplementary Information) by task, round and treatment.

Fig. 2
Fig. 2
Full size image

Percentage of players whose belief is partner cooperation by task, round, and treatment. Shaded areas represent 95% confidence intervals around the mean percentage.

Overall, the control group’s expectations of their partner’s cooperation increase across tasks, rising from an average of about 18% in the one-shot PD to 53% in RPD3. In contrast, participants in the clinical group maintain relatively stable expectations, averaging around 40% partner cooperation across all tasks. This pattern suggests that the limited behavioral adaptation observed among patients is accompanied by a reduced ability to update expectations about their partner’s intentions.

Eliciting participants’ beliefs allows us to measure how accurately they predicted their artificial agent (AA) partner’s behavior in each round - that is, whether they correctly expected cooperation when the partner cooperated or defection when the partner defected. Figure 3 shows the percentage of correct beliefs by task, round, and treatment. Overall, belief accuracy increases progressively across rounds in all tasks. However, the control group’s accuracy improves more rapidly than that of patients with schizophrenia. Consistent with our hypothesis, in the second half of each task, controls’ beliefs are significantly more accurate than those of patients with schizophrenia (see Table SI5.2 for detailed test results).

Fig. 3
Fig. 3
Full size image

Percentage of correct individual beliefs by task, round, and treatment. Shaded areas represent 95% confidence intervals around the mean percentage.

Result 3

The accuracy of patients with schizophrenia in predicting the artificial agent’s behavior in the PD games increases over rounds at a slower rate than that of the control group.

Emotions

As shown in Fig. 4, the large majority of participants report that their decisions are driven by reason rather than emotion. For analysis, self-reported emotions were grouped into three categories: “positive” (empathy, happiness, compassion, and excitement), “negative” (guilt, greed, fear, and regret), and “other.” Approximately 2% of participants in the control group and 8% in the clinical group reported “other” emotions. This difference may reflect the difficulties patients with schizophrenia experience in identifying and describing emotions compared with healthy controls21. These responses were excluded from Fig. 4.

Across tasks, an average of 88.4% of controls indicated “reason” as their main decision motive, compared with 73.4% in the clinical group—a statistically significant difference (Mann–Whitney z = − 3.541, p < 0.001). Furthermore, patients with schizophrenia reported positive emotions as a decision motive significantly more often than controls (Mann–Whitney z = 3.131, p = 0.002).

Fig. 4
Fig. 4
Full size image

Percentage of subjects reporting a reason- or emotion-based decision motive by treatment. “Other” emotions are excluded. Mann–Whitney test p-values: *** < 0.1%, ** < 1%.

Regression analysis

We further analyze the difference between our two treatments by means of panel logit regressions of the likelihood of individual cooperation on beliefs, emotions, tasks, and stage-game rounds within the tasks, controlling for age and gender. Table 1 describes the variables used in the analysis, while the estimated marginal effects for controls and patients with schizophrenia are presented in Table 2.

Table 1 Description of variables used in the regression analysis.
Full size table
Table 2 Marginal effects from random-effects panel logit regressions of individual cooperation by treatment.
Full size table

In the regression for the non-clinical sample (first column of Table 2), controls’ individual belief that their partner will cooperate and self-reported positive emotions significantly increase the probability of cooperation. Cooperation also rises significantly in the RPD tasks, using the one-shot PD game as the baseline. However, controls’ likelihood of cooperation decreases significantly as rounds progress within each task, with a pronounced “end-of-the-world” effect.

In contrast, as shown in column 2 of Table 2, the coefficients for task and round dummies are not significantly different from zero for the clinical sample. That is, patients with schizophrenia behave similarly across rounds and regardless of whether they interact once or repeatedly with other players, supporting the “lack of adaptation” hypothesis. Furthermore, self-reported emotions do not appear to influence cooperative behavior in this group. The only significant predictor is the belief that the AA partner will cooperate.

As a robustness check, Table SI5.3 in the Supplementary Information shows that including measures of symptom severity (the sum of items from the General Psychopathology scale of the Positive and Negative Syndrome Scale) and medication dosage (chlorpromazine-equivalent milligrams) in the regression for patients with schizophrenia does not alter the results reported in Table 2. Moreover, neither variable has a statistically significant effect on individual cooperation.

