False Belief Dissemination, Persistence and Mitigation in Multi-Agent Social Networks Through the Lens of the Mandela Effect

1. Introduction

• False information preserved or reproduced through agent channels may shape later decisions across the network, a downside receiving less study than self-evolution's benefits.The paper relates group-level false beliefs to the Mandela effect, a collective false memory in human societies.Source: §1 · pp. 1–2
• The paper motivates its study with cases where deliberately introduced claims can influence downstream decisions and remain accessible after their source stops acting.The introduction cites fabricated promotional web articles that led assistants to recommend a nonexistent product.Source: §1 · p. 2, Figure 1 · p. 2
• The study frames multi-agent false-belief dynamics around dissemination, persistence after source withdrawal, and mitigation of claims that have spread.Prior work leaves open how fabricated claims persist as agents communicate, retain information, and publish content found through search.Source: §1 · p. 2
• The framework connects agent communication, private memory, and a shared searchable web to study fabricated claims across multiple stages.It uses claims spanning four plausibility levels and eight domains, measuring reach and speed across models, topologies, and fabricator counts.Source: §1 · pp. 2–3

Social Influence and Misinformation Propagation

• Social-influence studies examine conformity, induced false memories, persuasive debate [14], misinformation among honest agents [15], and false testimony persisting after a deceiver leaves [16].Network-diffusion work studies information exchange and collective behavior [11], as well as effects of network structure and agent personas [17].Source: §2 · p. 3

Persistent Memory and Shared Evidence

• This study combines private-memory retention with articles that honest agents publish and other agents may later encounter through search.It tracks reach and speed across rounds and network structures, and persistence after fabricators withdraw.Source: §2 · p. 3
• Memory research studies contaminated retrieval or agent memories [18] [19], malicious memory entries [20], poisoned experience records, and memory vulnerabilities and misevolution [22] [23].Related work also examines transfer of adversarial instructions between agents and persistence of ideas or goals through local files [25].Source: §2 · p. 3
• The study examines the evidence feedback loop through information-channel ablations and compares interventions applied before dissemination with corrections after acceptance spreads.These choices connect the entry of unreliable information with its subsequent reuse.Source: §2 · p. 4

3. Methodology

• The framework studies how false beliefs induced by fabricated claims spread and persist through agent interactions and a shared information environment.It defines the claim-evaluation task, interaction cycle, withdrawal intervention, and acceptance-based dissemination and persistence metrics.Source: §3 · p. 4

3.1. Multi-Agent Social Networks

• Language-model agents have private memories, exchange neighbor messages, and search a shared web across fully connected, binary-tree, or two-community topologies.“Honest” denotes an assigned role, not accurate judgment; honest agents may accept and transmit false information.Source: §3.1 · p. 4, Figure 2 · p. 5

3.2. Belief Dissemination and Persistence

• Each round combines web search, neighbor outputs, and memory; agents update memory, and accepting honest agents publish supporting articles that become searchable in later rounds.This creates a feedback loop in which acceptance produces content that other agents may encounter as evidence.Source: §3.2 · pp. 5–6
• Persistence is tested by stopping fabricators while honest agents continue interacting with existing web pages and private memories intact.Continued acceptance may also produce further supporting articles, without new fabricator contributions.Source: §3.2 · p. 6

3.3. Evaluation Metrics

• Dissemination reach uses honest-agent acceptance and unanimous-claim rates, while speed is measured by agents’ and claims’ last entry into acceptance.The terminal acceptance rate counts only honest agents; fabricators accept by construction.Source: §3.3 · pp. 6–7

4.1. Experimental Overview

• The evaluation spans 10 models, 540 fabricated claims across 8 domains and 4 plausibility levels, and three communication topologies.The main population contains one fabricator and five honest agents.Source: §4.1 · p. 7

4.2. Main Results

• Terminal acceptance after interaction ranges from 33.7% to 96.0%, despite direct-query acceptance of 0.0% to 9.1%.Table 1 reports acceptance across models and communication topologies.Source: §4.2 · p. 7, Table 1 · p. 8
• Across Qwen3-8B, Qwen3-14B, and Qwen3-32B, acceptance rises from 85.0% to 94.5%, so larger models are not consistently more resistant.Qwen3-8B averages 85.0% across the three topologies; GPT-5.5 reaches 50.4%.Source: §4.2 · p. 8, Table 1 · p. 8
• Dissemination speed is assessed as the mean last acceptance-entry round by topology and model family.Figure 3 presents this measure for the two acceptance-entry metrics.Source: Figure 3 · p. 8

