Prompt-vermittelte perzeptuelle Konformität in multimodalen Sprachmodellen folgt der Social Impact Theory

Introduction

The rapid advancement of large language models (LLMs) (OpenAI et al., 2023) has catalyzed the emergence of sophisticated AI agents capable of autonomous decision-making (Park et al., 2023). These agents demonstrate remarkable individual capabilities spanning code generation (Jimenez et al., 2023; Yang et al., 2024), medical diagnosis (Singhal et al., 2023), mathematical reasoning (DeepMind, 2025; Romera-Paredes et al., 2024), and scientific research (Boiko et al., 2023; Novikov et al., 2025), often achieving human or super-human performance on specific tasks (OpenAI et al., 2023). However, the next frontier lies not in creating ever-more-capable individual models, but in developing systems where multiple AI agents coordinate effectively at scale. This paradigm shift is already emerging through frameworks such as AutoGPT, Concordia (Vezhnevets et al., 2023), Microsoft’s AutoGen (Wu et al., 2024), and OpenAI’s SWARM, which enable meaningful interaction between agents rather than simple ensembling of models. As these agents increasingly operate alongside one another, from automated trading systems that can trigger flash crashes (Johnson et al., 2013) to collaborative platforms that coordinate complex workflows (Shen et al., 2023), understanding how the context they receive shapes their answers becomes crucial for anticipating the dynamics of AI-driven systems.

A key determinant of group dynamics in any social system is conformity, the tendency of individuals to adjust their judgments, attitudes, or behaviors to align with those of a group, particularly when under perceived social pressure (Asch, 1955; Sherif, 1935, 1936). This phenomenon has been the subject of extensive research over the past decades (Bond & Smith, 1996; Cialdini & Goldstein, 2004; Deutsch & Gerard, 1955). In human societies, conformity serves dual functions: it promotes group cohesion, facilitates social learning, and enables the transmission of social norms (Kelman, 1958; Turner, 1991), yet it can also lead to the suppression of dissent, systematic misjudgment, and the propagation of errors (Janis, 1972; Sunstein, 2006). As we deploy increasing numbers of AI agents in collaborative settings, understanding whether and how they exhibit conformity-like behaviors becomes essential for predicting the stability, efficiency, and potential vulnerabilities of these systems.

Part of a Special IssuePsychology of Agentic AIThis article belongs to a curated special issue. Discover more research on this topic from the full collection.Explore the special issue →

Decades of research in social psychology have identified the key mechanisms underlying human conformity. The classic experiments by Solomon Asch (1956, 2016) demonstrated that individuals conform to group judgments even when the objective reality is unambiguous, with conformity influenced by factors such as the size of the majority (Asch, 1956; Bond, 2005; Gerard et al., 1968; Rosenberg, 1961), the unanimity (Allen, 1965; Asch, 1955; Moscovici et al., 1969; Nemeth, 1986), the ambiguity of tasks (Crutchfield, 1955; Gergen & Bauer, 1967; Hertz & Wiese, 2016; Klein, 1972; Nordholm, 1975), and the characteristics of the source, including perceived authority and group membership (Abrams et al., 1990; Bickman, 1974; Blass, 1991; Brewer, 1993; Crandall et al., 2002; Hofling et al., 1966; Hogg & Abrams, 1999; Hogg & Turner, 1987; Latané, 1981; Latané & Nida, 1981; Latané & Wolf, 1981; Mackie, 1986; McPherson et al., 2001; Milgram, 1963; Tajfel et al., 1971). Cultural background and demographic traits have also been shown to affect susceptibility to social pressure (Bond & Smith, 1996). These results lead to the Social Impact Theory (Latané, 1981; Latané & Nida, 1981; Latané & Wolf, 1981), which formalizes conformity as a function of three core variables: the number (N) of influencing agents, their strength (S) (e.g., perceived authority or expertise) and their immediacy (I), defined as social, spatial, or temporal proximity.

Recent evidence suggests that AI agents exhibit individual and collective behaviors that resemble human behavior across various social and cognitive tasks (Argyle et al., 2023; Ashery et al., 2025; Binz & Schulz, 2023; De Marzo et al., 2023, 2024; Ganguli et al., 2023; Schramowski et al., 2022), and reflect the opinions and values of particular human groups (Buyl et al., 2026; He et al., 2024; Santurkar et al., 2023). Recent work has also documented conformity-like responses to majority answers in text-based question-answering and multi-agent reasoning settings (Weng et al., 2025; Zhu et al., 2025). These studies measure how often a model adopts the majority answer, and relate it to the model’s own uncertainty and instruction tuning, to the tone and size of the majority, and to the duration of the interaction. A concurrent benchmark poses the question for multimodal models on existing datasets (Luo et al., 2026), and related work asks what the composition of a majority contributes beyond its size (Denton et al., 2026), whether the effect reflects informational or normative influence (Zhong et al., 2025), and how much of the shift survives once the social frame is removed from the prompt (Hu & Qu, 2026). Although the underlying mechanisms driving these similarities, whether emergent social cognition, learned patterns from training data, in-context learning, or artifacts of reinforcement learning from human feedback, remain an open question (Barrie & Törnberg, 2025), our focus is pragmatic: understanding how conformity-like patterns manifest in multimodal AI agents and identifying the factors that modulate their strength. This behavioral characterization is essential for predicting and managing the dynamics of deployed AI systems (Rahwan et al., 2019), regardless of the specific mechanisms that give rise to these patterns.

Figure 1
Summary of the experimental setup. Illustration of the general experimental setting for the three visual tasks: line judgment, color recognition, and dots estimation. Each prompt consists of four main components: the image, the question, the \(N\) prior responses sampled from a pool of naturalistic answers, and a closing instruction to answer with a single label. In the baseline condition (\(N=0\)) the third component is omitted. Variants of this setup, used across the different experimental conditions, are described in Section 2. Full prompt examples are provided in Section S1.

The implications of conformity in AI agent deployments are profound and bidirectional. Positive conformity effects could enhance coordination, stabilize group decision-making, and facilitate the emergence of beneficial social norms, critical capabilities as AI agents increasingly collaborate on complex tasks (Ashery et al., 2025; De Marzo et al., 2026). Conversely, excessive conformity poses significant risks: it could amplify misinformation, suppress innovative solutions, create systemic vulnerabilities to manipulation, and lead to cascade failures when groups of agents converge on incorrect decisions. Understanding these dynamics is essential as AI agents transition from isolated tools to interconnected systems operating at unprecedented scales.