We next perform panel logit regressions to estimate the likelihood of cooperation based on the type of PD game played (one-shot vs. repeated). Estimated marginal effects are reported in Table 3.

Table 3 Marginal effects from random-effects panel logit regressions of individual cooperation by nature of the game (one-shot vs. repeated PD).
Full size table

This regression indicates that, even after controlling for factors that may influence cooperation, patients with schizophrenia are significantly more likely to cooperate than controls. The effect is particularly pronounced and statistically significant in the one-shot Prisoner’s Dilemma.

Strategy analysis

In each task, we classified participants’ strategies according to those most commonly considered in cooperative games: tit-for-tat, always cooperate, always defect, grim, or other. The last two rounds of each task were excluded to avoid the “end-of-the-world” effect. Thus, “always cooperate” or “always defect” indicates that a participant consistently cooperated or defected in the first eight rounds of the game, respectively. Tit-for-tat was defined as reciprocating the partner’s previous cooperation or defection in more than 50% of rounds where this was possible, provided the participant did not qualify as always cooperating or always defecting. Grim was defined as ceasing to cooperate immediately after a partner’s defection and never returning to cooperation. Although tit-for-tat and grim strategies are not applicable to one-shot PD games, we retained this terminology for consistency and ease of comparison across tasks. Any behavior that did not fit the defined taxonomy was categorized as “other.”

As shown in Fig. 5, the clinical sample adopted the “Always Cooperate” strategy in 2.3% of cases in the one-shot PD, whereas the non-clinical sample never employed this strategy in that task. However, the latter used it much more frequently (14–19% of cases) in the RPD. The relatively high use of the “Always Defect” strategy among the non-clinical participants should be interpreted in light of the fact that they were interacting with an artificial agent rather than a human partner22.

Another noteworthy finding is that, in most cases, the strategy followed by the participants with schizophrenia could not be identified, while this occurred in at most 17% of cases in each task for the non-clinical sample.

Fig. 5
Fig. 5
Full size image

Relative frequencies of strategy adoption by task and treatment. Strategies are defined from behavior in rounds 1–8.

Discussion and concluding remarks

This study presents results from an incentivized experiment comparing cooperation in one-shot and finitely repeated Prisoner’s Dilemma games between patients with schizophrenia and a non-clinical control group. We find that patients with schizophrenia show limited adaptation to the type of PD game played, displaying similar cooperation rates in both one-shot and repeated settings. Although our design and framework differ, these results are consistent with Currie et al.16, who found that individuals with schizophrenia exhibit reduced flexibility in adapting to changing interaction environments. In contrast to Ellett et al.17, however, we observe that patients cooperate significantly more than controls in the one-shot PD. A possible explanation for this discrepancy is that, in our study, the patients did not experience current persecutory paranoia.

By eliciting the beliefs of patients with schizophrenia in the PD game, we provide behavioral evidence supporting Bitsch et al.’s13 neurological findings that updating social knowledge and forming forward predictions in social interactions are impaired in schizophrenia. Specifically, we find that patients’ belief accuracy regarding their partner’s intentions improves over time at a slower rate than that of the control group. Furthermore, the strategies they employ in response to their partner’s behavior are largely unidentifiable.

Our results highlight the difficulties that patients with schizophrenia face in thinking strategically and adapting their behavior to both the environment and the actions of others. By enabling the characterization of behavioral differences through observed decision-making rather than relying solely on self-reported measures, our game-theoretic experimental approach offers a valuable complement to traditional clinical assessments, which may sometimes lack accuracy and reliability23. In our experiment, the observed behavioral patterns allow for a clear distinction between patients with schizophrenia and non-clinical participants in our study. Further research using larger and more representative samples, as well as including other psychiatric conditions, is needed to assess the robustness and generalizability of these findings. More broadly, integrating decision-making data from economic experiments with standard psychometric assessments appears to be a promising avenue for improving our understanding of behavioral heterogeneity in mental health disorders.