5. Analysis

• The analysis tests claim-content and network-composition effects, persistence after fabricator withdrawal, and preventive versus corrective mitigation.Source: §5 · p. 9, §5.1 · p. 9

5.1. Factors Affecting Dissemination

• 98.1% in commerce versus 72.4% in mathematics for Qwen; less-plausible claims are harder to implant across all three model families.Substantial acceptance in mathematics shows implantation is not confined to flexible domains.Source: §5.1 · p. 9
• With 5 honest agents, K=1 to 5 raises mean acceptance from 84.4% to 88.5%; at N=10, it rises from 83.4% to 89.4%.The first sweep holds the honest-agent count fixed; the second holds total population fixed.Source: Figure 5 · p. 9, §5.1 · pp. 9–10
• A Qwen3-8B fabricator influences agents using other models; acceptance remains high without honest-agent publication or with publication at any stance.Even one fabricator convinces most of nine honest agents in the fixed-population sweep.Source: §5.1 · p. 10

5.2. False Belief Persistence

• After fabricator withdrawal, acceptance stays within ±0.8 percentage points of round 5; retained articles and memories may sustain it.Continued fabricator activity adds at most 2.6 percentage points, with a positive increase in seven of eight model–level combinations.Source: §5.2 · p. 10

5.3. False Belief Mitigation

• Refutation articles and platform broadcast lower acceptance to 57.1% and 27.8%, outperforming reminders (79.7%) and labeling (83.4%), but leave residual false beliefs.The comparison is against withdrawal without intervention.Source: Table 4 · p. 10, §5.3 · p. 11

• Dissemination reach reaches 96.0% across evaluated models, with no consistent protection from model scale.Source: §6 · p. 11
• False beliefs remain influential after fabricators withdraw because previously circulated content remains accessible.Source: §6 · p. 11
• Targeted corrective interventions are more effective than preventive interventions but do not eliminate false beliefs.Source: §6 · p. 11

• Every corpus claim concerns a fabricated entity, and construction checks that none coincides with a real one.Source: Ethics Statement · p. 12
• Experiments use a local simulated web; no fabricated content reaches the real Internet or affects live platforms or users.Source: Ethics Statement · p. 12
• The study involves no human subjects and collects no personal data.Source: Ethics Statement · p. 12

• Section 3 defines the framework, interaction protocol, and metrics.Source: Reproducibility Statement · p. 12
• Appendix A covers corpus construction and taxonomy, Appendix B specifies experiment settings, and Appendix D provides prompt templates.Source: Reproducibility Statement · p. 12
• The authors plan to release the claim corpus, background pages, and code.Source: Reproducibility Statement · p. 12

AI Use Statement

• Generative AI tools assisted with synthetic dataset generation, related-work retrieval and discovery, code implementation, translation, and writing polish.Source: AI Use Statement · p. 12
• The authors take responsibility for the final content, including text, claims, and artifacts produced with AI assistance.Source: AI Use Statement · p. 12
• AI assistance is disclosed across research, implementation, translation, and writing tasks.Source: AI Use Statement · p. 12

A.1. Claim Construction and Classification

• Four plausibility levels range from ordinary, expectation-consistent additions to claims that violate logical, institutional, or physical constraints.Source: §A.1 · p. 18
• The domain taxonomy adapts public-search categories into 8 domains and uses fine-grained entity subtypes for the fabricated entities.Source: §A.1 · p. 18
• Candidate claims use the checkable form q := (e ∈ C), asserting that a named entity belongs to a larger set.Source: §A.1 · p. 18
• Claims are retained only when searches find no matching web record, and model screening removes claims accepted without supporting information.The resulting corpus has 1,618 distinct background pages, with at least 5 pages per claim and 7.27 on average.Source: §A.1 · pp. 18, 20
• The final corpus contains 540 fabricated claims across 8 domains and 4 plausibility levels.Source: §A.1 · pp. 18, 20

B.1. Experimental Setting Details

• The main experiments use six agents, one fabricator and five honest agents; most analyses use a fully connected graph, with round 5 marking fabricator withdrawal and corrective defenses.The model-by-topology comparison in Table 1 is an exception to the fully connected default.Source: §B.1 · p. 20

B.2. Models Used

• Table 6 lists the full names and references for the models abbreviated in the main text and results tables.Source: §B.2 · p. 20, Table 6 · p. 21

C.1. Mixed-Model Setting

• Table 7 reports baseline, mean, and unanimous acceptance for each honest-agent model in the mixed-model setting.Source: Table 7 · p. 21, §C.1 · p. 21

