CoChat

Epistemic Communities under Active Inference

Open access

The DOI matches its Crossref record; 165 of the 165 references it cites were found in a registry or an index; 7 of the 18 sources its claims cite were checked for support.

Abstract

The spread of ideas is a fundamental concern of today’s news ecology. Understanding the dynamics of the spread of information and its co-option by interested parties is of critical importance. Research on this topic has shown that individuals tend to cluster in echo-chambers and are driven by confirmation bias. In this paper, we leverage the active inference framework to provide an in silico model of confirmation bias and its effect on echo-chamber formation. We build a model based on active inference, where agents tend to sample information in order to justify their own view of reality, which eventually leads to them to have a high degree of certainty about their own beliefs. We show that, once agents have reached a certain level of certainty about their beliefs, it becomes very difficult to get them to change their views. This system of self-confirming beliefs is upheld and reinforced by the evolving relationship between an agent’s beliefs and observations, which over time will continue to provide evidence for their ingrained ideas about the world. The epistemic communities that are consolidated by these shared beliefs, in turn, tend to produce perceptions of reality that reinforce those shared beliefs. We provide an active inference account of this community formation mechanism. We postulate that agents are driven by the epistemic value that they obtain from sampling or observing the behaviours of other agents. Inspired by digital social networks like Twitter, we build a generative model in which agents generate observable social claims or posts (e.g., ‘tweets’) while reading the socially observable claims of other agents that lend support to one of two mutually exclusive abstract topics. Agents can choose which other agent they pay attention to at each timestep, and crucially who they attend to and what they choose to read influences their beliefs about the world. Agents also assess their local network’s perspective, influencing which kinds of posts they expect to see other agents making. The model was built and simulated using the freely available Python package pymdp. The proposed active inference model can reproduce the formation of echo-chambers over social networks, and gives us insight into the cognitive processes that lead to this phenomenon.

Overview

CoChat Summary

## TL;DR

This paper presents a multi-agent active inference model of opinion dynamics in which agents infer both an abstract binary “Idea” and other agents’ beliefs from observable “Hashtags,” and use those inferences to choose whom to attend to and what to tweet. Simulations show that epistemic confirmation bias, network connectivity, inverse social volatility, and habit learning jointly shape whether groups polarize into echo-chambers or converge toward consensus.

## Key findings

  • The model reproduces confirmation-bias-like dynamics: higher epistemic confirmation bias \(\gamma\) makes agents preferentially sample like-minded neighbours and can drive echo-chamber formation.
  • In the network experiments, increasing connectivity \(p\) generally reduced polarisation, but high \(\gamma\) could counteract this: even at \(p=0.8\), many trials still ended in polarised dynamics.
  • For the \(\gamma\)–\(\omega_{\mathrm{Soc}}\) sweep, the authors report 15 values of \(\gamma\) over \((3,9)\), 15 values of \(\omega_{\mathrm{Soc}}\) over \((0.05,1.0)\), \(N=15\) agents, and \(p=0.4\); low inverse social volatility led to more re-attendance to neighbours.
  • In the habit-formation experiment, higher learning rate \(\eta\) increased polarisation; the average re-attendance rate was “seemingly maximised” when \(\eta \approx 0.3\), and the effect was strongest at high \(\gamma\) and \(\eta\).

## Methods

  • Multi-agent simulations were built under a discrete active inference/POMDP framework, with agents inferring an abstract binary hidden state (“Idea 1” vs “Idea 2”), neighbours’ meta-beliefs, and their own actions from observed tweets/hashtags.
  • The simulations used Erdős–Rényi networks and the freely available Python package `pymdp`; network size ranged from \(N=12\)–30 in general, and the main quantitative experiments used \(N=15\) agents with 100 independent realisations per condition and \(T=100\) timesteps.
  • Main analyses were parameter sweeps over epistemic confirmation bias \(\gamma\), inverse social volatility \(\omega_{\mathrm{Soc}}\), habit-learning rate \(\eta\), and connectivity \(p\), with outcomes summarized using polarisation index \(\rho\) and re-attendance rate \(r\).