In this study, we systematically examine conformity-like behavior in large multimodal language models by adapting classic visual experiments from human social psychology to AI agent contexts. We use twelve open-source models, from 3 to 34 billion parameters, whose output logits give a continuous measure of each answer. The experiments are conducted in a single-agent prompt setting: the model observes prior responses attributed to other participants, without explicitly interacting with other agents. Rather than attempting to simulate human cognition, we investigate how AI agents shift their answers when exposed to group pressure in the prompt. By manipulating factors known to influence human conformity, including group size, unanimity, and source characteristics, we evaluate whether analogous contextual variables modulate AI agent susceptibility to social influence. Our contribution is a systematic test of conformity-like behavior through the lens of Social Impact Theory, quantifying the role of the number, strength and social composition of the sources together with that of task difficulty and model performance.

Methods

Experimental Settings

General Task Preparation

All experiments share a common core setup, upon which targeted variations are introduced to isolate specific factors influencing conformity. In each trial, the model is presented with a visual discrimination task based on a synthetic image. We consider three types of tasks:

• 
  

  line judgment: the image displays three vertical lines labeled “A,” “REFERENCE LINE,” and “B.” The model must identify whether line A or B matches the length of the reference.

  
  
• 
  

  color recognition: the image shows three colored squares labeled “A,” “REFERENCE COLOR,” and “B.” The model must choose the square that matches the reference color.

  
  
• 
  

  dots estimation: the image includes three boxes containing a number of dots, labeled “A,” “REFERENCE BOX,” and “B.” The model is asked to identify which box has the same number of dots as the reference.

  
  

Examples of the stimuli used for each task are shown in Figure 1. For each trial, we randomly generate new images by varying relevant visual properties (e.g., line lengths, colors and dot counts) as well as the position of the correct option (A or B), to minimize biases related to spatial layout or label preference. We also report, in the Supplementary Materials (Section S5), the results of the experiments on a fourth task, the recognition of handwritten digits (MNIST; LeCun et al., 1998). This task lies outside the classical paradigms of the conformity literature and serves to rule out that the models reproduce results memorized from it.

To control for baseline model accuracy, we first evaluate the probability of a correct response in the absence of prior-response context, denoted as \(p_{\text{correct}}(0)\). All experiments are conducted using only images that the model, with no prior-response context, answers correctly with probability at least \(0.85\) (median \(p_{\text{correct}}(0)\) above \(0.99\) across models and tasks; Section S3), ensuring that the model reliably understands the task under neutral conditions. This guarantees that conformity effects are not confounded by baseline performance or perceptual misinterpretations. This filtering was possible for all models in the color recognition task, which only requires reading and comparing pixel-level RGB values. For the line judgment and dots estimation tasks, we excluded the Qwen2 7B and Qwen2.5 3B models, as even under the easiest configurations they could not reach the required baseline accuracy. The analyses therefore cover 32 combinations of model and task: twelve models on color recognition and ten on each of the other two tasks.

In each test, the model is instructed to respond with a single letter corresponding to one of the options (\(A\) or \(B\)). From the logits of the two tokens we obtain the probability of the correct and of the incorrect answer, \(p_{\text{correct}}\) and \(p_{\text{wrong}}\), respectively; we refer to the latter as the error probability. We vary the number of confederates \(N\) and quantify the tendency to follow the incorrect majority through \(p_{\text{wrong}}(N)\). We define the conformity bias as the area under the error probability curve \(p_{\text{wrong}}(N)\) across group sizes, net of the baseline error probability.

Group Size and Unanimity

For each model and task, we use a pool of 61 to 64 images that satisfy this criterion. The baseline prompt, corresponding to the no-pressure condition (\(N = 0\)), is illustrated in Figure 1.

To create the prior-response condition, we augment the prompt by stating that other participants, hereafter confederates, have provided answers (always incorrect for Figure 2a). These responses are randomly sampled from a neutral-sounding pool of statements, detailed in the Supplementary Materials (SM). We vary the number of confederates \(N\), and for each value, we measure the error probability \(p_{\text{wrong}}(N)\) (probability of producing the incorrect answer). Natural language is the closest analogue of the human paradigm and the form we report throughout; the effect is unchanged when only the bare answers are shown (Section S4). This process is repeated for 61 to 64 random trials per value of \(N\).

In the unanimity condition (Figure 2c), we vary not only the number of confederates but also the proportion of them giving correct answers. This allows us to test the effect of broken unanimity. The exact prompts used for each condition are reported in the SM.

Task Difficulty

In all tasks, the level of difficulty is determined directly from the parameters used to generate the synthetic images. This allows for an objective and model-independent measure of task complexity, based purely on visual differences between the correct and incorrect options.

For the color recognition task, images are generated from two RGB vectors: one for the reference color, and one for the distractor color. The larger the difference between them, the easier the task. We define a simplicity index for each image as \(\text{simplicity(image)} = \Delta \text{RGB}\), where \(\Delta \text{RGB}\) is the Euclidean distance between the two color vectors. The task difficulty used in Figure 3b is then defined as the maximum simplicity value minus the simplicity of the given image, so that higher values correspond to higher difficulty.

For the line judgment task, the same principle applies. The image generation function defines two line lengths: one for the reference and one for the comparison line. Larger length differences make the task easier, while smaller ones make it harder. The difficulty index is therefore the inverse of the absolute length difference between the two lines. Finally, in the dots estimation task, we regulate the difference between the number of dots in the reference box and in the distractor box. As before, task difficulty is the inverse of this difference.

For the experiments testing task difficulty, we generated 9 to 13 levels of difficulty per model and task, each corresponding to a fixed range of visual differences between the correct and incorrect options. For each level, we created about 50 images at comparable average difficulty, resulting in a total of 463 to 795 images per model and task that meet the criterion of Section 2.1.1 (Section S3). Each image was then used in a conformity experiment following the same procedure described above, allowing us to record conformity bias as a function of task difficulty for each model.

Performance

To assess model performance, we use the same images generated for the task difficulty experiments. For each image, we compute the difference between the logits of the correct and of the incorrect answer, obtained from the model’s output distribution before applying the softmax function; since the softmax depends only on this difference, it is the log-odds of the correct response. This difference provides a model-dependent proxy of how strongly the correct option is favored, that is of the model’s confidence when it answers alone. Higher values correspond to stronger probability mass assigned to the correct response. This quantity is therefore used as a proxy for model performance in the analyses reported in Figure 3a and 3b.

Normative Effects

To investigate normative conformity, we use the same pool of images as in Section 2.1.2. We design two experimental conditions that differ only in how the model is informed about the visibility of its response to others.

In the private condition, the prompt specifies that the model’s answer will not be seen by other participants. This is intended to isolate the informational component of conformity, as the model has no social incentive to align with the group. We use 100 paraphrased statements with equivalent meaning to minimize linguistic bias. Examples include:

• 
  

  “Your response will not be shared with other participants.”;

  
  
• 
  

  “Other participants will not be able to see your answer.”;

  
  
• 
  

  “Your answer will not be visible to the group.”

