PaperlayerView exact revision on arXiv

Source-linked AI summary

GenAI as an 'Invisible Hand' in Opinion Dynamics: Evidence from a Group-Level Simulation Study

Thomas Grisold, Johannes Schneider

arXiv:2610.37689v1cs.SIcs.AI

TL;DR

The paper examines when generative AI influences collective opinion dynamics, a concern raised by arguments that AI outputs may shape social beliefs and norms. Through simulations of repeated group discussions, it varies LLM agreeableness and personalization across social configurations. It finds that personalization increases influence for non-neutral LLMs, while neutral LLMs have stronger effects in homogeneous groups.

  • Problem

    The paper asks when and how genAI influences collective opinion dynamics, given concerns that AI outputs and tendencies may be taken up collectively.

    Source: Introduction · p. 1

  • Method

    The authors simulate repeated controversial group discussions using LLMs as human proxies and vary agreeableness, personalization, and social configurations.

    Source: Introduction · p. 2, Simulation Design Involving LLMs · p. 4

  • Results

    Profile-based personalization increases LLM influence when models agree or disagree with users, while neutral LLMs have stronger effects in homogeneous groups.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    The findings suggest that genAI’s influence on collective opinion making depends on model responses, user information, and the social environment.

    Source: Introduction · p. 2

    Because LLM agents may distort human behavior and flatten identity heterogeneity, these simulated patterns are not direct evidence about human collectives and require testing with human participants.

    Source: Limitations · p. 15

Abstract

from arXiv · show

Recent arguments suggest that genAI can act as an invisible hand that gradually shapes social structures, such as collective beliefs and norms. The argument is that if millions of people repeatedly consult genAI tools, their generated outcomes quietly shape what becomes acceptable, reasonable or conventional. We study this in the context of collective opinion dynamics, i.e., how groups form opinions around controversial dynamics. We present the findings of a simulation including 50,000 group discussions across 50 controversial statements, three LLMs, and different social configurations. We vary two characteristics of LLM interaction-agreeableness and personalization-and compare the influence of genAI in homogeneous and heterogeneous groups. Our findings indicate that profile-specific personalization increases LLM influence when models adopt a non-neutral stance, i.e., when they agree or disagree with the user. Furthermore, we find that neutral LLMs have a stronger effect in homogeneous groups. Based on our simulation, we generate a set of hypotheses to specify how, why and when AI-related features and social configurations jointly shape collective opinion making.

Introduction

The paper asks when generative AI influences collective opinion dynamics, given concerns that widely consulted outputs may shape shared beliefs and norms. It uses simulation to examine LLM agreeableness and personalization across homogeneous and heterogeneous groups, finding that personalization matters for non-neutral models and neutral models have stronger effects in homogeneous groups.

Introduction

  • Widespread consultation of human-like LLM outputs may gradually shape what people perceive as conventional, potentially affecting cultural practices, norms, and collective beliefs.The authors frame this gradual influence as an “invisible hand.”Source: Introduction · p. 1
  • The study asks how and when generative AI influences collective opinion dynamics, focusing on LLM agreeableness, personalization, and homogeneous versus heterogeneous social groups.Agreeableness refers to affirming or reinforcing a user’s position; personalization adapts responses to user preferences.Source: Introduction · pp. 1–2
  • The authors use a multi-factor simulation to explore opinion changes over repeated discussions and generate hypotheses about mechanisms and possible interventions.They present this approach as a way to examine the phenomenon under different conditions.Source: Introduction · p. 2
  • Profile-based personalization increases LLM influence when models agree or disagree with users, while neutral LLMs have stronger effects in homogeneous than heterogeneous groups.The authors suggest neutral LLMs may have stronger effects in homogeneous groups because interactions among similar users would otherwise reinforce existing views.Source: Introduction · p. 2

Background

The background frames genAI as a possible influence on social practices, norms, and collective beliefs, then focuses on agreeableness and personalization as LLM interaction dimensions relevant to opinion formation. The paper examines how these dimensions may shape collective opinion dynamics.

GenAI as an Invisible Hand: Collective Use and Social Change

  • The paper asks when and how genAI influences social collectives, building on arguments that repeated use can bring AI-generated outcomes and tendencies into collective life.The background describes possible effects on beliefs, norms, and social structures.Source: GenAI as an Invisible Hand: Collective… · pp. 2–3
  • GenAI outputs may become part of everyday conduct and influence social practices, conventions, and norms, while prior work reports effects on language and collective sense-making.The background also cites concerns about ideological bias in outputs and shifts in individual policy views.Source: GenAI as an Invisible Hand: Collective… · pp. 2–3
  • AI-based technologies may embody logics and rationales that threaten human values, and LLMs have been reported to influence sense-making in ethically problematic situations.The cited work also reports possible shifts in policy views and concerns about ideological tendencies in generated outputs.Source: GenAI as an Invisible Hand: Collective… · pp. 2–3

