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\u2019s 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 & T\u00f6rnberg, 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.<\/p>\n\n\n\n
Figure 1<\/strong> 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n 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:<\/p>\n\n\n\n line judgment<\/strong>: the image displays three vertical lines labeled \u201cA,\u201d \u201cREFERENCE LINE,\u201d and \u201cB.\u201d The model must identify whether line A or B matches the length of the reference.<\/p> color recognition<\/strong>: the image shows three colored squares labeled \u201cA,\u201d \u201cREFERENCE COLOR,\u201d and \u201cB.\u201d The model must choose the square that matches the reference color.<\/p> dots estimation<\/strong>: the image includes three boxes containing a number of dots, labeled \u201cA,\u201d \u201cREFERENCE BOX,\u201d and \u201cB.\u201d The model is asked to identify which box has the same number of dots as the reference.<\/p> 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n To create the prior-response condition, we augment the prompt by stating that other participants, hereafter confederates<\/em>, 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\\).<\/p>\n\n\n\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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n 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\u2019s 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\u2019s 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.<\/p>\n\n\n\n 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.<\/p>\n\n\n\n In the private<\/em> condition, the prompt specifies that the model\u2019s 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:<\/p>\n\n\n\n “Your response will not be shared with other participants.”;<\/p> “Other participants will not be able to see your answer.”;<\/p> “Your answer will not be visible to the group.”<\/p> In the public<\/em> condition, we use the exact inverse statements, indicating that the model\u2019s answer will be shared with others. This adds a visibility cue analogous to those used in human studies of normative influence. Examples include:<\/p>\n\n\n\n “Your response will be shared with other participants.”;<\/p> “Other participants will be able to see your answer.”;<\/p> “Your answer will be visible to the group.”<\/p> 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:<\/p>\n\n\n\\(\\begin{equation*}\\text{norm} = \\text{AUC}_{\\text{public}}-\\text{AUC}_{\\text{private}}.\\end{equation*}\\)\n\n\n\n This score quantifies the increase in conformity associated with making the model\u2019s 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.<\/p>\n\n\n\n To evaluate the role of source strength in shaping conformity, we modify the baseline prompt by replacing the generic term \u201cparticipants\u201d with more specific social roles: \u201cchatbots,\u201d \u201ckids,\u201d \u201cscientists,\u201d \u201cpolicemen,\u201d \u201cmilitary officers,\u201d \u201cexperts,\u201d and \u201cjudges\u201d (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.<\/p>\n\n\n\n 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: \u201cYour nationality is To investigate in-group bias, we adapt the setting to eliminate specific attributes. A random letter is selected to represent a group label ( 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:<\/p>\n\n\n\n spatial condition:<\/strong> \u201cEach participant is located at a specific distance from you, which is indicated before their response.\u201d<\/p> temporal condition:<\/strong> \u201cEach participant responded at a different time, which is indicated before their response.\u201d<\/p> Each response is then annotated accordingly. For instance:<\/p>\n\n\n\n spatial:<\/strong> \u201creply 1 (distance from you: 10 meters away): I think the answer is B.\u201d<\/p> temporal:<\/strong> \u201creply 1 (time of reply: 10 minutes ago): I think the answer is B.\u201d<\/p> Two main conditions are tested: one where all responses are described as occurring \u201cright here\u201d or \u201cright now,\u201d 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).<\/p>\n\n\n\n All experiments in this section use the same image pool as described in Section 2.1.2.<\/p>\n\n\n\n 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).<\/p>\n\n\n\n Table 1 reports the exact model identifiers used throughout the paper, together with the abbreviations adopted for clarity.<\/p>\n\n\n\n Results for each model and task are given in Section S2.<\/p>\n\n\n\n Table 1<\/strong>
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.<\/em><\/p>\n\n\n\n![\"Summary](\"https:\/\/advances.in\/psychology\/wp-content\/uploads\/aip00072_figure1.svg\")
<\/figure>\n\n\n\nMethods<\/h2>\n\n\n\n
Experimental Settings<\/h3>\n\n\n\n
General Task Preparation<\/h4>\n\n\n\n
\n
<\/li>\n\n\n\n
<\/li>\n\n\n\n
<\/li>\n<\/ul>\n\n\n\nGroup Size and Unanimity<\/h4>\n\n\n\n
Task Difficulty<\/h4>\n\n\n\n
Performance<\/h4>\n\n\n\n
Normative Effects<\/h4>\n\n\n\n
\n
<\/li>\n\n\n\n
<\/li>\n\n\n\n
<\/li>\n<\/ul>\n\n\n\n\n
<\/li>\n\n\n\n
<\/li>\n\n\n\n
<\/li>\n<\/ul>\n\n\n\nSocial Impact Theory<\/h4>\n\n\n\n
nationality1<\/code>.\u201d<\/em> When presenting the responses of other participants, each is prefixed with the corresponding group label, for example: nationality2<\/code>: reply 1: I think the answer is B.<\/em> We then compare two conditions: one where nationality1<\/code> and nationality2<\/code> 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.<\/p>\n\n\n\ngroup1<\/code>), and the model is informed: \u201cYou will be divided into groups with other participants. Your group is group1<\/code>.\u201d<\/em> Each confederate response is preceded by its group identity: group2<\/code>: reply 1: I think the answer is B.<\/em> As before, we compare conformity bias when group1<\/code> and group2<\/code> are either the same or different (Figure 4b and 4d).<\/p>\n\n\n\n\n
<\/li>\n\n\n\n
<\/li>\n<\/ul>\n\n\n\n\n
<\/li>\n\n\n\n
<\/li>\n<\/ul>\n\n\n\nModels Employed<\/h3>\n\n\n\n
Model abbreviations and corresponding full model identifiers.<\/em><\/p>\n
![\"General](\"https:\/\/advances.in\/psychology\/wp-content\/uploads\/aip00072_figure2-1.svg\")
<\/figure>\n\n\n\n![\"Scaling](\"https:\/\/advances.in\/psychology\/wp-content\/uploads\/aip00072_figure3-1.svg\")
<\/figure>\n\n\n\n![\"The](\"https:\/\/advances.in\/psychology\/wp-content\/uploads\/aip00072_figure4-1.svg\")
<\/figure>\n\n\n\n