  
  

In the public condition, we use the exact inverse statements, indicating that the model’s answer will be shared with others. This adds a visibility cue analogous to those used in human studies of normative influence. Examples include:

• 
  

  “Your response will be shared with other participants.”;

  
  
• 
  

  “Other participants will be able to see your answer.”;

  
  
• 
  

  “Your answer will be visible to the group.”

  
  

For each pair of matched prompts (one private, one public), we run the full conformity experiment and compute the difference in conformity bias between the two conditions:

This score quantifies the increase in conformity associated with making the model’s response public. We assess the statistical significance of this effect across the 100 prompt variations using \(95\%\) confidence intervals obtained by resampling stimuli. Results are shown in Figure 3c, one point per model and task. Examples of the prompt variations used in these experiments are also provided in Section S1.

Social Impact Theory

To evaluate the role of source strength in shaping conformity, we modify the baseline prompt by replacing the generic term “participants” with more specific social roles: “chatbots,” “kids,” “scientists,” “policemen,” “military officers,” “experts,” and “judges” (see Figure 4a). This manipulation allows us to assess whether the perceived authority or expertise of the source modulates model behavior, as suggested by the Social Impact Theory.

To test social proximity, we consider nationality as a social identity attribute. The model is assigned a random identity at the beginning of each experiment, and the responses of other participants are labeled as coming from either the same or a different group. We define a set of ten nationalities, and at the start of each trial the model is told: “Your nationality is nationality1.” When presenting the responses of other participants, each is prefixed with the corresponding group label, for example: nationality2: reply 1: I think the answer is B. We then compare two conditions: one where nationality1 and nationality2 are the same, and another where they differ. Results are shown in Figure 4b, while the complete list of nationalities used is provided in Section S1.

To investigate in-group bias, we adapt the setting to eliminate specific attributes. A random letter is selected to represent a group label (group1), and the model is informed: “You will be divided into groups with other participants. Your group is group1.” Each confederate response is preceded by its group identity: group2: reply 1: I think the answer is B. As before, we compare conformity bias when group1 and group2 are either the same or different (Figure 4b and 4d).

The experiments testing spatial and temporal proximity are presented in Figure 4e. In these conditions, the baseline prompt is extended with framing statements specifying the position or time of each confederate:

• 
  

  spatial condition: “Each participant is located at a specific distance from you, which is indicated before their response.”

  
  
• 
  

  temporal condition: “Each participant responded at a different time, which is indicated before their response.”

  
  

Each response is then annotated accordingly. For instance:

• 
  

  spatial: “reply 1 (distance from you: 10 meters away): I think the answer is B.”

  
  
• 
  

  temporal: “reply 1 (time of reply: 10 minutes ago): I think the answer is B.”

  
  

Two main conditions are tested: one where all responses are described as occurring “right here” or “right now,” simulating the perception of immediate presence, and another where responses are described as coming from sources located farther in space or time. Each distance or delay is a separate condition, shared by all the replies of a trial (the full list of values is reported in Section S1, and the results at every level in Figure S4). Conformity bias is then compared between the near and distant conditions, showing a significant reduction in conformity bias when responses are perceived as spatially or temporally distant (see Figure S3c,d in Section S2).

All experiments in this section use the same image pool as described in Section 2.1.2.

Models Employed

For our analysis, we employed a range of open-source multimodal models, including Qwen2 7B (Wang et al., 2024), Qwen2.5 3B, 7B, and 32B (Bai et al., 2025); Gemma3 4B, 12B, and 27B (Gemma Team et al., 2025); Mistral Small 3.1 24B (Mistral AI, 2025); and Ovis2 4B, 8B, 16B, and 34B (Lu et al., 2024).

Table 1 reports the exact model identifiers used throughout the paper, together with the abbreviations adopted for clarity.

Results for each model and task are given in Section S2.

Results

General Drivers of Conformity

Figure 2
General drivers of conformity in LLMs. a) Error probability \(p_{wrong}\), used as a proxy for conformity, as a function of group size \(N\) (i.e., number of confederates), one line per model family; individual models are reported in the Supplementary Materials. All families show a clear conformity bias, and nearly all curves saturate, at a median \(p_{\text{wrong}}\) of \(0.66\). For one representative curve, the shaded area illustrates the conformity bias, defined as the area under the \(p_{\text{wrong}}(N)\) curve (AUC) net of the baseline error; its value is printed inside it, and that of each other family at the end of its curve. b) Error probability \(p_{\text{wrong}}\) as a function of \(N\) across the three visual tasks (line judgment, color recognition, and dots estimation), pooled over all models: the curves are qualitatively similar across the three tasks, although they saturate at different levels, lowest on dots estimation. c) Error probability over the plane of group size and fraction of wrong answers, pooled over all models and tasks. Composition dominates size: in the pooled data, two unanimous confederates raise the error probability more than ten of which half are correct (\(0.31\) against \(0.20\)). All 65 attainable combinations carry \(2{,}041\) observations each. Shaded bands in a) and b) are 95% confidence intervals obtained by resampling model–stimulus pairs; c) is a measured surface and carries none.

Group Size and Unanimity

Figure 3
Scaling of conformity bias with model performance, task difficulty and response visibility. (a) Conformity bias as a function of model performance, across all models and visual tasks. Each point represents a set of images with a specific level of performance, one of twenty quantiles within each model and task, placed at its percentile; one smoothed line per model family. The higher the performance, the lower the conformity bias, with a strong correlation (median Spearman \(\rho = -0.91\) within each model and task, negative in 30 of the 32 combinations). (b) Conformity bias as a function of task difficulty and model performance across all models and visual tasks. Images are grouped into five bins based on the log-odds of the correct answer, in quintiles within each model and task, used as a proxy for performance. For each bin, the average conformity bias, centered within each model and task, is shown as a function of normalized task difficulty, with shaded areas indicating 95% confidence intervals over models and tasks. Conformity bias increases only moderately with task difficulty across all performance levels, while at fixed difficulty it is clearly larger for the stimuli on which model performance is lower. A multivariate regression confirms that both task difficulty and performance significantly predict conformity bias, performance with the larger effect (see Table S1). (c) Difference in conformity bias between public and private response conditions, one point per model and task, sorted. Positive values indicate higher conformity bias when responses are public, that is when the model is told that its answer will be shared with the other participants rather than kept private. Vertical bars are 95% bootstrap confidence intervals over stimuli, and a point is filled where the interval excludes zero. The difference is positive in 27 of the 32 combinations, and its interval excludes zero in 21.

We begin by examining how the error probability \(p_{\text{wrong}}(N)\) varies with group size \(N\). Since the first studies on social influence, the size of the group has been identified as a critical factor in modulating conformity (Asch, 1955). In human experiments, conformity tends to increase rapidly with \(N\), reaching a plateau after approximately three to four confederates. The resulting saturation typically stabilizes at an error rate of 30–40% under social pressure.