## Limitations

  • The paper states that the parameter search “was not exhaustive,” and that large regions of parameter space remain unexplored.
  • It states that network size was largely fixed/small for computational reasons: simulations mainly used \(N=30\) or \(N=15\), \(\omega_{\mathrm{Idea}}=9.0\) was fixed, and the authors say results are not guaranteed to generalise to larger networks.
  • It also says the outcome measures were relatively simple: the polarisation metric is one-dimensional and does not distinguish different belief-network configurations. Evidence from the text also makes clear that the model omits any true environmental signal and only models two mutually exclusive Ideas; this is a modeling simplification the authors themselves acknowledge as needing future extension.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/e24040476

Cites

165 of 165 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure A1 Exploration of the relationship between ‘agreement’ between an agent and one of its neighbours, the epistemic confirmation bias parameter γ, and the epistemic value of reading that neighbour’s tweet content. Here, the two marginal posteriors Q(s^Idea) and Q(s^MBk) are expressed as two Bernoulli distributions with respective parameters p and q, where ‘agreement’ is the case when p=q and hence (1−p)=(1−q). The top row shows heatmaps of the negative ambiguity H, entropy H[Q(o)], and the full epistemic value EV=H+H[Q(o)] for a fixed value of γ=15.0, under all possible values of p and q. The ‘epistemic confirmation bias‘ effect is seen in the negative ambiguity surface H (upper left plot), which is maximised when posterior beliefs about the validity of Idea 1, measured by p, are aligned with posterior beliefs about a neighbour’s meta-belief about Idea 1, q. The bottom row of plots shows a complementary perspective, demonstrating the effect of increasing γ on the epistemic value and its components, for different settings of q when p=0.0. The subplot on furthest to the right of the bottom row shows that increasing γ increases epistemic value most when q is on the same side of 0.5 as p (q=0.2,q=0.4), and the effect of γ on epistemic value deceases once q passes 0.5. Note that the epistemic value is 0 when p=q=0, because although the negative ambiguity is maximised in this case, it is counteracted by the entropy term which is 0 since both posteriors are certain.
Figure A1 Exploration of the relationship between ‘agreement’ between an agent and one of its neighbours, the epistemic confirmation bias parameter γ, and the epistemic value of reading that neighbour’s tweet content. Here, the two marginal posteriors Q(s^Idea) and Q(s^MBk) are expressed as two Bernoulli distributions with respective parameters p and q, where ‘agreement’ is the case when p=q and hence (1−p)=(1−q). The top row shows heatmaps of the negative ambiguity H, entropy H[Q(o)], and the full epistemic value EV=H+H[Q(o)] for a fixed value of γ=15.0, under all possible values of p and q. The ‘epistemic confirmation bias‘ effect is seen in the negative ambiguity surface H (upper left plot), which is maximised when posterior beliefs about the validity of Idea 1, measured by p, are aligned with posterior beliefs about a neighbour’s meta-belief about Idea 1, q. The bottom row of plots shows a complementary perspective, demonstrating the effect of increasing γ on the epistemic value and its components, for different settings of q when p=0.0. The subplot on furthest to the right of the bottom row shows that increasing γ increases epistemic value most when q is on the same side of 0.5 as p (q=0.2,q=0.4), and the effect of γ on epistemic value deceases once q passes 0.5. Note that the epistemic value is 0 when p=q=0, because although the negative ambiguity is maximised in this case, it is counteracted by the entropy term which is 0 since both posteriors are certain.CC BY · PubMed Central
Figure 1 Bayesian network representation of the POMDP generative model. Squares represent priors, likelihoods, or ‘factors’ that relate random variables to one another, and circles represent random variables (stochastic nodes). Different hidden state factors are represented as state variables and the different modality-specific A^(m) arrays of the observation model shown are side by side, since they lead independently to the observations generated in that modality, but are conjunctively dependent on hidden state factors. Note that the B array can be similarly decomposed into different sub-arrays, one per hidden state factor, but it is shown as a single square here for simplicity. The prior over policies is parameterised by E, which has separate prior over control states (E^Who and E^T) for each control state factor. The box at the top right contains mathematical descriptions of each component in the generative mode. Note that while it is included in the graphical model, we left out the C vector since it is not relevant for the current model.
