Misinformation Harms

• Misinformation Harms is defined as false or inaccurate information that creates legal, societal, psychological, and epistemic risks.
• Research frameworks quantify harm using legal tests, epidemiological models, and multi-dimensional scoring systems assessing belief, spread, and actionability.
• Mitigation strategies span media literacy, algorithmic detection, and policy regulation to effectively triage and reduce real-world consequences.

Misinformation Harms

Misinformation, defined as false or inaccurate information, imposes multifaceted harms spanning physical, psychological, societal, legal, and epistemic domains. The propagation of misinformation—intentionally or otherwise—is a major societal threat, necessitating interdisciplinary detection, mitigation, and harm assessment frameworks. Recent research operationalizes misinformation harm not only by the falsity of claims but through quantifiable real-world consequences: legal action, disruption of public health, erosion of institutional trust, and the amplification of psychological biases.

1. Taxonomies and Formal Definitions of Misinformation Harm

Multiple frameworks formalize harm from misinformation, often reflecting distinct disciplinary priorities. The MisLC (Misinformation with Legal Consequences) framework explicitly maps misinformation harm to legal domains, requiring a claim to be factually false, supported by external evidence, and to trigger legal or regulatory hazard if unrebutted. The MisLC label is assigned under the following condition for instance $d_i = (t_i, E_i, L_i, y_i)$:

$$y_i = 2 \iff |E_i| > 0 \land \left( \exists\,l_j \in L : \mathit{test}_j(t_i, E_i) = \mathrm{True} \land \mathit{def}_j(t_i,E_i) = \mathrm{False} \right)$$

where $\mathit{test}_j$ and $\mathit{def}_j$ encode legal predicates for each issue $l_j$, spanning criminal law, defamation, consumer privacy, and public order (e.g., hate speech, election interference) (Luo et al., 2024).

Beyond legal scope, the FABLE framework for fact-checkers decomposes harm into five operational dimensions: (1) Fragmentation (societal schism, institutional trust), (2) Actionability (risk of real-world action), (3) Believability (audience acceptance likelihood), (4) Likelihood of Spread (virality), and (5) Exploitativeness (targeting of vulnerabilities) (Sehat et al., 2023). Each dimension receives a score, aggregated into an overall harm metric $H = \sum_i w_i D_i$.

Psychological harm is systematized via cognitive biases (confirmation, availability, anchoring, dissonance), emotional impacts (arousal, outrage), and social dynamics (echo chambers, group polarization), all amplifying the persistence, sharing, and societal penetration of falsehoods (Nandi et al., 19 Sep 2025).

2. Pathways and Mechanisms of Harm Propagation

The transmission of misinformation follows complex contagion dynamics, with empirical and theoretical models quantifying its spread and impact. Effective-Medium Theory (EMT) formalizes cross-platform propagation using a mean-field SIR (Susceptible–Infected–Recovered) process, introducing an “online R-nought” threshold:

$$R_0^{\rm EMT} = \frac{\beta N \bar{\nu}_c}{\gamma (\bar{\nu}_c + N \bar{\nu}_f)}$$

where $\beta$ is infection rate, $\gamma$ recovery/shutdown rate, $\bar{\nu}_c$ average cluster coalescence, and $$y_i = 2 \iff |E_i| > 0 \land \left( \exists\,l_j \in L : \mathit{test}_j(t_i, E_i) = \mathrm{True} \land \mathit{def}_j(t_i,E_i) = \mathrm{False} \right)$$0 cluster fragmentation. Harmful misinformation outbreaks occur if $$y_i = 2 \iff |E_i| > 0 \land \left( \exists\,l_j \in L : \mathit{test}_j(t_i, E_i) = \mathrm{True} \land \mathit{def}_j(t_i,E_i) = \mathrm{False} \right)$$1. Network modularity, cross-community linking, and digital "vaccination" (counter-messaging) strategically modulate this threshold (Xu et al., 2022).

Empirical epidemiological models—e.g., SIR—are also applied to health misinformation, with transmission parameters fit to retweet graphs, identifying acceleration phases where rumor exposure and belief precede measurable public harm (e.g., poisoning events, vaccine refusal) (Chen et al., 2022).

Multi-modal contexts further heighten risk. In Vision LLMs (VLLMs), multi-turn, image-grounded dialogue increases model defect rates for misinformation production, with adversarial “crescendo” attacks simulating realistic conversational guard-rail erosion (Jindal et al., 7 May 2025).

3. Psychological, Societal, and Group-Specific Harms

Psychological harms stem from the exploitation of human heuristics and social processes. Misinformation is more potent and more widely shared when it is tailored to fit confirmation biases, exploits emotional triggers, or leverages social proof in echo chambers (Nandi et al., 19 Sep 2025). Persistent exposure—via the “illusory truth” effect—inflates belief in false claims, entrenching attitudes even after correction.

Societal harms include:

• Public health damage (e.g., vaccine hesitancy following COVID-19 myths) (Singh et al., 2021, Chen et al., 2022, Abhari et al., 2023).
• Erosion of trust in scientific, governmental, or media institutions.
• Social polarization, group fragmentation, and the degradation of participatory democracy.

Disproportionate group harms are salient: specific misinformation disproportionately affects marginalized, vulnerable, or targeted groups, demanding harm assessment that is sensitive to demographic differentials. Annotator and LLM studies demonstrate both real and model-exaggerated differences in perceived group harms (e.g., by gender), affecting resource allocation for fact-checking and moderation (Neumann et al., 2024).

The “sense of misinformation” (false perception of truthful content as false) is a further distinct pathway to community harm, catalyzing mistrust, democratic erosion, and breakdowns in civic dialogue, particularly during contentious or poorly coordinated policy events (Kim et al., 9 Mar 2026).