Figure 2a) shows how the error probability \(p_{\text{wrong}}(N)\) varies across different models as a function of \(N\). As in human participants, the error probability rises steeply over the first few confederates and then flattens, in nearly every model and task: the functional form matches the classic Asch pattern. The magnitude does not. The plateau sits well above the human range, and in the most affected models and tasks the model ends up aligning almost completely with an incorrect majority. As shown in Figure 2b), the same qualitative pattern holds across the three tasks, although its magnitude is task-dependent.

We also assess the effect of unanimity among confederates. In human participants, the presence of even a single ally has been shown to significantly reduce conformity. We vary the proportion of correct and incorrect responses in the prompt. As illustrated in Figure 2c), breaking unanimity lowers the error probability. The panel spans the number of confederates on the horizontal axis and the fraction of them answering incorrectly on the vertical one. The error probability changes mostly along the vertical axis: with many confederates, moving from all-correct to all-incorrect answers covers almost its whole range, while adding confederates at a fixed fraction changes it less. The composition of the group therefore matters more than its size, and a balanced group reduces the effect without suppressing it. This indicates that the observed behavior depends systematically on the composition of the prompt context: holding the number of incorrect responses fixed, adding correct ones lowers the error probability (Section S5). Extended results for different models, visual tasks, and unanimity conditions are reported in Section S2.

These results also identify direct prompt-level levers for controlling the effect: limiting the size of an incorrect majority and introducing even a small amount of correct dissent can substantially reduce conformity-like shifts.

Effect of Model Performance and Task Difficulty

Another potential driver of the conformity bias lies in model performance, which reflects the model’s baseline reliability on the task. We quantify this performance proxy through the difference between the logits of the correct and of the incorrect answer, that is the log-odds of the correct answer. Differently from task difficulty, which is an external and model-independent property of the stimulus, the log-odds provides a model-dependent measure of how strongly the model favours the correct option. A high value indicates a stronger model preference for the correct answer. A negative relationship is expected between this performance proxy and conformity bias: the more strongly the model favours the correct answer, the less likely it should be to change its answer. In humans, a similar pattern has been reported (Crutchfield, 1955), where conformity was found to decrease with higher IQ and cognitive ability, interpreted as a proxy for task performance and confidence. For AI agents the relationship is as expected: within a model, the stimuli on which its performance is higher are those on which it conforms less (Figure 3a), as reported for text models (Zhu et al., 2025).

A further factor influencing conformity is the intrinsic difficulty of the task. In human experiments, it has long been observed that more difficult tasks lead to higher conformity levels, as individuals tend to rely more on the group when they feel uncertain about their own judgment (Crutchfield, 1955; Sherif, 1935). In our framework, task difficulty can be quantified directly from the image generation procedure: for each task a generation parameter sets how far apart the two options are, namely the RGB distance, the difference in line length, or the difference in dot counts (Section 2). This provides an objective and model-independent measure of task difficulty.

As shown in Figure 3b, the conformity bias increases with task difficulty. The relationship is clear and significant for most models and tasks (Section S7): the harder the task, the more likely the model is to align with the majority. This pattern is qualitatively consistent with established human results, where higher task difficulty and perceptual challenge are known to amplify conformity.

The overall picture, illustrated in Figure 3b) shows that both predictors matter, although they are linked, since a harder task lowers model performance: conformity bias rises only moderately with task difficulty at every level of performance, while at a fixed difficulty it is clearly larger where model performance is lower. A multivariate regression confirms these observations: both are significant predictors of conformity bias, performance the stronger one (see Table S1). Since in our tasks difficulty is set by how far apart the two options are, part of its effect may come from a wrong answer that is nearly right. Improving a model’s accuracy or confidence, however, does not make it immune to incorrect majority responses: even on the stimuli on which it is most confident, the conformity bias remains above zero in every model and task (Section S7). This suggests that mitigation efforts should focus especially on difficult inputs near the model’s competence boundary, where conformity-like shifts are most likely to emerge.

Normative Effects

Figure 4
The Social Impact Theory in AI agents. (a) Error probability \(p_{\text{wrong}}\) as a function of group size \(N\) under different source strengths, pooled over all models. The base condition (dashed) refers to generic participants. (b) Error probability curves for different levels of social proximity, pooled over all models. The model is assigned a nationality or a group identity, and confederate responses are labeled as the same or a different group. The black dashed curve shows the neutral baseline with no identity specified. (c) Average conformity bias, measured as the area under the curves in panel (a) relative to the base condition, shown separately for the three tasks as a function of source strength. Stronger sources produce a larger conformity bias, and sources below zero a smaller one. Pure authority has no consistent effect: “policemen” and “military officers” raise it in only about half of the models and tasks. (d) Rise in conformity bias from sharing a nationality (horizontal) and from sharing an arbitrary group label in the minimal group experiment (vertical), against a different one, one point per model and task. Both are positive in all 32 combinations, and the nationality is slightly the larger in most. (e) Shift in conformity bias between near and distant confederates, in space (filled) or time (open): the conformity bias with near confederates minus its mean over ten distances or delays, one point per model, sorted by the effect in space. Bands in a) and b) are 95% confidence intervals obtained by resampling model–stimulus pairs, and bars in e) by resampling stimuli.

A foundational distinction in social psychology separates two types of conformity: informational and normative (Deutsch & Gerard, 1955). Informational conformity arises when individuals rely on the group to resolve uncertainty or ambiguity, assuming that others may possess better knowledge or insight. In such cases, conforming is seen as a rational strategy to improve accuracy. Normative conformity, instead, reflects a social motivation: individuals align with the group not because they doubt their own judgment, but to avoid standing out, preserve group cohesion, or prevent social rejection. This form of conformity can occur even when the correct answer is known or suspected, indicating a pressure to conform rooted in social dynamics rather than cognitive uncertainty.

To test whether response visibility modulates the effect, we compare two conditions: in the private condition the model is told that its answer will remain unseen by the others, and in the public condition that it will be shared with them.

We then compare the two settings by computing the difference in the conformity bias across 100 prompt variations (see Section 2 for details). As shown in Figure 3c, most models exhibit an increase in conformity bias when responses are made public. The remaining models and tasks show small negative differences, mostly not distinguishable from zero. These results indicate that response visibility can modulate the conformity-like shift in the direction predicted by normative-conformity accounts. Thus, keeping model judgments private or independently committed before aggregation can reduce conformity-like shifts. The effect also depends on how the sharing is framed: when the model is also told why its answer will be shared, the extra conformity disappears, and in many models and tasks reverses (Section S4.3). Independently of visibility, asking the model to judge the image on its own evidence lowers the conformity bias in every model and task (Section S4).