Methods

Participants and procedures

Our experiment was conducted with two distinct populations: undergraduate university students (for the non-clinical sample) and hospitalized patients with schizophrenia (for the clinical sample). We first collected data on 59 students in our Laboratory for Experimental Economics (LEE) at Jaume I University (Spain). Given the magnitude of the difference between first-round cooperation in the PD task and in RPD3 for our non-clinical sample (difference = 0.2373, correlation = 0.0857), an ex-ante power analysis recommended a sample size of 43 to achieve 80% power with 5% significance in a one-tailed McNemar paired-proportions test. Thus, we collected data from a clinical sample of 44 patients with schizophrenia, recruited from the Short-Term Hospitalization Unit (UHB) at the Provincial Hospital of Castellón, Spain. Both samples were gender-balanced (49.15% male in the non-clinical sample and 50.0% male in the schizophrenia sample). The mean age of the clinical sample was 36.5 years, while that of the student sample was 22 years. As this difference might have an effect on observed behavior, we control for it in our regression analysis.

Undergraduate students were recruited from various university degree programs using the ORSEE platform24. Participants provided informed consent to take part in economic experiments upon registration, with the assurance that their personal data would be protected in accordance with the General Data Protection Regulation (GDPR). The experiment was fully computerized and conducted in a specialized computer laboratory (LEE, Universitat Jaume I) using software developed with the z-Tree toolbox25. Specifically, 59 participants played ten one-shot and three finitely repeated Prisoner’s Dilemma (RPD) games against an artificial agent (AA) previously trained on data from “all-human” sessions to mimic human decisions. Both decisions and belief elicitation tasks were monetarily incentivized and paid in cash at the end of the session. The experimental instructions, screenshots of the computer interface, along with details about the artificial agent, are provided in the Supplementary Information (Sections SI1 and SI3 respectively).

Informed consent was obtained from all participants, with assurances that their medical information would remain strictly confidential and be protected in accordance with applicable data protection regulations. The confidentiality of participants’ data and identities was strictly ensured. No minors participated in this study. The research was conducted in accordance with the guidelines of the World Medical Association’s Declaration of Helsinki. The experimental protocol was approved by the Ethics Committees of the Provincial Hospital of Castellón (CEIM-42-2.1.1, November 25, 2022) and Universitat Jaume I (CEISH/17/2022, March 24, 2023).

The entire clinical sample was evaluated and classified by clinical personnel as adult patients who met criteria for schizophrenia according to the Diagnostic and Statistical Manual of Mental Disorders-IV (DSM-IV). Patients with other different psychiatric diagnoses (excluding addictive disorders) or those with physical or neurological limitations preventing the completion of the experimental task were excluded from the study.

Sociodemographic information was collected through an admission interview. For clinical variables, the Psychiatric Research Interview for Substance and Mental Disorders for DSM-IV (PRISM-IV) was used to exclude patients with diagnoses other than schizophrenia. If negative, the Positive and Negative Syndrome Scale (PANSS26, was used to evaluate the positive and negative syndrome of schizophrenia from both a dimensional (severity) and categorical (schizophrenia) perspective (schizophrenia positive, negative and mixed). Research assessments were administered after a mean of 7–10 days of pharmacological treatment, once psychopathological stabilization had been achieved. Descriptive statistics of the clinical sample are presented in section SI4 of the Supplementary Information.

We considered that patients’ capacities for concentration and attention were sufficient to undertake the experimental tasks. Each patient was individually assisted to ensure full understanding of the instructions and the software. At the beginning of the experimental session, patients were informed that they could opt out of the experiment at any time. The same software and procedures used for the non-clinical sample were applied to elicit behavior in the PD games through incentive-compatible methods for the sample of 44 schizophrenia patients. Heerey et al.27 observed normal sensitivity to monetary rewards in patients with schizophrenia.

A possible source of bias in the observed behavior could be demand effects due to the participants’ willingness to comply with perceived rules and actions they might consider desirable by the experimenters. To minimize or at least equalize any possible demand effects across both populations, we conducted the experiments in the presence of the same group of experimenters and technical support personnel.