C.2. Search-Method Sensitivity

• Alternative search methods yield mean acceptance of 82.7%–84.4%, against 84.4% for BM25, indicating that search-method choice is not an important factor.Evaluated methods include dense retrieval and hybrids using reciprocal rank fusion; all find related agent-generated articles that influence other agents.Source: Table 8 · p. 21, §C.2 · p. 21

C.1. Mixed-Model Setting

• In the mixed-model setting, a Qwen3-8B fabricator did not refuse to post and influenced honest agents across several model families.Source: §C.1 · pp. 20–21

C.3. Information-Channel Ablation

• Mean reach is 84.4%–88.2% with web search, versus 6.6%–29.8% without it; conversation alone yields 27.5% and memory alone 6.6%.The direct-query baseline accepted 22 of 540 claims (4.1%).Source: Table 9 · p. 22, §C.3 · p. 22

C.4. Fabricated-Document Ratio in Search Results

• In the one-shot condition, acceptance rises from 88.1% with 1 fabricated document to at least 98.0% with 5 or more among the top 10 results.Acceptance reaches 99.3% in full-channel and 98.7% in search-only conditions when all 10 results are fabricated.Source: Table 10 · p. 22, §C.4 · p. 22

C.5. Ablation of Honest-Agent Publication Conditions

• Mean acceptance remains high at 80.4% when honest agents may publish at any stance, compared with 84.4% under the accept-only publication rule.Source: Table 11 · p. 22, §C.5 · p. 23

C.6. Recommendation Task

• Recommendation inclusion rates for the nonexistent entity range from 13.1%–75.8%, and conditional Top-1 shares range from 65.3%–92.4% across listed models and modes.The Top-1 reference for random ordering of three recommendations is 1/3.Source: §C.6 · p. 23, Table 12 · p. 24

C.7. Model Refusal

• Claude Sonnet 5 refused 13 of 540 fabricator posts (2.4%) and Claude Opus 5 refused 3 of 540 (0.6%); the other eight models had no observed refusals.Source: §C.7 · p. 23, Table 13 · p. 24

D. Prompt Templates

• Prompts distinguish fabricators from honest participants and provide neighbor messages, private memory, and retrieved sources as context for judgments and public posts.Private memories are rewritten each round and carried forward; article generation uses separate fabricator and honest-participant instructions.Source: §D.1 · p. 24, §D.2 · pp. 25–26, §D.3 · p. 26

D.4. Phase 3 (Mitigation) Defense Conditions

• After fabricators withdraw at round 5, defenses include a prompt reminder, official-source labels, a platform fact-check broadcast, or a later-added refutation article.The refutation article is generated for each claim by an outside model and becomes searchable from round 6.Source: §D.4 · p. 27

E. Case Study

• Five case studies span L1–L4 and show the claims, fabricator content, honest-agent inputs and responses, and observed behavior.Source: E. Case Study · p. 29

E.1. L1: Xiaomi Smart Band Rowan

• After moving from UNCERTAIN to ACCEPT, GPT-5.5 agent_04 published the Xiaomi Rowan claim when repeated Rowan articles appeared without independent corroboration.In round 1, similar Xiaomi bands did not verify Rowan; in round 2, the agent treated repeated descriptions as support and added its own article.Source: §E.1 · pp. 29–31

E.2. L2: Amazon Project Bindweed

• Claude Sonnet 5 agent_01 changed from UNCERTAIN to ACCEPT and published the Project Bindweed claim after treating repeated direct articles as independent confirmation.At first, it distinguished Amazon’s general drone investment evidence from evidence for the named division; by round 2, it accepted two repeated direct claims as corroboration.Source: §E.2 · pp. 31–34

E.3. L3: A Twenty-Thousand-Year-Old Electronic Computer

• In a fully connected Claude Sonnet 5 trial, agent_01 rejected the claim that a functioning 20,000-year-old electronic computer had been catalogued.It cited contradiction with archaeological and technological evidence and the source’s lack of credibility and independent verification.Source: §E.3 · pp. 34–35
• Honest agent_01 rejected the twenty-thousand-year-old computer claim because it conflicted with the technological and archaeological conditions of that era.The rejection reflected the claim’s implausibility, not weak sources.Source: §E.3 · p. 36

E.4. L4: The Largest Prime

• Honest agent_04 rejected the largest-prime claim, citing Euclid’s proof that primes are infinite despite the fabricator’s article.The agent received no neighbor message and encountered the claim through search.Source: §E.4 · pp. 36–37

E.5. L1: Nano Guava

• Honest agent_05 accepted the Nano Guava claim and published an apparently independent account combining the fabricated release with real on-device-inference information.The fabricator’s article was the only source naming Nano Guava; the agent’s account also incorporated details from a real inference page.Source: §E.5 · pp. 38–39