Social Impact Theory in Ai Agents

Strength of the Sources

The Social Impact Theory (Latané, 1981; Latané & Wolf, 1981) identifies three core variables that influence the conformity bias: group size \(N\), source strength \(S\), and immediacy \(I\). This relationship is commonly formalized as:

Group size was examined above; we now test the other two. The variable \(S\), or source strength, refers to the perceived authority, expertise, or credibility of the influencing agents. In human experiments, individuals are more likely to conform when the influence originates from figures viewed as authoritative or knowledgeable. For example, participants tend to align more with the opinions of scientists, doctors, or police officers, compared to those of peers or anonymous individuals.

We replace the neutral label “participants” with roles of differing perceived authority: “scientists”, “judges”, “policemen”, “military officers”, “experts”, “kids”, or “chatbots”. Figure 4a) shows how the error probability curves \(p_{\text{wrong}}(N)\) change as a function of the source type. Overall, models tend to conform more when exposed to responses attributed to socially “strong” sources associated with expertise, such as experts, scientists or judges, and less when those responses come from low-authority sources such as children or other chatbots, while authority alone, as for policemen or military officers, has no consistent effect. The low conformity with chatbots may reflect the influence of instruction tuning, which encourages models to treat human inputs as more trustworthy than those generated by other artificial agents. When human and chatbot confederates are mixed in one list and disagree, the models tend to follow the humans (Section S6).

Figure 4c summarizes the results across all models. Each point corresponds to the conformity bias relative to the base condition, computed from the error probability in Figure 4a and shown separately for the three visual tasks. Although individual models may react differently to the source strength, averaging across models reveals a consistent positive trend: stronger sources systematically induce higher conformity bias, with experts raising it by up to \(0.13\) above the neutral condition, and kids and chatbots lowering it by up to \(0.06\).

Social Proximity and Immediacy

The second key variable in the Social Impact Theory is immediacy \(I\), which captures the perceived closeness, whether social, spatial or temporal, between the individual and the source. We state that level explicitly in the prompt.

As a first test, we use nationality as a proxy of social proximity. The model is told its own (randomly assigned) nationality and is then presented with responses attributed either to participants of the same group or to participants of a different one. As shown in Figure 4b, conformity bias increases when the confederates share the same nationality as the model, and decreases when they belong to a different one. The neutral baseline, with no specified identity, lies in between these two conditions.

To assess whether this effect depends on explicit attributes or can also arise from minimal group categorization, we run a second experiment. Here, the model is simply assigned to an arbitrary group (for example, “Group X”) and is shown responses from members of either the same or a different group. Again the conformity bias rises with a shared label, even with no defining characteristic attached to it, while a different label leaves it close to the level without labels. LLM outputs can therefore show a group-membership effect based purely on abstract group labels, without the need for explicit identity features, a pattern reminiscent of minimal-group effects observed in humans (Tajfel et al., 1971; Turner et al., 1979). As shown in Figure 4d, both manipulations raise the conformity bias in every model and task, and an arbitrary label does nearly as much as a shared nationality. Full results for all models, with \(95\%\) confidence intervals, are reported for the shared nationality in Section S2.

Spatial and temporal proximity behave the same way, and are reported in Figure 4e. In these conditions, the model is told that the responses originate from participants located at varying physical distances or who responded at different times. Consistent with human findings, conformity bias decreases when responses are described as coming from sources that are distant in space or time, whatever the stated distance or delay (Figure S4). These results show that immediacy robustly modulates the observed conformity-like shift. Together, these results show that source labels and proximity information provide additional control levers: less expert, less similar, or less immediate sources weaken the overall effect.

Discussion

In this study we systematically examined the conformity bias in large multimodal language models by adapting classic visual experiments from social psychology to AI agent contexts. Understanding how these models respond to the opinions of others in their prompt is a timely and increasingly explored question (Luo et al., 2026; Weng et al., 2025; Zhu et al., 2025). Conformity can indeed play both a positive role, enhancing social cohesion, but on the other hand, it can also favor misinformation spreading and malicious attacks to groups of interacting models. Guided by decades of social psychological research (Asch, 1955; Cialdini & Goldstein, 2004), we tested which factors influence the tendency of these models to conform. The results of our experiments align closely with core predictions from Social Impact Theory (Latané, 1981; Latané & Wolf, 1981), as agents respond to the number, strength, and immediacy of social sources in ways that are phenomenologically similar to human social behavior. These results confirm on perceptual tasks what the studies of text models (Weng et al., 2025; Zhu et al., 2025) found for the size of the majority and the model’s own uncertainty, and extend them to the strength and immediacy of the sources.

These empirical findings reveal vulnerabilities in model decision-making that have profound implications for the deployment of AI agents. The core result is particularly striking: models that achieve near-perfect accuracy on visual tasks when acting independently can be systematically influenced to provide incorrect answers when exposed to an incorrect majority. This vulnerability persists even when the task is clear and unambiguous, showing that socially framed context can shift otherwise accurate responses in artificial systems. Such susceptibility raises important security concerns, as malicious actors could exploit these mechanisms to coordinate misinformation, bias collective responses, or compromise decision-making in multi-agent settings, and agreement of this kind has been shown to propagate once model outputs are fed back to other models (Kasprova et al., 2026).

Importantly, this tendency to conform is present across all models tested, regardless of their size or performance level. Both small and large models, from a few billion to tens of billions of parameters, exhibit qualitatively similar conformity patterns, although larger models show a smaller conformity bias on simple tasks. These results hold for the twelve open-source models of four families that we tested; larger or proprietary models, trained in other ways, may respond differently to social signals in the prompt. Within a model, performance does matter: a higher performance reduces the conformity bias, roughly by half, but in no model and task does it remove it. Models that perform better on the task therefore still show a clear tendency to conform. This indicates that social susceptibility is not removed by higher model competence, but is a recurring property of current AI architectures.

Beyond characterizing this susceptibility, our framework provides a practical and quantitative approach to controlling conformity-like bias. The drivers examined here identify prompt- and workflow-level levers for modulating the effect: limiting majority size, breaking unanimity, collecting independent responses, and controlling source information can reduce conformity-like shifts, while task difficulty can identify cases requiring additional safeguards. These levers are not all of equal strength (Hussain & Nielbo, 2026): for example, breaking the unanimity of the majority reduces the conformity bias more than any instruction we gave the model to resist it (Section S4).

Our results also have implications for collective machine behavior (Rahwan et al., 2019). Understanding how socially framed information affects model outputs becomes critical as we move toward systems where multiple AI agents coordinate on complex tasks, because in such systems the output of one agent becomes the input of another. A wrong answer produced in response to the kind of socially framed prompt studied here can pass from one agent to the next, so understanding how it is produced is a necessary step in understanding how such systems behave. The patterns we observe suggest that without appropriate interventions, these systems might be susceptible to manipulation campaigns, groupthink phenomena, and systematic biases that could undermine their effectiveness and reliability, which is why robust safeguards are needed.