Dimensions of LLMs Driving Opinion Dynamics

  • Agreeableness is the extent to which a model confirms a user’s position rather than challenging it; prior work finds models may affirm reasoning even when it conflicts with social norms.Agreeableness may be deliberately prompted or arise implicitly from model training.Source: Dimensions of LLMs Driving Opinion… · p. 3
  • Personalization adapts LLM responses to user-specific information, and prior research associates user-memory profiles with larger increases in agreement-oriented behavior.Personalization may use explicit user information or implicit inferences from prior interactions.Source: Dimensions of LLMs Driving Opinion… · p. 3
  • The study examines how agreeableness and personalization influence opinion formation in social collectives, extending questions commonly inferred from individual-level research.The authors note that personalization’s effect on whether models question their own positions may depend on the interaction role they adopt.Source: Dimensions of LLMs Driving Opinion… · pp. 3–4

Research Design: Simulations for Hypothesis Generation

The study uses LLM-based simulations of repeated group discussions to explore how model behavior and social configurations shape opinion dynamics, with the aim of generating hypotheses for future research. It varies agreeableness and personalization, samples multiple models, statements, and network structures, and compares outcomes with human-only discussions.

Research Design: Simulations for Hypothesis Generation

  • The simulation explores the possibility space of genAI influence on social collectives, using multiple conditions to generate hypotheses for future studies.The authors describe this as an initial investigation of genAI’s influence on social collectives.Source: Research Design: Simulations for… · p. 4

Simulation Design Involving LLMs

  • LLMs simulate human utterances in controversial group discussions, allowing the study to examine whether and how LLM use steers interactions over multiple rounds.The authors note that such simulations can distort behavior and flatten social identities.Source: Simulation Design Involving LLMs · p. 4

Simulation Setup

  • Discussions proceed in sequential rounds of ratings with brief explanations; participants remember interactions, which can alter their current views.Human–LLM interactions follow the same general pattern, with instructions and user information varying by setup.Source: Simulation Setup · p. 4

Simulation Parameters

  • The design varies agreeableness and personalization, including agree, disagree, or neutral instructions and prompts with no user details, user profiles, or prior responses.It also includes a default LLM condition without explicit agreeableness instructions or personalization.Source: Simulation Parameters · p. 6

Simulation Runs

  • The study runs 50,000 group discussions across 50 controversial statements, three LLMs, and homogeneous or heterogeneous networks.Homogeneous groups contain an ego and three similar primary contacts; heterogeneous groups include primary, secondary, and tertiary contacts.Source: Simulation Runs · p. 7

Analysis and Metrics

  • The analysis compares average Likert-scale ratings with human-to-human interaction as a reference and tests setting contrasts across statements.The authors also use redundant settings as robustness checks and examine correlations between models.Source: Analysis and Metrics · p. 9

Findings

LLM influence depended on both personalization and group structure: profile information strengthened influence under non-neutral stances, while neutral models had larger effects in homogeneous groups. Model results were highly aligned in aggregate, but correlations were lower for individual questions.

Findings

  • In heterogeneous groups, a neutral LLM without personalization was closest to human-only discussion; paired differences across prompt variants were nonsignificant.The authors attribute small differences between ostensibly redundant neutral prompts to prompt sensitivity.Source: Findings · p. 9
  • Profile information produced the strongest LLM influence under agreeable and disagreeable stances; both comparisons against no-profile conditions had p-values below 0.001.The profile-based influence was stronger than that from generic agreeableness instructions or a user's prior response.Source: Findings · p. 9, Figure 3 but computed on differences… · p. 11
  • Disagreement generally shifted opinions more than agreement, though its added impact over neutrality was modest without profile information.For homogeneous groups without personalization, the disagreement-versus-neutral gap was 0.005, compared with 0.054 for neutral versus no personalization.Source: Findings · p. 10, Figure 3 but computed on differences… · p. 11
  • Aggregate model alignment was high, with the weakest correlation between the European and US models.Table 1 reports model correlations for differences shown in Figure 3.Source: Findings · p. 10, Table 1 · p. 11
  • Final-response correlations across individual questions were 0.3 to 0.4, lower than aggregate-level correlations.The authors characterize this range as the higher end of medium for psychological studies.Source: Table 2 · p. 11, Findings · p. 11

Figure 3 but computed on differences using individual statements

Hypothesis Generation

The paper uses a simulation paradigm that substitutes LLMs for humans to explore collective opinion dynamics. It formulates hypotheses along two dimensions: AI-related features and the social configurations in which AI has an effect.