Figure 1 Bayesian network representation of the POMDP generative model. Squares represent priors, likelihoods, or ‘factors’ that relate random variables to one another, and circles represent random variables (stochastic nodes). Different hidden state factors are represented as state variables and the different modality-specific A^(m) arrays of the observation model shown are side by side, since they lead independently to the observations generated in that modality, but are conjunctively dependent on hidden state factors. Note that the B array can be similarly decomposed into different sub-arrays, one per hidden state factor, but it is shown as a single square here for simplicity. The prior over policies is parameterised by E, which has separate prior over control states (E^Who and E^T) for each control state factor. The box at the top right contains mathematical descriptions of each component in the generative mode. Note that while it is included in the graphical model, we left out the C vector since it is not relevant for the current model.CC BY · PubMed Central
Figure 2 Belief dynamics and actions of a single agent in response to a sequence of Hashtag observations from two fictive neighbours. Shown are the history of Bernoulli parameters defining three marginal posterior beliefs of the focal agent: the belief about the truth value of Idea 1 (Q(st^Idea=Idea1), in red), and its beliefs about the beliefs of its two neighbours regarding Idea 1 (Q(st^MB1=Idea1) and Q(st^MB2=Idea1), shown in two shades of blue). Through its generative model, the focal agent believes that its Hashtag observations are caused by two neighbour ‘meta-belief’ states. The focal agent is exposed to a sequence of Hashtag observations for 100 timesteps, where in case of attending to the first neighbour (ut^Who=0), the agent receives observation ot^NT1=Hashtag 1, ot^NT2=Null, and in case of sampling the other neighbour (ut^Who=0), the agent receives observation ot^NT1=Null, ot^NT2=Hashtag 2. Due to the ‘Hashtag semantics’ matrix in its generative model, these two Hashtags, respectively, lend evidence for the two levels of s^Idea. At each timestep the focal agent performs inferences with respect to hidden states Q(st) as well as policies (control states) Q(ut), and then samples a Neighbour Attendance action from the posterior over control states Q(u^Who=0,u^Who=1). Below each subplot is a heatmap showing the temporal evolution of the probability of sampling Neighbour 1 vs. Neighbour 2 over time. Subfigure (a) shows an agent with low γ (3.0) and low ω^Soc (3.0). The agent’s beliefs about both of their neighbors does not lead it to converge on an idea being true or not. Subfigure (b) shows an agent shows an agent with low γ (3.0) and high ω^Soc (9.0). The agent will be more certain about the beliefs of their neighbors, attend less often to their neighbors, quickly converging to neighbour 2. Subfigure (c) shows an agent shows an agent with high γ (9.0) and low ω^Soc (3.0). This agent believes in high volatility and will be driven to continue sampling their neighbors, which will lead them to take longer to converge towards an idea. However, given their γ, the agent does converge towards the first sampled idea. Subfigure (d) shows an agent shows an agent with high γ (9.0) and high ω^Soc (9.0). This agent believes in low volatility and will be driven to sample the same neighbor very quickly, which will lead them to converge towards an idea quickly.
Figure 2 Belief dynamics and actions of a single agent in response to a sequence of Hashtag observations from two fictive neighbours. Shown are the history of Bernoulli parameters defining three marginal posterior beliefs of the focal agent: the belief about the truth value of Idea 1 (Q(st^Idea=Idea1), in red), and its beliefs about the beliefs of its two neighbours regarding Idea 1 (Q(st^MB1=Idea1) and Q(st^MB2=Idea1), shown in two shades of blue). Through its generative model, the focal agent believes that its Hashtag observations are caused by two neighbour ‘meta-belief’ states. The focal agent is exposed to a sequence of Hashtag observations for 100 timesteps, where in case of attending to the first neighbour (ut^Who=0), the agent receives observation ot^NT1=Hashtag 1, ot^NT2=Null, and in case of sampling the other neighbour (ut^Who=0), the agent receives observation ot^NT1=Null, ot^NT2=Hashtag 2. Due to the ‘Hashtag semantics’ matrix in its generative model, these two Hashtags, respectively, lend evidence for the two levels of s^Idea. At each timestep the focal agent performs inferences with respect to hidden states Q(st) as well as policies (control states) Q(ut), and then samples a Neighbour Attendance action from the posterior over control states Q(u^Who=0,u^Who=1). Below each subplot is a heatmap showing the temporal evolution of the probability of sampling Neighbour 1 vs. Neighbour 2 over time. Subfigure (a) shows an agent with low γ (3.0) and low ω^Soc (3.0). The agent’s beliefs about both of their neighbors does not lead it to converge on an idea being true or not. Subfigure (b) shows an agent shows an agent with low γ (3.0) and high ω^Soc (9.0). The agent will be more certain about the beliefs of their neighbors, attend less often to their neighbors, quickly converging to neighbour 2. Subfigure (c) shows an agent shows an agent with high γ (9.0) and low ω^Soc (3.0). This agent believes in high volatility and will be driven to continue sampling their neighbors, which will lead them to take longer to converge towards an idea. However, given their γ, the agent does converge towards the first sampled idea. Subfigure (d) shows an agent shows an agent with high γ (9.0) and high ω^Soc (9.0). This agent believes in low volatility and will be driven to sample the same neighbor very quickly, which will lead them to converge towards an idea quickly.CC BY · PubMed Central
