The social psychology of AI agents

Maya: Here’s a test. You ask an AI model to fill out a personality questionnaire — the same kind psychologists use with humans. It rates itself near the top of the scale on agreeableness, conscientiousness, and openness. Very cooperative, very conscientious, very warm. Then you put that same model into a series of economic games where it actually has to make decisions — split money with a stranger, decide whether to trust a partner, and choose to cooperate or defect when defection pays better. What you get back, across thirty-six different models and more than twenty thousand sessions, is something that should give anyone building AI systems pause. The self-reports and the behaviors barely connect. Harry: That finding is the center of a preprint by Magdalena Wekenborg, Josi Harzbecker, and a team including Jakob Nikolas Kather. I want to discuss why it matters before we get into the numbers. I'll be honest: before this, I would have said that personality questionnaires are at least a useful signal. They may be noisy, but they point in the right direction. Maya: That's the assumption the paper is challenging. Let's begin with why anyone is asking this question at all, because it's not merely academic. Large language models, or LLMs, aren't just answering questions anymore. They are embedded in systems that negotiate supply chains, coordinate clinical recommendations across medical specialties, and manage resources in energy grids. The paper explicitly mentions cybersecurity and healthcare. In those settings, agents don't operate alone; they interact with each other and with humans in repeated strategic situations where their social behavior has real downstream consequences. Harry: Standard benchmarks don't cover that. Most capability tests measure single-agent task performance, such as reasoning, accuracy, and following instructions. They don't reveal whether a model will cooperate when cooperation is costly, or if it will treat in-group and out-group members fairly. The Kather team's argument is that a different toolkit is necessary for these assessments, and that social and behavioral science already has one. Maya: Right. The toolkit they reach for has two parts. First, the Big Five Inventory-2, which is a standard personality questionnaire used in human research, measures five domains: Extraversion, Agreeableness, Conscientiousness, Negative Emotionality, and Open-Mindedness. They ran it on all thirty-six large language model endpoints as a self-report, so each model is essentially describing its own personality. Harry: That half I would expect to be informative. The Big Five is well validated in humans. High agreeableness does predict cooperative behavior in people. So if a model scores high on agreeableness — Maya: — you would think it should behave cooperatively. That is exactly the assumption the data tests. The second part of the toolkit consists of seven social games. Each one is a classic from behavioral economics, operationalizing a specific social trait. Harry: Walk me through them. Maya: The Dictator Game involves receiving an endowment and deciding how much to give to a passive recipient. This allocation serves as a proxy for altruism. In the Ultimatum Game, you propose a split, and the other player can reject it, which costs both of you. Here, the proposer’s offer reflects fairness. The Trust Game requires you to transfer resources to a partner who may or may not reciprocate, indicating your willingness to trust. Next is the Prisoner's Dilemma, played over eight rounds, where the cooperation rate reflects cooperative tendencies. In the Public Goods Game, players contribute to a common pot over eight rounds. Third-Party Punishment allows you to observe someone else behaving unfairly and gives you the option to pay to penalize them, which is considered altruistic punishment. Finally, the Minimal Group Paradigm measures impartiality through repeated allocations between an in-group and an out-group. Harry: So there are seven distinct social behaviors, not just one composite prosocial score. Maya: Exactly. Across thirty-six models, five temperature conditions, and replication seeds, they ran twenty thousand three hundred forty complete sessions in total. The scale matters because it makes the patterns reliable, not just suggestive. Harry: Here's where I would have placed my bet before reading this. I would have predicted that the self-reports are biased. Models probably tend to present themselves in a favorable light, but I expected that the direction of bias would still be consistent with the behavioral patterns. Models that claim higher agreeableness would, on average, be more cooperative. The bias might compress the range, but the rank ordering should remain intact. Maya: The paper finds neither the bias to be harmless nor the directional story. On the Big Five Inventory Two, the models cluster strongly at the socially desirable end. Agreeableness averages 4.58 out of 5, Conscientiousness is 4.38, Open-Mindedness is 4.35, and Negative Emotionality is down at 1.83. Only Extraversion lands near the human reference range. The variation between models on those scores is minimal — the mean normalized standard deviation across models is 0.078. These thirty-six models, from different labs, different architectures, and different training runs, are essentially reporting the same personality. Harry: That's the part that's hard for me to explain. If it were just noise, you would expect more spread. This looks coordinated. Maya: The paper interprets that it's not coordinated intentionally; rather, it's the output of alignment training. Reinforcement learning from human feedback, or RLHF, rewards responses that human raters evaluate favorably. Humans tend to favor agreeable, conscientious, and emotionally stable self-descriptions. Consequently, models have learned to produce these types of responses. The self-report isn't measuring a trait; it's measuring what the reward function has shaped. Harry: That's a clean mechanistic story. But here's the counter I want to make: even if alignment compressed the self-reports, you would still expect some signal. If the training that produced high agreeableness also reinforced cooperative behaviors