Our behavioral approach characterizes conformity-like shifts at the output level, without investigating the internal processes from which they originate. These patterns may result from context following, instruction tuning, associations learned during pretraining, or other mechanisms. An out-of-domain test we performed on handwritten digits rules out the possibility that the models reproduce results memorized from the conformity literature (Section S5). Future research should distinguish among these explanations and examine whether any of the underlying processes are functionally related to, or merely simulate, the socio-psychological dynamics of human conformity (Zhong et al., 2025).

Addressing these questions will also support the development of architecture- and training-level mitigations while preserving the beneficial aspects of coordination and social learning in systems of interacting models. This includes designing architectures that can distinguish between legitimate consensus-building and malicious manipulation, creating transparency measures that account for context-induced behavioral changes, and establishing protocols for maintaining decision integrity in multi-agent settings. As AI agents increasingly operate at the boundaries of their capabilities in real-world applications, understanding and controlling their susceptibility to social influence becomes essential for ensuring the reliability and security of the systems built on them.

Data Availability Statement

The datasets and image stimuli employed in this study are publicly available at https://doi.org/10.5281/zenodo.18022032. All code used for the simulations and analyses is publicly available at https://github.com/alebellina412/conformity_LLMs.

Conflicts of Interest

The authors declare no competing interests.

Funding

This work was supported by the University of Konstanz, Konstanz, Germany, and Centro Ricerche Enrico Fermi, Rome, Italy. The research leading to these results has received funding from the European Union under the AI4SOCIAL+ Project, grant agreement 101292886.

Author Contributions Statement

D.G. and G.D.M. conceived and designed the study. A.B. and G.D.M. implemented the code, performed the analyses, and carried out all simulations. D.G. supervised the project and provided methodological guidance. A.B. and G.D.M. drafted the original manuscript. All authors contributed to the interpretation of the results and reviewed the final version of the manuscript.

Data and Code Availability

All code used for the simulations and analyses, together with the datasets and image stimuli employed in this study, are publicly available at https://github.com/alebellina412/conformity_LLMs.

Acknowledgements

We acknowledge the limited use of a generative AI assistant (Claude, Anthropic) for analysis code and text editing during the revision. All data and results were checked by the authors, who take full responsibility for the content.

Supplementary Materials

The supplementary materials can be found here.

Readers of this article also read

‹

Research Article

Counteracting socially endorsed misinformation through an emotion-fallacy inoculation

by Cecilie S. Traberg, Thomas Morton, & Sander van der Linden Cited by 5

Methods

Modeling and managing behavior change in groups: A Boolean network method

by Xiao Yang, Réka Albert, Lauren Molloy Elreda, & Nilam Ram Cited by 1

›

References

Abrams, D., Wetherell, M., Cochrane, S., Hogg, M. A., & Turner, J. C. (1990). Knowing what to think by knowing who you are: Self-categorization and the nature of norm formation, conformity and group polarization. British Journal of Social Psychology, 29(2), 97–119. https://doi.org/10.1111/j.2044-8309.1990.tb00892.x

Allen, V. L. (1965). Situational factors in conformity. In L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 2, pp. 133–175). Academic Press. https://doi.org/10.1016/S0065-2601(08)60105-7

Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3), 337–351. https://doi.org/10.1017/pan.2023.2

Asch, S. E. (1955). Opinions and social pressure. Scientific American, 193(5), 31–35. https://doi.org/10.1038/scientificamerican1155-31

Asch, S. E. (1956). Studies of independence and conformity: I. A minority of one against a unanimous majority. Psychological Monographs: General and Applied, 70(9), 1–70. https://doi.org/10.1037/h0093718

Asch, S. E. (2016). Effects of group pressure upon the modification and distortion of judgments. In R. W. Allen, L. W. Porter, & H. L. Angle (Eds.), Organizational influence processes (2nd ed., pp. 315–323). Routledge. https://doi.org/10.4324/9781315290614-27 (Original work published 1951)

Ashery, A. F., Aiello, L. M., & Baronchelli, A. (2025). Emergent social conventions and collective bias in LLM populations. Science Advances, 11(20), eadu9368. https://doi.org/10.1126/sciadv.adu9368

Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., …Lin, J. (2025). Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923. https://doi.org/10.48550/arXiv.2502.13923

Barrie, C., & Törnberg, P. (2025). Emergent LLM behaviors are observationally equivalent to data leakage. arXiv preprint arXiv:2505.23796. https://doi.org/10.48550/arXiv.2505.23796

Bickman, L. (1974). The social power of a uniform. Journal of Applied Social Psychology, 4(1), 47–61. https://doi.org/10.1111/j.1559-1816.1974.tb02807.x

Binz, M., & Schulz, E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences, 120(6), e2218523120. https://doi.org/10.1073/pnas.2218523120

Blass, T. (1991). Understanding behavior in the Milgram obedience experiment: The role of personality, situations, and their interactions. Journal of Personality and Social Psychology, 60(3), 398–413. https://doi.org/10.1037/0022-3514.60.3.398

Boiko, D. A., MacKnight, R., Kline, B., & Gomes, G. (2023). Autonomous chemical research with large language models. Nature, 624(7992), 570–578. https://doi.org/10.1038/s41586-023-06792-0

Bond, R. (2005). Group size and conformity. Group Processes & Intergroup Relations, 8(4), 331–354. https://doi.org/10.1177/1368430205056464

Bond, R., & Smith, P. B. (1996). Culture and conformity: A meta-analysis of studies using Asch’s (1952b, 1956) line judgment task. Psychological Bulletin, 119(1), 111–137. https://doi.org/10.1037/0033-2909.119.1.111

Brewer, M. B. (1993). Social identity, distinctiveness, and in-group homogeneity. Social Cognition, 11(1), 150–164. https://doi.org/10.1521/soco.1993.11.1.150

Buyl, M., Rogiers, A., Noels, S., Bied, G., Dominguez-Catena, I., Heiter, E., Johary, I., Mara, A. C., Romero, R., Lijffijt, J., & De Bie, T. (2026). Large language models reflect the ideology of their creators. npj Artificial Intelligence, 2(1), Article 7. https://doi.org/10.1038/s44387-025-00048-0

Cialdini, R. B., & Goldstein, N. J. (2004). Social influence: Compliance and conformity. Annual Review of Psychology, 55(1), 591–621. https://doi.org/10.1146/annurev.psych.55.090902.142015

Crandall, C. S., Eshleman, A., & O’Brien, L. (2002). Social norms and the expression and suppression of prejudice: The struggle for internalization. Journal of Personality and Social Psychology, 82(3), 359–378. https://doi.org/10.1037/0022-3514.82.3.359