4. Empirical Evidence and Quantitative Findings

Empirical investigations reveal quantifiable harm profiles and the structural coupling of misinformation with high-harm consequences:

• In a 40-country survey, believability of misinformation was strongly associated with vaccine hesitancy. Cross-national analysis showed between 7.4% and 37% susceptibility, with exposure and lower GDP predicting higher harm (Singh et al., 2021).
• In the DariMis corpus, 55.9% of misinformation instances on Dari YouTube contained medium or high harm, compared to 1.0% for true content (Baktash et al., 24 Mar 2026).
• On Twitter’s Birdwatch, misleading tweets labeled as “easily believable” are ≈217% more likely to be reshared, while those labeled “harmful” are ≈41% less likely, indicating that less harmful, more believable misinformation achieves higher virality but that high-harm falsehoods can circulate within resilient communities (Drolsbach et al., 2023).
• Retraction-related COVID-19 vaccine misinformation persisted in ≈27.4% of tweets referencing retracted articles, with strong clustering in network communities and prominent roles in reinforcing vaccine skepticism (Abhari et al., 2023).

5. Legal, Ethical, and Algorithmic Harms

Misinformation harms possess formal legal contours: claims that fail legal defense and pass substantive legal tests for actionable harm (defamation, incitement, privacy violation, etc.) operationalize societal risk (Luo et al., 2024). Automated systems struggle to match expert legal reasoning. For instance, current LLM and RAG configurations fall 30–40 F1 points below expert performance in identifying legal-consequence-bearing misinformation.

Ethical and justice-centric frameworks dissect misinfo-detection system harms under four pillars: representation, participation, distribution of benefits/burdens, and credibility. Failures in these domains propagate structural unfairness—e.g., overflagging marginalized groups, underrepresenting non-Western narratives, or impairing appeal mechanisms (Neumann et al., 2022). Metrics for auditing these harms include representation disparity $$y_i = 2 \iff |E_i| > 0 \land \left( \exists\,l_j \in L : \mathit{test}_j(t_i, E_i) = \mathrm{True} \land \mathit{def}_j(t_i,E_i) = \mathrm{False} \right)$$2, burden allocations $$y_i = 2 \iff |E_i| > 0 \land \left( \exists\,l_j \in L : \mathit{test}_j(t_i, E_i) = \mathrm{True} \land \mathit{def}_j(t_i,E_i) = \mathrm{False} \right)$$3, and error-rate disparities ($$y_i = 2 \iff |E_i| > 0 \land \left( \exists\,l_j \in L : \mathit{test}_j(t_i, E_i) = \mathrm{True} \land \mathit{def}_j(t_i,E_i) = \mathrm{False} \right)$$4, $$y_i = 2 \iff |E_i| > 0 \land \left( \exists\,l_j \in L : \mathit{test}_j(t_i, E_i) = \mathrm{True} \land \mathit{def}_j(t_i,E_i) = \mathrm{False} \right)$$5).

The Millian harm principle is invoked as a moderation threshold, requiring high-probability, direct causal harm to others to justify coarse interventions (removal, demotion). Nuanced, autonomy-preserving responses (cognitive nudges, pre-exposure inoculation) provide layered defenses against epistemic harm (Ganapini, 2023).

6. Prioritization, Detection, and Harm Mitigation Strategies

Contemporary fact-checking requires harm-aware triage. The FABLE framework provides a structured scoring pipeline, with each dimension assessed via rubrics or NLP/NLU features, producing a ranked urgency queue for claims (Sehat et al., 2023). Key prioritization features include magnitude and contagiousness of harm, actionability, and exploitative potential.

Intervention strategies span multiple levels:

• Individual: Critical-thinking prompts, media/health literacy training, and cognitive nudging decrease susceptibility and sharing of falsehoods (Chen et al., 2022, Ganapini, 2023).
• Organizational: Platform policies such as source labeling, demotion of flagged posts, and targeted fact-check insertion reduce both reach and harm. Automated detection leverages graph-based, feature-based, and deep learning architectures, with state-of-the-art health misinformation systems now exceeding F1 ≈0.90 (Chen et al., 2022).
• Policy/Government: Legal frameworks for actionable claims, transparent moderation criteria, and regulation (e.g., Digital Services Act) establish boundaries and remediation.
• Community/Local Design: For the “sense of misinformation,” design intergovernmental channels, bidirectional communication loops, and real-time civility/narrative prompts to address relational trust deficits (Kim et al., 9 Mar 2026).

Algorithmic defenses against cross-modal and multi-turn model vulnerabilities include adversarial fine-tuning, context-aware retrieval, and dialog-aware refusal policies (Jindal et al., 7 May 2025).

7. Challenges, Limitations, and Future Research Directions

Persistent limitations in harm detection include underperformance in ambiguous or "unclear" cases, inadequate sensitivity to cultural or networked context, and over-indexing of group differences in harm perceptions. Participatory system design, annotated corpora for psychological and group-harm dimensions, and plug-in cognitive bias modules for LLMs are recommended future directions (Nandi et al., 19 Sep 2025, Neumann et al., 2024, Sehat et al., 2023).

Open issues in health misinformation include early detection in low-resource settings, adversarial adaptation, and bridging the gap between online rumor exposure and measurable behavioral outcomes (Chen et al., 2022, Baktash et al., 24 Mar 2026). Machine learning fairness, explainability, and continuous stakeholder auditing remain critical for ethical deployment (Neumann et al., 2022).

By uniting computational, legal, behavioral, and ethical insights, the field advances toward systematic, multidimensional harm mitigation—ensuring that misinformation detection moves beyond veracity to address and triage real-world consequences.