Figure 3 Collective belief dynamics of multi-agent active inference simulations under different generative model parameterisations. Above each panel are listed the parameter values of γ, ω^Soc, and η used in the simulation. Other parameters were fixed with T=50 timesteps, N=30, network connectivity p=0.2, and inverse environmental volatility ω^Idea=9.0. At the beginning of each simulation, every agent’s beliefs about Idea 1 were sampled from a uniform distribution over the interval Q(s^Idea=Idea1)∈(0.4,0.6). Each panel displays the evolving beliefs of all agents about Idea 1 (the Bernoulli parameter of each agent’s respective posterior over s^Idea), with proximity of the belief to 1.0 indicated by colouring along the green-to-blue spectrum (blue beliefs indicate Q(s^Idea=Idea1)>0.5). Panels (A,D) demonstrate polarisation, where two subsets of agents end up believing in two different levels of the Idea hidden state with high certainty. Panels (B,C) on the other hand show examples of consensus, where the whole network converges to the same opinion by the end of the simulation.
Figure 3 Collective belief dynamics of multi-agent active inference simulations under different generative model parameterisations. Above each panel are listed the parameter values of γ, ω^Soc, and η used in the simulation. Other parameters were fixed with T=50 timesteps, N=30, network connectivity p=0.2, and inverse environmental volatility ω^Idea=9.0. At the beginning of each simulation, every agent’s beliefs about Idea 1 were sampled from a uniform distribution over the interval Q(s^Idea=Idea1)∈(0.4,0.6). Each panel displays the evolving beliefs of all agents about Idea 1 (the Bernoulli parameter of each agent’s respective posterior over s^Idea), with proximity of the belief to 1.0 indicated by colouring along the green-to-blue spectrum (blue beliefs indicate Q(s^Idea=Idea1)>0.5). Panels (A,D) demonstrate polarisation, where two subsets of agents end up believing in two different levels of the Idea hidden state with high certainty. Panels (B,C) on the other hand show examples of consensus, where the whole network converges to the same opinion by the end of the simulation.CC BY · PubMed Central
Figure 4 The interaction between epistemic confirmation bias and network connectivity in determining collective outcomes. (Top) a heatmap of the mean polarisation index across S=100 independent realisations of the multi-agent opinion dynamics simulations, for unique combinations of network connectivity p and epistemic confirmation bias precision γ. (Bottom) selected line plots show extreme settings of p (p=0.2 and p=0.8) and γ (γ=3.5 and γ=9.0). Shaded areas around each line represent the standard deviation of the polarisation index across independent realisations.
Figure 4 The interaction between epistemic confirmation bias and network connectivity in determining collective outcomes. (Top) a heatmap of the mean polarisation index across S=100 independent realisations of the multi-agent opinion dynamics simulations, for unique combinations of network connectivity p and epistemic confirmation bias precision γ. (Bottom) selected line plots show extreme settings of p (p=0.2 and p=0.8) and γ (γ=3.5 and γ=9.0). Shaded areas around each line represent the standard deviation of the polarisation index across independent realisations.CC BY · PubMed Central
Figure 5 (Above left) a heatmap of the polarisation index for all 225 combinations of inverse belief volatility and epistemic confirmation bias precision. (Above right) a heatmap of the re-attendance rate for all 225 combinations of inverse belief volatility and epistemic confirmation bias precision. (Below left) a line plot of the most extreme rows of the polarisation heatmap. (Below right) a line plot of the most extreme columns of the re-attendance rate heatmap.
Figure 5 (Above left) a heatmap of the polarisation index for all 225 combinations of inverse belief volatility and epistemic confirmation bias precision. (Above right) a heatmap of the re-attendance rate for all 225 combinations of inverse belief volatility and epistemic confirmation bias precision. (Below left) a line plot of the most extreme rows of the polarisation heatmap. (Below right) a line plot of the most extreme columns of the re-attendance rate heatmap.CC BY · PubMed Central
Figure 6 (Above left) a heatmap of the polarisation index for all 225 combinations of learning rate and epistemic confirmation bias precision. (Above right) a heatmap of the re-attendance rate for all 225 combinations of learning rate and epistemic confirmation bias precision. The parameters represent the centres of the normal distribution sampled from across trials for each configuration. (Below left) the most extreme row of the polarisation index heatmap. (Below right) the most extreme column of the re-attendance rate heatmap.
Figure 6 (Above left) a heatmap of the polarisation index for all 225 combinations of learning rate and epistemic confirmation bias precision. (Above right) a heatmap of the re-attendance rate for all 225 combinations of learning rate and epistemic confirmation bias precision. The parameters represent the centres of the normal distribution sampled from across trials for each configuration. (Below left) the most extreme row of the polarisation index heatmap. (Below right) the most extreme column of the re-attendance rate heatmap.CC BY · PubMed Central