Hypothesis Generation

  • The paper uses a simulation paradigm that substitutes LLMs for humans to examine collective opinion dynamics.The simulation is used to generate observations and formulate hypotheses about genAI's influence.Source: Hypothesis Generation · p. 11
  • The hypotheses address how AI-related features shape collective opinion making.Source: Hypothesis Generation · p. 11
  • The hypotheses also address how social configurations shape the effects of AI on collective opinion dynamics.Source: Hypothesis Generation · p. 11

AI-related Features

The paper proposes hypotheses about how AI-related features, especially profile-based personalization and model stance, shape opinion influence. It also proposes that disagreement and the depth of user information condition that influence.

AI-related Features

  • The proposed AI-feature hypotheses center on personalization and how user profile information compares with simply knowing a person's current response.The authors find profile information matters more, particularly when the AI is instructed to agree or disagree.Source: AI-related Features · p. 11

Personalization-based Influence

  • Profile-personalized responses are proposed to exert greater influence on opinion making than generic responses.The authors suggest user profiles may help the LLM tailor arguments to individual characteristics.Source: Personalization-based Influence · p. 12

Personalization and Non-neutral Stance

  • Personalization is proposed to have a stronger influence when the LLM adopts a non-neutral stance than when it remains neutral.Personalized neutral responses produced comparatively limited changes, unlike responses that agree or disagree.Source: Personalization and Non-neutral Stance · p. 12

Disagreement and Personalization

  • Profile-personalized disagreement is proposed to produce greater opinion change than profile-personalized agreement.The authors suggest that users' limited room to move further in an existing extreme and models' baseline tendency toward agreement may contribute to this asymmetry.Source: Disagreement and Personalization · p. 12

Depth of Personalization-related Information

  • Personalization based on a broader user profile is proposed to produce greater opinion change than personalization based only on a user's previous response.The authors distinguish prior responses as information about current concerns and profiles as broader context about user characteristics.Source: Depth of Personalization-related… · p. 13

Social Configurations

The simulations link collective opinion dynamics to social structure: LLM influence varies with group homogeneity and the availability of alternative human viewpoints, while out-of-trajectory LLM perspectives may interrupt within-group reinforcement.

Social Configurations

  • The authors contrast homogeneous and heterogeneous networks to examine how social-group diversity shapes AI’s influence, and develop three hypotheses about social context.Source: Social Configurations · p. 13

Social Homogeneity

  • Neutral LLMs have more influence in homogeneous than heterogeneous groups, while profile-personalized agreeing or disagreeing models show the highest effect across both structures.The authors propose that homogeneous groups reinforce existing views, making a neutral LLM perspective more distinct; they state this as hypothesis H5.Source: Social Homogeneity · p. 13

Alternative Viewpoints

  • The authors propose that LLM influence increases as viewpoint diversity from human interaction decreases, while noting that this assumption needs testing independently of network structure.Their simulation contrasts heterogeneous groups with diverse contacts against homogeneous groups of mainly similar peers.Source: Alternative Viewpoints · pp. 13–14

Disruption of Echo Chamber

  • In homogeneous groups, a neutrally instructed LLM moves group ratings toward moderation, potentially weakening reinforcement among similar individuals.The authors propose that LLM perspectives differing from the group’s prevailing trajectory weaken within-group opinion reinforcement (H7).Source: Disruption of Echo Chamber · p. 14

Implications

The paper examines how agreeableness and personalization shape opinion formation across social settings. Its simulation suggests that model stance and profile-based personalization matter, and it discusses possible implications alongside directions for further study.

Implications

  • Profile-based personalization increases LLM influence when models take a non-neutral stance, while neutrality has effects similar to disagreement except when profile information is used.The authors suggest that requiring a “be neutral” system-prompt instruction could counteract extremist views, but could also reduce diverse views.Source: Implications · p. 14
  • The simulation varies agreeableness and personalization across social settings and interaction rounds to examine their role in collective opinion making.These dimensions have been associated with user behavior and social outcomes.Source: Implications · p. 14
  • The authors present their hypotheses as starting points for experimental research and call for studying mixed network structures, adoption patterns, more realistic agents, and broader topic coverage.They also identify larger-scale social norms and conventions as a possible direction for future work.Source: Implications · pp. 14–15

Limitations

The findings are patterns produced under the simulation’s assumptions, not direct evidence of how human collectives respond to LLMs; the modeled opinions and social settings are also simplified.

Limitations

  • LLM-based simulated participants can distort responses and flatten identity heterogeneity, so findings require empirical testing with human participants.Opinion formation is represented with a five-point Likert scale, repeated interactions, and memory of the four most recent encounters.Source: Limitations · p. 15
  • The study measures influence magnitude rather than direction and models only selected personalization and network conditions, so its hypotheses are not established claims about real-world opinion formation.The authors note that real-world relationships, identities, histories, network structures, and contexts are more nuanced.Source: Limitations · p. 15

Related papers

Loading 2610.37689v1…