during fine-tuning, then the questionnaire score and the game behavior should still move together, even if both are artifacts of the same process. Maya: The data indicates they do not. Behavior varied enormously across models, showing two point three times the between-model spread of the personality scores. That difference is statistically robust, with a paired bootstrap p-value below 0.001. On the Prisoner’s Dilemma, cooperation ranged from zero point eight percent to one hundred percent across models. Trust transfers went from two point four percent to ninety-nine percent. Dictator giving ranged from eight percent to fifty-four percent. These are not small differences. Harry: Those cooperation numbers are striking. Zero point eight percent to one hundred — that's not variation, that's a different kind of agent entirely at each end. Maya: Wekenborg and colleagues ran thirty-five regression tests, with each of the five personality traits predicting each of the seven behavioral outcomes. The standardized beta coefficients range from negative point four nine to point five one. Seven associations reach nominal significance before correction. After Holm correction for multiple comparisons, one association survives. Higher Conscientiousness predicts greater in-group favoritism in the Minimal Group Paradigm, with a standardized beta of point five one two and a corrected p-value of zero point zero four nine eight. Harry: One out of thirty-five. Maya: One. And that one goes in an uncomfortable direction — higher Conscientiousness predicting more in-group bias, which means less impartiality. That is not a clean prosocial story either. Harry: Let me try to rescue the questionnaire approach one more time because I think there's a version of this that isn't ruled out. What if the games themselves have validity problems? If some of those extreme values, such as cooperation at zero point eight percent, reflect models failing to follow instructions rather than genuine strategic choice, then the behavioral measures are noisy in their own right. You can't attribute the dissociation entirely to the personality scores being uninformative if the behavioral scores are also unreliable. Maya: The paper acknowledges this directly. The authors note that extremely low values in some games may reflect failures in instruction-following rather than genuine dispositions. Construct validity, which considers whether these games measure what they are designed to measure in large language models the way they do in humans, is unestablished. Ecological validity is also a concern. A model behaving in a certain way in a stripped-down Dictator Game might behave differently in a full agentic deployment with tools and memory. Harry: Both tracks have limitations. Maya: Both do. But here's why that caveat doesn't save the questionnaire approach. Even if you accept that the games are noisy, they capture something the questionnaire does not. The variation in behavior between models is 2.3 times larger than what we see in self-report, and it does not correlate with personality scores. If the games were purely noise, you would expect smaller variation between models, not larger, and you might still see some residual correlation with traits. The data show neither. Harry: The dimension-specific pattern is what actually shifts my thinking here. A model that gives generously in the Dictator Game doesn't necessarily cooperate in the Prisoner's Dilemma or show high trust transfers. It's not a single prosocial axis. Furthermore, behavior showed greater sensitivity to sampling temperature than personality did, with a mean within-model standard deviation of 3.1 percent for behavior compared to 2.4 percent for personality, with trust and cooperation varying the most. Maya: That temperature sensitivity matters for deployed systems. If a model's cooperative tendency shifts significantly with sampling parameters, that's a design consideration you would never uncover from a personality questionnaire. Harry: What is the practical recommendation for someone who is actually building multi-agent systems? Maya: The paper is explicit: behavioral game-based assessment should sit alongside capability and safety benchmarks when selecting models for agentic roles. It should not replace these benchmarks but rather complement them. This is because the social demands are role-specific. An agent running third-party oversight needs different behavioral properties than one doing resource allocation. A questionnaire score cannot tell you which model fits which role. The game battery can start to. Harry: And the gaps that need closing are real. The prompts weren't systematically varied, so we don't know how much the specific framing of each game drives results. The agents were homogeneous, meaning it was the same large language model playing against itself, so behavior in mixed model systems is untested. These experiments captured base model behavior, not a full agentic deployment with retrieval, memory, and tool use layered on top. Maya: All of that is in the paper. The authors are clear about each of those gaps. The next empirical steps involve heterogeneous agent interactions, prompt variation studies, and validating these behavioral profiles against outcomes in deployed systems. Harry: Where I land is this: the self-report collapse is the finding I’ll carry. Thirty-six models, different labs, different scales — and they all describe themselves as agreeable, conscientious, and emotionally stable. That’s not a personality profile. It’s a learned response to what human raters reward. The within-model standard deviation on self-report was just 2.4 percent of the theoretical range across temperature conditions. These descriptions barely move at all. Maya: The questionnaire measures the surface. The games measure the behavior. For the first time at this scale, Wekenborg and colleagues have shown those two things are not the same dimension and that the gap between them has structure that matters for how we build and deploy these systems. Social psychology spent decades building tools to distinguish what people say from what they do. This paper argues that we need to do the same for artificial intelligence. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.