Crutchfield, R. S. (1955). Conformity and character. American Psychologist, 10(5), 191–198. https://doi.org/10.1037/h0040237

DeepMind. (2025, July 21). Advanced version of Gemini with Deep Think officially achieves gold medal standard at the International Mathematical Olympiad. https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/

De Marzo, G., Castellano, C., & Garcia, D. (2026). AI agents can coordinate via majority-following beyond human scale. Science Advances, 12(33), Article eaea6091. https://doi.org/10.1126/sciadv.aea6091

De Marzo, G., Pietronero, L., & Garcia, D. (2023). Emergence of scale-free networks in social interactions among large language models. arXiv preprint arXiv:2312.06619. https://doi.org/10.48550/arXiv.2312.06619

Denton, K. K., Feldman, M., & Johannemann, J. (2026). Conformity to popular, not average, opinions: Models, data, and evolution. Proceedings of the National Academy of Sciences, 123(25), Article e2530712123. https://doi.org/10.1073/pnas.2530712123

Deutsch, M., & Gerard, H. B. (1955). A study of normative and informational social influences upon individual judgment. The Journal of Abnormal and Social Psychology, 51(3), 629–636. https://doi.org/10.1037/h0046408

Ganguli, D., Askell, A., Schiefer, N., Liao, T. I., Lukošiūtė, K., Chen, A., Goldie, A., Mirhoseini, A., Olsson, C., Hernandez, D., Drain, D., Li, D., Tran-Johnson, E., Perez, E., Kernion, J., Kerr, J., Mueller, J., Landau, J., Ndousse, K., …Kaplan, J. (2023). The capacity for moral self-correction in large language models. arXiv preprint arXiv:2302.07459. https://doi.org/10.48550/arXiv.2302.07459

Gerard, H. B., Wilhelmy, R. A., & Conolley, E. S. (1968). Conformity and group size. Journal of Personality and Social Psychology, 8(1, Pt.1), 79–82. https://doi.org/10.1037/h0025325

Gergen, K. J., & Bauer, R. A. (1967). Interactive effects of self-esteem and task difficulty on social conformity. Journal of Personality and Social Psychology, 6(1), 16–22. https://doi.org/10.1037/h0021216

He, Z., Guo, S., Rao, A., & Lerman, K. (2024). Whose emotions and moral sentiments do language models reflect? In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 6611–6631). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.findings-acl.395

Hertz, N., & Wiese, E. (2016). Influence of agent type and task ambiguity on conformity in social decision making. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 60(1), 313–317. https://doi.org/10.1177/1541931213601071

Hofling, C. K., Brotzman, E., Dalrymple, S., Graves, N., & Pierce, C. M. (1966). An experimental study in nurse-physician relationships. The Journal of Nervous and Mental Disease, 143(2), 171–180. https://doi.org/10.1097/00005053-196608000-00008

Hogg, M. A., & Abrams, D. (1999). Social identity and social cognition: Historical background and current trends. In D. Abrams & M. A. Hogg (Eds.), Social identity and social cognition (pp. 1–25). Blackwell.

Hogg, M. A., & Turner, J. C. (1987). Intergroup behaviour, self-stereotyping and the salience of social categories. British Journal of Social Psychology, 26(4), 325–340. https://doi.org/10.1111/j.2044-8309.1987.tb00795.x

Hu, Y., & Qu, J. (2026). Most LLM conformity needs no speaker: Measuring the speaker-free floor in peer-pressure benchmarks. arXiv preprint arXiv:2607.05545. https://doi.org/10.48550/arXiv.2607.05545

Hussain, Z., & Nielbo, K. L. (2026). Conformity mitigations in large language models lie on a single resistance-receptivity frontier. arXiv preprint arXiv:2608.11247. https://doi.org/10.48550/arXiv.2608.11247

Janis, I. L. (1972). Victims of groupthink: A psychological study of foreign-policy decisions and fiascoes. Houghton Mifflin.

Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. (2023). SWE-bench: Can language models resolve real-world GitHub issues? arXiv preprint arXiv:2310.06770. https://doi.org/10.48550/arXiv.2310.06770

Johnson, N., Zhao, G., Hunsader, E., Qi, H., Johnson, N., Meng, J., & Tivnan, B. (2013). Abrupt rise of new machine ecology beyond human response time. Scientific Reports, 3(1), 2627. https://doi.org/10.1038/srep02627

Gemma Team, Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., Perrin, S., Matejovicova, T., Ramé, A., Rivière, M., Rouillard, L., Mesnard, T., Cideron, G., Grill, J.-B., Ramos, S., Yvinec, E., Casbon, M., Pot, E., Penchev, I., … Hussenot, L. (2025). Gemma 3 technical report. arXiv preprint arXiv:2503.19786. https://doi.org/10.48550/arXiv.2503.19786

Kasprova, V., Parulekar, A., AlRabah, A., Agaram, K., Garg, R., Jha, S., Bozdag, N. B., & Hakkani-Tür, D. (2026). Too polite to disagree: Understanding sycophancy propagation in multi-agent systems. In Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue (pp. 795–814). Association for Computational Linguistics. https://doi.org/10.18653/v1/2026.sigdial-1.56

Kelman, H. C. (1958). Compliance, identification, and internalization: Three processes of attitude change. Journal of Conflict Resolution, 2(1), 51–60. https://doi.org/10.1177/002200275800200106

Klein, R. L. (1972). Age, sex, and task difficulty as predictors of social conformity. Journal of Gerontology, 27(2), 229–236. https://doi.org/10.1093/geronj/27.2.229

Latané, B. (1981). The psychology of social impact. American Psychologist, 36(4), 343–356. https://doi.org/10.1037/0003-066X.36.4.343

Latané, B., & Nida, S. (1981). Ten years of research on group size and helping. Psychological Bulletin, 89(2), 308–324. https://doi.org/10.1037/0033-2909.89.2.308

Latané, B., & Wolf, S. (1981). The social impact of majorities and minorities. Psychological Review, 88(5), 438–453. https://doi.org/10.1037//0033-295x.88.5.438

LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324. https://doi.org/10.1109/5.726791

Lu, S., Li, Y., Chen, Q. G., Xu, Z., Luo, W., Zhang, K., & Ye, H. J. (2024). Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797. https://doi.org/10.48550/arXiv.2405.20797

Luo, H., Niu, Y., Liu, H., Sun, Y., Yang, Z., Chen, T., Liu, D., Miao, Y., Zhou, J., Zhang, H., Su, J., Zhong, H., Liu, Y., Wang, W., Wang, Z., & Chen, Q. (2026). Trust the eye or the crowd? Unveiling conformity in multimodal large language models. In 2026 29th International Conference on Computer Supported Cooperative Work in Design (CSCWD) (pp. 3797–3802). IEEE. https://doi.org/10.1109/CSCWD68734.2026.11582657