Verification

Claims the paper makes

5 Valid · 2 Not valid · 5 Not checked · 112 Own results and interpretations

82 more claims

Create a free account to see all 124 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 21
  • It may state two separate claims as one. 19

Create a free account to see these claims.

Citations

Cites

Show all 165 references

Research works printed but not linked: 5

  1. Heerwagen, 1995, Humans, habitats, Biophilia Hypothesis, 138, 138
  2. Oswald, M.E., and Grosjean, S. (2004). Confirmation bias. Cognitive Illusions: A Handbook on Fallacies and Biases in Thinking, Judgement and Memory, Psychology Press.
  3. Ladegard, 2006, Accessing Social Capital through Conformity: Investigating Network Cultures, Adv. Organ. Stud., 20, 153
  4. Gumperz, 2009, The speech community, Linguist. Anthropol. A Read., 1, 66
  5. Sutton, R.S., and Barto, A.G. (2018). Reinforcement Learning: An Introduction, MIT Press.

Cited by

Metadata

Authors
Albarracin, M., Demekas, D., Ramstead, M. J. D., & Heins, C.
Affiliations
  • Université du Québec à Montréal (CA)
  • Imperial College London (GB)
  • Wellcome Centre for Human Neuroimaging (GB)
  • University of Konstanz (DE)
Corresponding author
  • Albarracin, Mahault
  • Heins, Conor
Venue
Entropy
Publisher
MDPI AG
Published
2022-03-29
Volume
24
Issue
4
Pages
476
ISSN
1099-4300 (ISSN-L)
OpenAlex
W4220746083
Cited by
60
References
157
Citations by year
2026: 15 · 2025: 10 · 2024: 14 · 2023: 15 · 2022: 6
FWCI
5.00
Type
journal article
Language
en
Open access
Locations
  • Entropy (publishedVersion, cc-by)
  • Preprints.org (submittedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • KOPS (University of Konstanz) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
U.S. Office of Naval Research (N00014-19-1-2556)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Embodied and Extended Cognition
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
  • Physical Sciences › Physics and Astronomy › Statistical and Nonlinear Physics › Opinion Dynamics and Social Influence
Keywords
active inference, confirmation bias, echo chambers, epistemic communities, belief updating, information dissemination, social networks, in silico models, belief certainty, generative models

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-06, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-06 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.