Psychiatric measures

The Psychiatric Research Interview for Substance and Mental Disorders version IV (PRISM-IV) was developed by Torrens et al.28 as a semi-structured clinician-administered interview to measure the major Axis I DSM-IV diagnoses (current and past) of alcohol, drug and psychiatric disorders. It provides clear guidelines for differentiating among the expected effects of intoxication and withdrawal, substance-induced disorders, and primary disorders. The PRISM-IV also covers two Axis II disorders: Borderline Personality Disorder and Antisocial Personality Disorder. Although primarily designed as a research instrument, the PRISM-IV includes coverage of alcohol -and drug- related experiences and symptoms that may be relevant for identifying areas of focus for medical treatment. In this study, the interview is used to provide the differential diagnosis between schizophrenia and other psychiatric disorders.

The Positive and Negative Syndrome Scale (PANSS) is a standardized clinical interview developed by Kay et al.26 that rates the presence and severity of positive and negative symptoms, as well as general psychopathology, in people with schizophrenia. Of the 30 items, seven assess positive symptoms, seven assess negative symptoms, and 16 assess general psychopathology symptoms. Symptom severity for each item is rated on a 7-point scale (1 = absent; 7 = extreme).

Experimental economics measures

As reported in the introduction, previous experimental studies have found that patients with schizophrenia exhibit reduced adaptability to their environment13,16. Further empirical evidence indicates that deficits in theory of mind are common among these patients, as reported in the meta-analyses by29,30,31. Nassar et al.32 observed that these patients are more prone to ignore new information and rely excessively on prior knowledge. This updating pattern hinders “the integration of information over time, reducing both the flexibility and precision of beliefs.” Similarly, Achim et al.33 found that impaired theory of mind in patients with schizophrenia is associated with difficulties inferring others’ mental states and generating appropriate responses to perceived intentions and behaviors. Consequently, we employed both one-shot games with randomly matched partners and repeated Prisoner’s Dilemma games with fixed partners to design an environment that assesses participants’ adaptability. Moreover, we elicited participants’ beliefs using an incentivized procedure.

PD games

Each PD session began with a set of training questions to ensure that participants fully understood the rules and payoff structure of the game. After completing the training phase, participants played four consecutive PD tasks. They were informed that only one of these four tasks - randomly selected at the end of the session - would determine their final payment based on their decisions.

One-shot PD games

The first task involved a series of ten one-shot Prisoner’s Dilemma (PD) games. Participants interacted with an artificial agent trained on data from earlier sessions where only human subjects played the same games. Importantly, the agent’s choices were made independently of any prior interactions with the participant.

Table 4 Payoffs of the one-shot game.
Full size table

Table 4 presents the payoff matrix for the one-shot PD game. As shown in the table, option A represents the decision to cooperate, while option B corresponds to the decision not to cooperate. To avoid endowment effects across the sequence of one-shot PD games, the Random Lottery Incentive (RLI) system was employed as the payment mechanism. Specifically, if this task was selected for payment, only one randomly chosen PD game determined the participant’s remuneration.

Finitely repeated PD games (tasks RPD1 to RPD3)

In the last three tasks, participants played a repeated PD game, in which each subject played 10 rounds of the same game. From a design perspective, using multiple RPD games allows us to observe how participants adapt their behavior across distinct repeated interactions. This structure enables us to capture learning effects, changes in beliefs, and the evolution of cooperation both within a single repeated game and across multiple repeated experiences.

Each subject played each task and round with an AA that was trained to probabilistically implement decisions analogous to those made in the same task, round and with the same history of play by humans in previous sessions.

For all three RPD tasks, the payoff of each round was equivalent to one-tenth of the payoff from a round of the one-shot PD game. If one of the three RPD tasks was randomly selected for payment, participants received the total accumulated payoff from that task.

Beliefs and self-reported reason for the decision taken

To obtain more detailed insights into participants’ strategic reasoning, we elicited their beliefs about the partner’s likelihood of cooperation at the beginning of each round in every game, as well as their assessment of the probability of cooperation by both human and artificial agents in that round. These belief elicitation responses were incentivized. Immediately after each decision, we also asked participants whether their choice had been driven mainly by reason or a particular emotion. Further details are provided in the Supplementary Information (section SI2).