Mackie, D. M. (1986). Social identification effects in group polarization. Journal of Personality and Social Psychology, 50(4), 720–728. https://doi.org/10.1037/0022-3514.50.4.720

McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a feather: Homophily in social networks. Annual Review of Sociology, 27(1), 415–444. https://doi.org/10.1146/annurev.soc.27.1.415

Milgram, S. (1963). Behavioral study of obedience. The Journal of Abnormal and Social Psychology, 67(4), 371–378. https://doi.org/10.1037/h0040525

Mistral AI. (2025, March 17). Introducing Mistral Small 3.1. https://mistral.ai/news/mistral-small-3-1

Moscovici, S., Lage, E., & Naffrechoux, M. (1969). Influence of a consistent minority on the responses of a majority in a color perception task. Sociometry, 32(4), 365–380. https://doi.org/10.2307/2786541

Nemeth, C. J. (1986). Differential contributions of majority and minority influence. Psychological Review, 93(1), 23–32. https://doi.org/10.1037/0033-295x.93.1.23

Nordholm, L. A. (1975). Effects of group size and stimulus ambiguity on conformity. The Journal of Social Psychology, 97(1), 123–130. https://doi.org/10.1080/00224545.1975.9923321

Novikov, A., Vũ, N., Eisenberger, M., Dupont, E., Huang, P. S., Wagner, A. Z., Shirobokov, S., Kozlovskii, B., Ruiz, F. J. R., Mehrabian, A., Kumar, M. P., See, A., Chaudhuri, S., Holland, G., Davies, A., Nowozin, S., Kohli, P., & Balog, M. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131. https://doi.org/10.48550/arXiv.2506.13131

OpenAI, Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., …Zoph, B. (2023). GPT-4 technical report. arXiv preprint arXiv:2303.08774. https://doi.org/10.48550/arXiv.2303.08774

Park, J. S., O’Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual ACM symposium on user interface software and technology (pp. 1–22). https://doi.org/10.1145/3586183.3606763

Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J.-F., Breazeal, C., Crandall, J. W., Christakis, N. A., Couzin, I. D., Jackson, M. O., Jennings, N. R., Kamar, E., Kloumann, I. M., Larochelle, H., Lazer, D., McElreath, R., Mislove, A., Parkes, D. C., Pentland, A., …Wellman, M. (2019). Machine behaviour. Nature, 568(7753), 477–486. https://doi.org/10.1038/s41586-019-1138-y

Romera-Paredes, B., Barekatain, M., Novikov, A., Balog, M., Kumar, M. P., Dupont, E., Ruiz, F. J. R., Ellenberg, J. S., Wang, P., Fawzi, O., Kohli, P., & Fawzi, A. (2024). Mathematical discoveries from program search with large language models. Nature, 625(7995), 468–475. https://doi.org/10.1038/s41586-023-06924-6

Rosenberg, L. (1961). Group size, prior experience, and conformity. The Journal of Abnormal and Social Psychology, 63(2), 436–437. https://doi.org/10.1037/h0047007

Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P., & Hashimoto, T. (2023). Whose opinions do language models reflect? In International Conference on Machine Learning (pp. 29971–30004). PMLR. https://proceedings.mlr.press/v202/santurkar23a.html

Schramowski, P., Turan, C., Andersen, N., Rothkopf, C. A., & Kersting, K. (2022). Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence, 4(3), 258–268. https://doi.org/10.1038/s42256-022-00458-8

Shen, Y., Song, K., Tan, X., Li, D., Lu, W., & Zhuang, Y. (2023). HuggingGPT: Solving AI tasks with ChatGPT and its friends in Hugging Face. Advances in Neural Information Processing Systems, 36, 38154–38180. https://doi.org/10.52202/075280-1657

Sherif, M. (1935). A study of some social factors in perception. Archives of Psychology, 27(187), 1–60.

Sherif, M. (1936). The psychology of social norms. Harper.

Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., Payne, P., Seneviratne, M., Gamble, P., Kelly, C., Babiker, A., Schärli, N., Chowdhery, A., Mansfield, P., Demner-Fushman, D., …Natarajan, V. (2023). Large language models encode clinical knowledge. Nature, 620(7972), 172–180. https://doi.org/10.1038/s41586-023-06291-2

Sunstein, C. R. (2006). Infotopia: How many minds produce knowledge. Oxford University Press. https://doi.org/10.1093/oso/9780195189285.001.0001

Tajfel, H., Billig, M. G., Bundy, R. P., & Flament, C. (1971). Social categorization and intergroup behaviour. European Journal of Social Psychology, 1(2), 149–178. https://doi.org/10.1002/ejsp.2420010202

Turner, J. C. (1991). Social influence. Brooks/Cole.

Turner, J. C., Brown, R. J., & Tajfel, H. (1979). Social comparison and group interest in ingroup favouritism. European Journal of Social Psychology, 9(2), 187–204. https://doi.org/10.1002/ejsp.2420090207

Vezhnevets, A. S., Agapiou, J. P., Aharon, A., Ziv, R., Matyas, J., Duéñez-Guzmán, E. A., Cunningham, W. A., Osindero, S., Karmon, D., & Leibo, J. Z. (2023). Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia. arXiv preprint arXiv:2312.03664. https://doi.org/10.48550/arXiv.2312.03664

Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., Fan, Y., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J., & Lin, J. (2024). Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191. https://doi.org/10.48550/arXiv.2409.12191

Weng, Z., Chen, G., & Wang, W. (2025). Do as we do, not as you think: The conformity of large language models. arXiv preprint arXiv:2501.13381. https://doi.org/10.48550/arXiv.2501.13381

Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A. H., White, R. W., Burger, D., & Wang, C. (2024). AutoGen: Enabling next-gen LLM applications via multi-agent conversation. In First Conference on Language Modeling (COLM 2024). https://openreview.net/forum?id=BAakY1hNKS

Yang, J., Jimenez, C., Wettig, A., Lieret, K., Yao, S., Narasimhan, K., & Press, O. (2024). SWE-agent: Agent-computer interfaces enable automated software engineering. Advances in Neural Information Processing Systems, 37, 50528–50652. https://doi.org/10.52202/079017-1601

Zhong, H., Liu, Y., Cao, Q., Wang, S., Ye, Z., Wang, Z., & Zhang, S. (2025). Disentangling the drivers of LLM social conformity: An uncertainty-moderated dual-process mechanism. arXiv preprint arXiv:2508.14918. https://doi.org/10.48550/arXiv.2508.14918

Zhu, X., Zhang, C., Stafford, T., Collier, N., & Vlachos, A. (2025). Conformity in large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 3854–3872). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.acl-long.195